Add web map with zone generation API, side job queue, and restore historical DTM hole rendering

- webapp.py: FastAPI serving the continuous map (port 8973) with
  /api/preview, /api/generate and /api/status; tiles are downloaded
  from IGN and processed in a logged subprocess, tracked live in a
  side "File de génération" panel that survives page reloads
- fetch_ign.py: download missing 1 km LiDAR HD tiles from the IGN
  geoplateforme before processing
- index.py: tile thumbnails and 500 m subtiles are now invalidated by
  mtime so regenerating a tile refreshes its cached images; progress
  logging per tile
- dtm.py: back to the historical gap handling (small gaps filled by
  fillnodata only, larger holes left as nodata rendered black);
  lowest-return floor only via --bare-earth, IGN class selection via
  --ign-classes
- cli.py: positional input now optional (--rebuild-index works alone)
- docker-compose.yml: serve (GPU, port 8973) and process services;
  launch via docker compose only (documented in AGENTS.md/AGENTS.md)
- tests: 131 passing, incl. regressions for thumbnail staleness,
  --rebuild-index without input, and nodata rendering

AGENTS.md

@@ -8,6 +8,9 @@
 - lint: not configured
 - format: not configured
 - after every edit: `./run.sh --test`
+- **RÈGLE 1 — toujours lancer via docker compose** (jamais `docker run` direct) : carte/API → `docker compose up -d serve` (port 8973) ; traitement ponctuel → `docker compose run --rm process [options]` ; logs → `docker compose logs -f serve` ; arrêt → `docker compose down`.
+- **RÈGLE 2 — après chaque édition de code : rebuild de l'image puis relance du conteneur.** Le code est baké dans l'image (jamais monté) : sans `docker compose build` suivi d'un `docker compose up -d serve` (recrée le conteneur), l'ANCIEN code continue de tourner. Toujours reconstruire avant de faire tester/valider une modif par l'utilisateur.
+- test rapide sans rebuild (code monté par-dessus l'image): `docker run --rm -e PYTHONPATH=/app -v $(pwd)/lidar_pipeline:/app/lidar_pipeline lidar-lidar python3 -m pytest --pyargs lidar_pipeline.tests`
 - debug: `./run.sh --debug` (file:line logging); container shell: `docker run --rm -it -v $(pwd)/input:/data/input -v $(pwd)/output:/data/output --entrypoint bash lidar-lidar`
 
 ## Conventions
@@ -21,6 +24,7 @@
 - **Default output is AVIF**, not WebP. Use `--format webp` for WebP. Quality default is 98.
 - **Tests use lazy imports inside each test function**, never at module top, to avoid importing CuPy/GDAL at import time.
 - **`_`-prefixed names are critical private**: `_create_ground_pipeline`, `_fallback_to_smrf`, `_fill_nans`, `_init_gpu`, `_process_file_standalone` — do not call from outside their module.
+- **`build_index()` writes 3 files**: `output/index.html` (data shell, `const TILES` embedded), `output/assets/app.css` and `output/assets/app.js` (source: `_APP_CSS`/`_APP_JS` constants in `index.py`). `webapp.py` serves `/assets` with no-cache headers. Each tile carries `meta` — ground method read from `DTM/*_dtm{_rXpY}_method.txt` (falls back to the primary-resolution sidecar) + per-viz dates/sizes.
 
 ## Commit & Pull Request Guidelines
 

docker-compose.yml

@@ -1,38 +1,61 @@
-version: '3.8'
+# Lancement du pipeline LiDAR — TOUJOURS via docker compose :
+#   docker compose build            # après chaque édition de code (code baké dans l'image)
+#   docker compose up -d serve      # carte interactive + API sur http://localhost:8973
+#   docker compose logs -f serve    # journal du serveur
+#   docker compose down             # arrêt
+# Traitement ponctuel (sans serveur) :
+#   docker compose run --rm process [-r 0.5,0.2 | --force | --file ...]
+# Jupyter (opt-in) :
+#   docker compose --profile interactive up -d jupyter
 
 services:
-  lidar:
+  # Carte interactive + API de génération (mode ./run.sh --serve)
+  serve:
     build: .
-    container_name: lidar-archeo
+    image: lidar-lidar
+    container_name: lidar-serve
+    init: true
     user: "1000:1000"
+    gpus: all
+    ports:
+      - "8973:8973"
     volumes:
-      # Mount your LAZ files directory here
+      # input/ en écriture : l'API y télécharge les dalles IGN manquantes
-      - ./input:/data/input:ro
+      - ./input:/data/input
-      # Output directory
       - ./output:/data/output
-      # Optional: Mount a large data directory
-      # - /path/to/your/laz/files:/data/input:ro
     environment:
       - TZ=Europe/Paris
-      # Processing parameters
+      - LIDAR_INPUT_DIR=/data/input
-      - RESOLUTION=0.5
+      - LIDAR_OUTPUT_DIR=/data/output
-      - WHITEBOX_THREADS=4
+      # Les générations lancées depuis la carte utilisent le GPU
-    # Resource limits (adjust based on your system)
+      - LIDAR_GPU=1
-    deploy:
+      - LIDAR_WORKERS=2
-      resources:
+    command: python3 -m uvicorn lidar_pipeline.webapp:app --host 0.0.0.0 --port 8973
-        limits:
+    restart: unless-stopped
-          cpus: '4'
-          memory: 8G
-        reservations:
-          cpus: '2'
-          memory: 4G
-    # Override default command
-    command: ["process_lidar.py", "/data/input", "-o", "/data/output", "-r", "0.5"]
 
-  # Optional: Jupyter notebook for interactive exploration
+  # Traitement ponctuel des dalles input/ (une passe puis arrêt)
+  process:
+    build: .
+    image: lidar-lidar
+    container_name: lidar-process
+    init: true
+    user: "1000:1000"
+    gpus: all
+    volumes:
+      - ./input:/data/input
+      - ./output:/data/output
+    environment:
+      - TZ=Europe/Paris
+    command: ["python3", "-m", "lidar_pipeline", "/data/input", "-o", "/data/output", "-r", "0.5,0.2", "-g", "all"]
+    profiles:
+      - process
+
+  # Exploration interactive (opt-in : --profile interactive)
   jupyter:
     build: .
+    image: lidar-lidar
     container_name: lidar-jupyter
+    init: true
     ports:
       - "8888:8888"
     volumes:

docs/GROUND_CLASSIFICATION.md

@@ -0,0 +1,141 @@
+# Classification du sol — options et références
+
+Note de synthèse pour choisir/améliorer l'algorithme de détection du sol.
+Contexte : tiles LiDAR HD IGN (Lambert 93), zones de relief fort / rochers /
+forêt dense où le sol est sous-classifié et le MNT présente de grands trous.
+
+Tile de référence : `LHD_FXX_0999_6778_PTS_LAMB93_IGN69`
+- 65 047 919 points, ~1 km², résolutions 0.5 m et 0.2 m.
+- Répartition classes (pré-classification fournisseur) :
+  classe 2 (sol) **25.79 %**, classe 5 (veg haute) **63.2 %**,
+  classe 3 (veg basse) 7.85 %, classe 4 (veg moyenne) 2.66 %,
+  classe 1 0.47 %, classe 6 0.01 %. Aucun point en classe 0.
+- Trous dans le MNT existant (avant correction) : **43.5 %** à 0.5 m,
+  **44.1 %** à 0.2 m (surface du tile, bornes du header).
+
+## Benchmark mesuré (PDAL, 1 km²)
+
+| Méthode | Temps | Points sol | Surface sol* | Trous* |
+|---|---|---|---|---|
+| **IGN** (pré-classif.) | **9.4 s** | 25.8 % | 84.6 % | 15.4 % |
+| **SMRF** | 326.3 s | 36.1 % | 90.4 % | 9.6 % |
+| **CSF** | 355.2 s | 17.6 % | 46.9 % | 53.1 % |
+
+\* « Surface sol » calculée sur l'emprise des points (bounding box du nuage),
+à 0.5 m. Les % de trous du MNT final (bornes du header, plus grandes) sont
+supérieurs : voir le tile de référence ci-dessus.
+
+## A. Filtres géométriques (stack PDAL actuelle)
+
+- **IGN** (pré-classification fournisseur, classe 2) : le plus rapide (~9 s).
+  Fiable là où le fournisseur a confiance ; trous sous forêt dense / relief.
+  Aucun paramètre à régler.
+- **SMRF** — Pingel, Clarke & McBride 2013, *ISPRS J. Photogramm. Remote
+  Sens.* 77:21-30. Filtre **raster** (opère sur un DSM, pas sur les points),
+  donc plus rapide que les filtres point-based ; **minimise les erreurs de
+  type I** (omission de sol) → bien adapté quand le sol est rare (forêt).
+  Meilleure couverture des trois ici (90.4 %) mais ~5.4 min/tile.
+- **CSF** — Zhang et al. 2016, *Remote Sensing* 8(6):501. Toile inversée
+  drapée sur le nuage ; simple, précis, mais **la toile ne touche plus le sol
+  en terrain raide/vallonné** → mauvaise classification. Le plus lent ici et
+  le pire sur ce tile. À réserver aux zones urbaines.
+- **PTD/PTIN** (Progressive TIN Densification) — Axelsson 2000, ISPRS
+  Congress. **Gagnant de la littérature** : le plus robuste sur terrain
+  complexe + forêt (Moudrý et al. 2020, *Measurement* 150:107047 ; Cai et al.
+  2019, *Remote Sensing* 11(9):1037) et le plus rapide (benchmark lidR :
+  PTD ~20 s vs CSF ~156 s vs PMF ~1800 s). **NON disponible dans la version
+  PDAL de cette image** (`filters.ground` / TIN absents) — à ajouter pour
+  l'utiliser (ou via lidR / une implémentation maison).
+
+## B. Hybride rapide (choisi pour implémentation)
+
+Principe **PTD / Wack & Wimmer** (Wack & Wimmer 2002, *ISPRS Archives*
+XXXIV/3A:293-296 : MNT par retour le plus bas, en excluant le 1 % le plus bas
+par cellule pour écarter les outliers) :
+
+1. **Base = pré-classification IGN** (classe 2, ~9 s, fiable et officielle).
+2. **Comblement mesuré des trous** : pour chaque cellule sans point sol,
+   prendre le **retour le plus bas robuste** (min du 99 % des points de la
+   cellule) → ajoute du sol *mesuré* là où le fournisseur a échoué
+   (rochers, clairières, sol forestier).
+3. **Inpainting topographique** des derniers vides (interpolation
+   terrain-aware déjà implémentée dans `dtm.py:_interpolate_holes`).
+
+Attendu : MNT **continu** (0 % de trous), robuste en forêt/relief,
+**~10-15 s/tile** au lieu de 326-355 s. Aucune dépendance GPU, aucun
+entraînement.
+
+## C. Modèles IA / ML (supervisés — nécessitent des labels)
+
+Avertissement (Qin et al. 2023, *ISPRS J. Photogramm. Remote Sens.*
+202:246-261) : **tout est supervisé** ; le principal risque est la
+**généralisation** — un modèle entraîné sur une région dégrade ailleurs.
+Aucun filtre DL entièrement non-supervisé publié à date.
+
+**Basés sur les points (3D) :**
+
+| Modèle | Année | Archi | Précision | Vitesse (~/km², GPU) |
+|---|---|---|---|---|
+| KPConv / RandLA-Net (Qin, OpenGF) | 2021 | KPConv / RandLA-Net | 97.8 % OA, RMSE DTM 0.20 m, IoU sol 95 % | 0.5-2.5 min |
+| PFCN (Jin, *IEEE JSTARS* 13:3958) | 2020 | point-FCN | Te 1.73 %, Kappa 93.9 % | ~1/3 du coût PointNet++ |
+| Terrain-Net (Li, *Remote Sensing* 14(22):5798) | 2022 | KPConv + self-attention | OA 98 %, mIoU 0.933 | param-free au transfert |
+| MSVC (Štroner, *Remote Sensing* 17(4):615) | 2025 | DNN voxel 9x9x9 | bat CSF en F-score | — |
+
+**Rasterisés (sortent directement le MNT — le plus proche du besoin) :**
+
+| Modèle | Année | Archi | Résultat |
+|---|---|---|---|
+| Precursor (Rizaldy, *ISPRS Annals* IV-2:231) | 2018 | 2D FCN | Te 5.22 %, 78x plus rapide |
+| DeepTerRa / ALS2DTM (Lê, *IEEE JSTARS* 15:2778) | 2022 | GAN pix2pix (U-Net) | RMSE MNT < 1 m, filtre + interp en 1 passe |
+| DSM2DTM (Bittner, *ISPRS Annals* X-1/W1-2023:925) | 2023 | U-Net (EfficientNet) | masque non-sol + hauteur sol/pixel |
+
+**Jeu de données d'entraînement** : OpenGF (Qin et al., CVPRW 2021,
+arXiv:2101.09641 — 47.7 km², 542 M pts) ; ALS2DTM (Lê et al. 2022,
+arXiv:2206.03778 — 52 km², 1.66 Md pts, urbain/forêt/montagne).
+
+**Coûts / obstacles pour notre cas** : (1) labels → à générer en
+pseudo-labels (sortie SMRF/PTD haute qualité sur un échantillon représentatif
+de nos tiles) ou pré-entraînement OpenGF/ALS2DTM ; (2) généralisation sur le
+terrain divers de LiDAR HD (plaine/forêt/montagne/urbain) ; (3) infra :
+checkpoint + chemin d'inférence GPU dans l'image Docker.
+
+**Meilleur fit si on part sur l'IA** : un **U-Net rasterisé (style
+DSM2DTM)** — rasteriser le nuage en grilles multi-canaux (altitude, pente,
+courbure, densité, stats de retours), sortir masque sol + hauteur sol.
+2D = très rapide et trivial à déployer sur GPU, fusionne filtrage +
+interpolation. Le KPConv/RandLA-Net est plus précis en 3D pur mais plus lourd
+à déployer.
+
+## Synthèse / décision
+
+- « Rapide » contrainte dure + faible maintenance → **hybride (B)**
+  (~10-40 s/tile, zéro entraînement, zéro GPU). ← **choix retenu, IMPLÉMENTÉ**
+  - Base = pré-classification IGN (rapide, ~10 s). `auto` la préfère dès que
+    ≥ 20 % des points sont classés sol (seuil abaissé de 30 % à 20 %, car le
+    MNT est ensuite complété — voir ci-dessous).
+  - Le MNT est **toujours** complété dans `create_dtm_fast` : comblement par le
+    **retour le plus bas par cellule** (`_min_return_grid`, Wack & Wimmer 2002)
+    pour les trous, puis **interpolation terrain-aware** (`_interpolate_holes`).
+    Résultat : MNT continu (0 % de trous) pour n'importe quelle base.
+  - Vérifié sur le tile 0999_6778 : base CSF + comblement → 1,9 M trous
+    comblés par retour le plus bas, 5,3 % interpolés, MNT 0 % de trous.
+- Qualité max dans les cas durs (raide + dense), ~1-2 min/tile + GPU +
+  entraînement acceptés → **U-Net rasterisé (C)**. (non implémenté)
+- Meilleur filtre géométrique disponible dans PDAL → **SMRF (A)** (meilleure
+  couverture 90.4 % mais 5.4 min/tile). Sélectionnable via `--ground-classification smrf`.
+- « Gagnant » absolu de la littérature (rapide + robuste) → **PTD/PTIN
+  (A)** : à intégrer (pas dans la stack PDAL actuelle).
+
+## Références
+
+- Axelsson (2000), PTIN/PTD, ISPRS Congress.
+- Pingel, Clarke & McBride (2013), SMRF, ISPRS J. P&RS 77:21-30.
+- Zhang et al. (2016), CSF, Remote Sensing 8(6):501.
+- Wack & Wimmer (2002), DMT par retour le plus bas, ISPRS Archives XXXIV/3A.
+- Moudrý et al. (2020), comparaison CSF/PTIN/PMF/SMRF, Measurement 150:107047.
+- Cai et al. (2019), CS+PTD, Remote Sensing 11(9):1037.
+- Qin et al. (2021), OpenGF, CVPR Workshops (arXiv:2101.09641).
+- Qin et al. (2023), dataset + evaluation + survey, ISPRS J. P&RS 202:246-261.
+- Lê et al. (2022), DeepTerRa/ALS2DTM, IEEE JSTARS 15:2778 (arXiv:2206.03778).
+- Bittner et al. (2023), DSM2DTM, ISPRS Annals X-1/W1-2023:925.
+- lidR book (comparaison PTD/CSF/PMF) : https://r-lidar.github.io/lidRbook/gnd.html

lidar_pipeline/cli.py

@@ -97,7 +97,10 @@ def main():
     )
     parser.add_argument(
         "input",
-        help="Dossier contenant les fichiers LAZ/LAS"
+        nargs="?",
+        default="/data/input",
+        help="Dossier contenant les fichiers LAZ/LAS (défaut: /data/input ; "
+             "optionnel pour --rebuild-index)"
     )
     parser.add_argument(
         "-o", "--output",
@@ -134,7 +137,16 @@ def main():
     parser.add_argument(
         "--force-classification",
         action="store_true",
-        help="Reclassifier le sol même si le fichier .las existe déjà"
+        help="Reclassifier le sol même si la méthode est inchangée (régénère aussi le DTM "
+             "et les images). Sans ce flag, changer --ground-classification suffit : la "
+             "méthode enregistrée est comparée et un changement déclenche la reclassification."
+    )
+    parser.add_argument(
+        "--bare-earth",
+        action="store_true",
+        help="Sol nu : ramener le DTM au retour le plus bas de chaque cellule. "
+             "Requalifie le point le plus bas de chaque colonne en terrain — utile sous "
+             "végétation dense ou en relief raide où la classification du sol sous-couvre le terrain."
     )
     parser.add_argument(
         "--keep-tif",
@@ -143,9 +155,22 @@ def main():
     )
     parser.add_argument(
         "--ground-classification",
-        choices=["auto", "smrf", "csf"],
+        choices=["auto", "ign", "smrf", "csf"],
         default="auto",
-        help="Méthode de classification du sol : auto (détection), smrf, csf (défaut: auto)"
+        help="Méthode de classification du sol : auto (préfère la pré-classification IGN si "
+             "présente — base rapide — sinon détection SMRF/CSF), ign, smrf, csf. "
+             "Avec ign, le MNT est la rasterisation pure des classes choisies "
+             "(--ign-classes) sans aucune retouche ; avec smrf/csf, il est complété "
+             "par le retour le plus bas par cellule + interpolation des trous. (défaut: auto)"
+    )
+    parser.add_argument(
+        "--ign-classes",
+        default="sol",
+        help="Classes LAS extraites pour le MNT avec la classification IGN (méthode "
+             "ign/auto) : liste noms ou codes séparés par virgules — "
+             "sol(2), unclassified(1), eau(9), virtuel(66), pont(17), sursol(64). "
+             "Ex: --ign-classes sol,unclassified. Changer la liste reclassifie les "
+             "dalles concernées. (défaut: sol)"
     )
     parser.add_argument(
         "--quality",
@@ -185,6 +210,14 @@ def main():
         default=None,
         help="Traiter un ou plusieurs fichiers LAZ/LAS (nom complet sans extension, ex: LHD_FXX_1000_6882_PTS_LAMB93_IGN69.copc)"
     )
+    parser.add_argument(
+        "--fetch-tiles",
+        nargs="+",
+        default=None,
+        metavar="COL,ROW",
+        help="Télécharger ces dalles LiDAR HD depuis l'IGN avant traitement "
+             "(tuiles non encore générées, ex: --fetch-tiles 1055,6882 1056,6883)"
+    )
     parser.add_argument(
         "-v", "--verbose",
         action="store_true",
@@ -253,6 +286,21 @@ def main():
                 logger.warning("Aucune tuile traitée trouvée — carte globale non générée")
             return
 
+        # Téléchargement des dalles IGN manquantes avant le traitement
+        if args.fetch_tiles:
+            from .fetch_ign import fetch_tiles, parse_tile_specs
+            try:
+                specs = parse_tile_specs(args.fetch_tiles)
+            except ValueError as e:
+                logger.error(str(e))
+                return
+            logger.info(f"Téléchargement de {len(specs)} dalle(s) LiDAR HD depuis l'IGN...")
+            fetched = fetch_tiles(args.input, specs, args.output)
+            if fetched:
+                logger.info(f"{len(fetched)} dalle(s) téléchargée(s) — traitement...")
+            else:
+                logger.warning("Aucune dalle téléchargée (déjà présentes ou introuvables)")
+
         quality = 100 if args.lossless else args.quality
         # Parse --only and --skip: accept comma-separated values
         only_viz = None
@@ -268,8 +316,10 @@ def main():
             workers=args.workers,
             force=args.force,
             ground_method=args.ground_classification,
+            ign_classes=args.ign_classes,
             force_classify=args.force_classification,
             keep_tif=args.keep_tif,
+            bare_earth=args.bare_earth,
             quality=quality,
             only_viz=only_viz,
             skip_viz=skip_viz,
@@ -315,30 +365,8 @@ def main():
             logger.info(f"Traitement de {len(unique_files)} fichier(s) sélectionné(s)")
             for laz_file in unique_files:
                 logger.info(f"  → {laz_file.name}")
-            for laz_file in unique_files:
+            # Réutilise process_all : workers parallèles, résumé, index, nettoyage
-                pipeline.process_file(laz_file)
+            pipeline.process_all(files=unique_files)
-
-            # Clean up temporary files
-            logger.info("Nettoyage des fichiers temporaires...")
-            try:
-                if pipeline.temp_dir.exists():
-                    shutil.rmtree(pipeline.temp_dir)
-                temp_base = pipeline.output_dir / "temp"
-                if temp_base.exists():
-                    shutil.rmtree(temp_base)
-                logger.info("  ✓ Fichiers temporaires supprimés")
-            except Exception as e:
-                logger.warning(f"  Note: Impossible de supprimer les fichiers temporaires: {e}")
-
-            # Génère la carte globale après traitement --file
-            if not args.no_index:
-                try:
-                    from .index import build_index
-                    index_path = build_index(pipeline.output_dir, pipeline.output_format)
-                    if index_path:
-                        logger.info(f"Carte globale générée : {index_path}")
-                except Exception as e:
-                    logger.warning(f"Index global non généré: {e}")
         else:
             pipeline.process_all()
     except Exception as e:

lidar_pipeline/dtm.py

@@ -1,7 +1,10 @@
 """DTM generation from classified LiDAR point clouds.
 
-Handles ground classification via PDAL (SMRF or CSF) and DTM rasterisation
+Handles ground classification via PDAL (IGN supplier pre-classification,
-using scipy binned_statistic_2d. Zones without LiDAR data remain as NaN.
+SMRF or CSF) and DTM rasterisation
+using scipy binned_statistic_2d. Gaps without LiDAR data (common in
+complex/rocky terrain) are filled with a terrain-aware interpolation so the
+DTM stays continuous.
 """
 
 import json
@@ -16,8 +19,67 @@ from scipy.stats import binned_statistic_2d
 
 logger = logging.getLogger("lidar")
 
+# Classes LAS exploitables de la pré-classification LiDAR HD (noms → codes)
+IGN_CLASS_NAMES = {
+    "sol": 2,
+    "unclassified": 1,
+    "non-classe": 1,
+    "eau": 9,
+    "virtuel": 66,
+    "pont": 17,
+    "sursol": 64,
+}
 
-def _create_ground_pipeline(input_laz, output_las, method):
+
+def parse_ign_classes(spec):
+    """Convertit une liste de classes IGN (noms ou codes) en codes LAS triés.
+
+    Args:
+        spec: Chaîne séparée par virgules, ex. "sol,unclassified" ou "2,1".
+
+    Returns:
+        Liste triée de codes LAS uniques.
+
+    Raises:
+        ValueError: Si un élément n'est ni un nom connu ni un code LAS 0-255,
+            ou si la liste est vide.
+    """
+    codes = set()
+    for token in str(spec).split(","):
+        token = token.strip().lower()
+        if not token:
+            continue
+        if token in IGN_CLASS_NAMES:
+            codes.add(IGN_CLASS_NAMES[token])
+        else:
+            try:
+                code = int(token)
+            except ValueError:
+                raise ValueError(
+                    f"Classe IGN inconnue: '{token}' "
+                    f"(noms: {', '.join(sorted(IGN_CLASS_NAMES))} ou code LAS 0-255)")
+            if not 0 <= code <= 255:
+                raise ValueError(f"Code LAS hors bornes (0-255): {code}")
+            codes.add(code)
+    if not codes:
+        raise ValueError("Aucune classe IGN fournie")
+    return sorted(codes)
+
+
+def ign_method_label(codes):
+    """Étiquette de méthode encodant les classes IGN (ex. 'ign_1_2').
+
+    'ign' seul = sol uniquement (code 2), rétrocompatible avec les fichiers de
+    classification existants. Toute autre combinaison est encodée dans le nom
+    pour invalider le cache et déclencher la reclassification.
+    """
+    codes = sorted(codes)
+    if codes == [2]:
+        return "ign"
+    return "ign_" + "_".join(str(c) for c in codes)
+
+
+def _create_ground_pipeline(input_laz, output_las, method, ign_codes=None):
     """Create a PDAL pipeline JSON for ground classification.
 
     All methods include a ReturnNumber/NumberOfReturns >= 1 filter to handle
@@ -33,7 +95,10 @@ def _create_ground_pipeline(input_laz, output_las, method):
     Args:
         input_laz: Path to input LAZ/LAS file.
         output_las: Path to output classified LAS file.
-        method: Ground classification method ('smrf' or 'csf').
+        method: Ground classification method ('ign', 'smrf' or 'csf').
+        ign_codes: LAS class codes to extract with the 'ign' method
+            (default: [2] = sol). Multiple ranges on Classification are
+            logically ORed by filters.range (documented PDAL semantics).
 
     Returns:
         JSON string of the PDAL pipeline.
@@ -44,6 +109,38 @@ def _create_ground_pipeline(input_laz, output_las, method):
         "limits": "ReturnNumber[1:],NumberOfReturns[1:]"
     }
 
+    # Classification filter (ground points only)
+    ground_filter = {
+        "type": "filters.range",
+        "limits": "Classification[2:2]"
+    }
+
+    # LiDAR HD IGN : le fichier est pré-classifié par le fournisseur.
+    # On réutilise la classification telle quelle (mode pur) : les classes
+    # extraites sont paramétrables — par défaut le sol seul (2), mais on peut
+    # ajouter p.ex. unclassified (1) pour combler les trous sans retouche.
+    # Les plages multiples sur Classification sont combinées en OU logique
+    # par filters.range (sémantique PDAL documentée).
+    if method == 'ign':
+        codes = sorted(ign_codes) if ign_codes else [2]
+        ign_filter = {
+            "type": "filters.range",
+            "limits": ",".join(f"Classification[{c}:{c}]" for c in codes)
+        }
+        pipeline = {
+            "pipeline": [
+                str(input_laz),
+                return_filter,
+                ign_filter,
+                {
+                    "type": "writers.las",
+                    "filename": str(output_las),
+                    "extra_dims": "all"
+                }
+            ]
+        }
+        return json.dumps(pipeline)
+
     # Reset Classification to 0 before preprocessing
     reset_classification = {
         "type": "filters.assign",
@@ -68,12 +165,6 @@ def _create_ground_pipeline(input_laz, output_las, method):
         "multiplier": 3.0
     }
 
-    # Classification filter (ground points only)
-    ground_filter = {
-        "type": "filters.range",
-        "limits": "Classification[2:2]"
-    }
-
     # Method-specific ground classification filter
     if method == 'smrf':
         ground_step = {
@@ -85,9 +176,12 @@ def _create_ground_pipeline(input_laz, output_las, method):
             "scalar": 1.25
         }
     elif method == 'csf':
+        # resolution 1.0 m : un cloth à 0.5 m (= 4 M particules pour 1 km²)
+        # rend la classification ~4× plus lente sans gain visible sur le MNT
+        # (la résolution finale du MNT vient de la rasterisation, pas du cloth).
         ground_step = {
             "type": "filters.csf",
-            "resolution": 0.5,
+            "resolution": 1.0,
             "rigidness": 3,
             "smooth": True,
             "threshold": 0.5
@@ -119,6 +213,11 @@ def create_smrf_pipeline(input_laz, output_las):
     return _create_ground_pipeline(input_laz, output_las, 'smrf')
 
 
+def create_ign_pipeline(input_laz, output_las):
+    """Create a PDAL pipeline JSON using the IGN supplier pre-classification."""
+    return _create_ground_pipeline(input_laz, output_las, 'ign')
+
+
 def create_csf_pipeline(input_laz, output_las):
     """Create a PDAL pipeline JSON for CSF ground classification."""
     return _create_ground_pipeline(input_laz, output_las, 'csf')
@@ -258,7 +357,7 @@ def detect_ground_method(laz_file):
         laz_file: Path to input LAZ/LAS file.
 
     Returns:
-        String: 'smrf' or 'csf'
+        String: 'ign', 'smrf' or 'csf'
     """
     import laspy
 
@@ -280,6 +379,22 @@ def detect_ground_method(laz_file):
         logger.warning(f"  Nuage vide (0 points) — méthode par défaut: SMRF")
         return 'smrf'
 
+    # LiDAR HD IGN : les données livrées sont pré-classifiées par le fournisseur
+    # (classe 2 = sol). C'est la base la plus rapide (~10 s) et de référence.
+    # Le MNT est ensuite complété par le retour le plus bas par cellule +
+    # interpolation (voir create_dtm_fast), ce qui « rattrape » les trous de la
+    # pré-classification (forêt dense / relief). On la préfère donc dès qu'une
+    # part raisonnable des points est classée sol, plutôt que de refiltrer.
+    try:
+        cls = np.asarray(las.classification, dtype=np.int32)
+        ground_ratio = float(np.mean(cls == 2))
+    except Exception:
+        ground_ratio = 0.0
+    if ground_ratio >= 0.2:
+        logger.info(f"  → Méthode: IGN (pré-classification fournisseur — "
+                    f"{ground_ratio * 100:.1f}% de points classe 2)")
+        return 'ign'
+
     z = np.array(las.z)
 
     # Height variance (always available)
@@ -317,14 +432,17 @@ def detect_ground_method(laz_file):
     return method
 
 
-def classify_ground(laz_file, temp_dir, method='auto', force=False):
+def classify_ground(laz_file, temp_dir, method='auto', force=False, ign_classes="sol"):
     """Classify ground points using PDAL ground classification filter.
 
     Args:
         laz_file: Path to input LAZ/LAS file.
         temp_dir: Directory for temporary files (pipeline.json, ground.las).
-        method: Ground classification method ('auto', 'smrf', or 'csf').
+        method: Ground classification method ('auto', 'ign', 'smrf' or 'csf').
         force: If True, reclassify even if output file already exists.
+        ign_classes: Classes LAS extraites par la méthode IGN (noms ou codes
+            séparés par virgules, ex. "sol,unclassified"). Ignoré pour les
+            autres méthodes.
 
     Returns:
         Path to classified ground LAS file, or None on failure.
@@ -338,11 +456,17 @@ def classify_ground(laz_file, temp_dir, method='auto', force=False):
     else:
         logger.info(f"  Classification sol: {method.upper()} (forcé)")
 
+    # Les classes IGN sont encodées dans le nom de fichier (ex. ign_1_2)
+    # pour qu'un changement de classes invalide le cache et déclenche la
+    # reclassification.
+    ign_codes = parse_ign_classes(ign_classes) if method == 'ign' else None
+    method_label = ign_method_label(ign_codes) if ign_codes else method
+
     # Use shared basename extraction function
     from .pipeline import _file_basename
     laz_base = _file_basename(laz_file)
 
-    output_las = temp_dir / f"{laz_base}_ground_{method}.las"
+    output_las = temp_dir / f"{laz_base}_ground_{method_label}.las"
 
     if output_las.exists() and not force:
         logger.info(f"  Classification {method.upper()} déjà effectuée — fichier existant réutilisé")
@@ -352,8 +476,8 @@ def classify_ground(laz_file, temp_dir, method='auto', force=False):
         logger.info(f"  Reclassification forcée — suppression de {output_las.name}")
         output_las.unlink()
 
-    pipeline_json = _create_ground_pipeline(laz_file, output_las, method)
+    pipeline_json = _create_ground_pipeline(laz_file, output_las, method, ign_codes=ign_codes)
-    pipeline_file = temp_dir / f"pipeline_{method}.json"
+    pipeline_file = temp_dir / f"pipeline_{method_label}.json"
 
     with open(pipeline_file, 'w') as f:
         f.write(pipeline_json)
@@ -367,9 +491,9 @@ def classify_ground(laz_file, temp_dir, method='auto', force=False):
         if output_las.exists() and output_las.stat().st_size < 100:
             logger.error(f"  ✗ Fichier ground vide (taille < 100 octets)")
             output_las.unlink(missing_ok=True)
-            # Fallback: if CSF produced no ground points, retry with SMRF
+            # Fallback: si la méthode ne produit aucun point sol, réessayer avec SMRF
-            if method == 'csf':
+            if method in ('csf', 'ign'):
-                return _fallback_to_smrf(laz_file, temp_dir, laz_base, force)
+                return _fallback_to_smrf(laz_file, temp_dir, laz_base, force, source=method_label)
             return None
         logger.info(f"  ✓ Classification sol {method.upper()} terminée")
         return output_las
@@ -377,9 +501,9 @@ def classify_ground(laz_file, temp_dir, method='auto', force=False):
         error_msg = e.stderr.decode() if e.stderr else str(e)
         logger.warning(f"  ✗ Erreur classification PDAL ({method.upper()}): {error_msg}")
 
-        # Fallback: if CSF failed, retry with SMRF
+        # Fallback: si CSF ou la pré-classification échouent, réessayer avec SMRF
-        if method == 'csf':
+        if method in ('csf', 'ign'):
-            return _fallback_to_smrf(laz_file, temp_dir, laz_base, force)
+            return _fallback_to_smrf(laz_file, temp_dir, laz_base, force, source=method_label)
 
         # Try repairing file with laspy if PDAL fails on EVLR/VLR
         if 'VLR' in error_msg or 'Invalid' in error_msg:
@@ -405,28 +529,30 @@ def classify_ground(laz_file, temp_dir, method='auto', force=False):
         return None
 
 
-def _fallback_to_smrf(laz_file, temp_dir, laz_base, force=False):
+def _fallback_to_smrf(laz_file, temp_dir, laz_base, force=False, source='csf'):
-    """Retry ground classification with SMRF when CSF fails.
+    """Retry ground classification with SMRF when CSF/IGN fails.
 
     CSF (Cloth Simulation Filter) can fail on certain terrain types where
-    SMRF (Simple Morphological Filter) succeeds. This fallback ensures
+    SMRF (Simple Morphological Filter) succeeds, and a file without usable
-    processing continues even when auto-detection selects CSF incorrectly.
+    pre-classification produces an empty ground extract. This fallback ensures
+    processing continues even when the selected method fails.
 
     Args:
         laz_file: Path to input LAZ/LAS file.
         temp_dir: Directory for temporary files.
         laz_base: Base name for the file.
         force: If True, reclassify even if output exists.
+        source: Method that failed ('csf' or 'ign').
 
     Returns:
         Path to classified ground LAS file, or None on failure.
     """
-    logger.info(f"  → Basculement CSF → SMRF (fallback)")
+    logger.info(f"  → Basculement {source.upper()} → SMRF (fallback)")
 
-    # Clean up failed CSF output if it exists
+    # Clean up failed output if it exists
-    csf_output = temp_dir / f"{laz_base}_ground_csf.las"
+    failed_output = temp_dir / f"{laz_base}_ground_{source}.las"
-    if csf_output.exists():
+    if failed_output.exists():
-        csf_output.unlink(missing_ok=True)
+        failed_output.unlink(missing_ok=True)
 
     output_las = temp_dir / f"{laz_base}_ground_smrf.las"
 
@@ -481,7 +607,118 @@ def _repair_laz_with_laspy(input_laz, output_las):
         return False
 
 
-def create_dtm_fast(las_file, basename, dtm_dir, resolution, force=False, output_suffix=""):
+def _interpolate_holes(dtm, downsample=8):
+    """Fill remaining NaN holes with a terrain-aware surface interpolation.
+
+    In complex / rocky terrain the ground under-classification leaves interior
+    holes far too large for a 1 m gap fill, which otherwise become flat
+    nearest-neighbor patches in the downstream layers. This helper triangulates
+    the valid cells on a downsampled grid (linear, nearest as a fallback for
+    cells outside the data hull) and bilinearly upsamples the result, keeping
+    the operation fast even for large rasters.
+
+    Args:
+        dtm: 2-D float array (may contain NaN holes).
+        downsample: Coarsening factor for the interpolation grid.
+
+    Returns:
+        Tuple (filled_array, filled_count).
+    """
+    holes = np.isnan(dtm)
+    if not holes.any():
+        return dtm, 0
+    valid = ~holes
+    if not valid.any():
+        return dtm, 0
+
+    height, width = dtm.shape
+    step = max(1, downsample)
+    coarse = dtm[::step, ::step].astype(np.float64)
+    c_valid = ~np.isnan(coarse)
+    c_holes = np.isnan(coarse)
+    if not c_holes.any() or not c_valid.any():
+        return dtm, 0
+
+    from scipy.interpolate import griddata
+    from scipy.ndimage import map_coordinates
+
+    cy, cx = np.where(c_valid)
+    c_coords = np.column_stack([cx, cy]).astype(np.float64)
+    c_vals = coarse[c_valid]
+    hy, hx = np.where(c_holes)
+    h_coords = np.column_stack([hx, hy]).astype(np.float64)
+
+    interp = griddata(c_coords, c_vals, h_coords, method='linear')
+    bad = np.isnan(interp)
+    if bad.any():
+        interp[bad] = griddata(c_coords, c_vals, h_coords[bad], method='nearest')
+    coarse_filled = coarse.copy()
+    coarse_filled[c_holes] = interp
+
+    # Coarse cell i represents fine column/row i*step, so fine index c maps to
+    # coarse coordinate c/step (no half-cell offset).
+    rows = np.arange(height) / step
+    cols = np.arange(width) / step
+    grid_y, grid_x = np.meshgrid(rows, cols, indexing='ij')
+    upsampled = map_coordinates(coarse_filled, [grid_y, grid_x], order=1)
+
+    filled = dtm.copy()
+    filled[holes] = upsampled[holes]
+    return filled, int(holes.sum())
+
+
+def _min_return_grid(laz_file, width, height, bounds, chunk_size=2_000_000):
+    """Rasterize the per-cell minimum z (lowest return) of the full point cloud.
+
+    In complex/forested terrain the ground is under-classified, leaving DTM
+    holes. Filling them with the *lowest measured return* of the cell (Wack &
+    Wimmer 2002) recovers a real ground surface (forest floor, rock, clearing)
+    instead of a pure interpolation. The read is streamed in chunks so memory
+    stays bounded to the output grid regardless of the point count.
+
+    Args:
+        laz_file: Path to the full (unclassified) LAZ/LAS file.
+        width, height: Output grid dimensions (pixels).
+        bounds: (min_x, min_y, max_x, max_y) the grid covers.
+        chunk_size: Points per streaming chunk.
+
+    Returns:
+        (height, width) float32 array of per-cell min z (NaN where no point).
+    """
+    import laspy
+    min_x, min_y, max_x, max_y = bounds
+    grid = np.full((height, width), np.nan, dtype=np.float32)
+    rng = [[min_x, max_x], [min_y, max_y]]
+
+    def process(points):
+        if len(points) == 0:
+            return
+        x = np.asarray(points.x, dtype=np.float64)
+        y = np.asarray(points.y, dtype=np.float64)
+        z = np.asarray(points.z, dtype=np.float64)
+        st = binned_statistic_2d(x, y, z, statistic='min',
+                                 bins=[width, height], range=rng)
+        # Match the DTM convention: .T then flip Y (north at top).
+        cell_min = st.statistic.T[::-1, :].astype(np.float32)
+        # fmin ignores NaN so cells without a point in this chunk stay NaN.
+        np.fmin(grid, cell_min, out=grid)
+
+    try:
+        with laspy.open(str(laz_file)) as las:
+            for chunk in las.chunk_iterator(chunk_size):
+                process(chunk)
+    except Exception as e:
+        logger.warning(f"    Lecture streaming impossible ({e}) — lecture complète")
+        las = _read_with_pdal(laz_file)
+        if las is None:
+            return grid
+        process(las)
+    return grid
+
+
+def create_dtm_fast(las_file, basename, dtm_dir, resolution, force=False,
+                    output_suffix="", source_laz=None, bare_earth=False,
+                    pure=False):
     """Create DTM using fast binning method with gap filling.
 
     Args:
@@ -491,6 +728,15 @@ def create_dtm_fast(las_file, basename, dtm_dir, resolution, force=False, output
         resolution: Grid resolution in meters per pixel.
         force: If True, regenerate even if DTM already exists.
         output_suffix: Suffix for output filename (e.g. '_r0p2' for additional resolutions).
+        source_laz: Optionnel : chemin du LAZ complet (non classé). Utilisé
+            uniquement avec bare_earth (plancher au retour le plus bas).
+        bare_earth: If True, pull the DTM down to the lowest measured return of
+            each cell (bare-earth floor). This requalifies the lowest point of
+            every column as terrain, recovering the ground under dense
+            vegetation / steep relief that the ground classifier rejected.
+        pure: Sans effet (conservé pour compatibilité). Fonctionnement
+            historique rétabli : petits trous comblés par fillnodata, grands
+            trous laissés en nodata (rendus en noir dans les rendus).
 
     Returns:
         Path to output DTM GeoTIFF, or None on failure.
@@ -542,7 +788,19 @@ def create_dtm_fast(las_file, basename, dtm_dir, resolution, force=False, output
         dtm = stat.statistic.T
         dtm = dtm[::-1, :]  # Flip Y so north is at top
 
-        # Fill small gaps (< 1m from existing data) while keeping large gaps as NaN
+        # Comblement « historique » (fonctionnement d'origine, rétabli) :
+        # seuls les petits trous proches des données sont remplis ; les grands
+        # trous restent en nodata et apparaissent en noir dans les rendus.
+        # Le plancher au retour le plus bas n'est appliqué qu'à la demande
+        # explicite (--bare-earth).
+        if bare_earth and source_laz is not None:
+            min_grid = _min_return_grid(source_laz, width, height,
+                                        (min_x, min_y, max_x, max_y))
+            lower = ~np.isnan(min_grid) & (min_grid < dtm)
+            dtm = np.where(lower, min_grid, dtm)
+            logger.info(f"    Sol nu : {int(lower.sum()):,} cellules raménées au retour le plus bas")
+
+        # Fill small gaps (< 1 m from data) precisely — comme avant
         nan_count = np.count_nonzero(np.isnan(dtm))
         if nan_count > 0:
             total = dtm.size
@@ -558,8 +816,6 @@ def create_dtm_fast(las_file, basename, dtm_dir, resolution, force=False, output
             if filled_count > 0:
                 dtm = np.where(small_gap_mask, dtm_filled, dtm)
                 logger.info(f"    {filled_count:,} petits trous comblés (< {max_gap_pixels}px)")
-            remaining = np.count_nonzero(np.isnan(dtm))
-            logger.info(f"    {remaining:,} pixels restent sans données (grands écarts)")
 
         # Save as GeoTIFF
         output_tif = dtm_dir / f"{basename}_dtm{output_suffix}.tif"

lidar_pipeline/fetch_ign.py

@@ -0,0 +1,162 @@
+"""Téléchargement des dalles LiDAR HD de l'IGN pour les tuiles non générées.
+
+Catalogue STAC (à jour) : https://browser.stac.teledetection.fr/collections/lidarhd
+API                     : https://api.stac.teledetection.fr/collections/lidarhd/items
+Fichiers (géoplateforme): https://data.geopf.fr/telechargement/download/...
+
+Chaque dalle couvre 1 km × 1 km en Lambert 93 et est nommée par son coin
+nord-ouest : LHD_FXX_{col}_{row}_PTS_LAMB93_IGN69.copc.laz
+(col = X ouest en km, row = Y nord en km, cf. propriété STAC
+"lidarhd:coordonnees_NW" au format "0816-6847").
+"""
+
+import json
+import logging
+import time
+import urllib.parse
+import urllib.request
+from pathlib import Path
+
+logger = logging.getLogger("lidar")
+
+_STAC_ITEMS_URL = "https://api.stac.teledetection.fr/collections/lidarhd/items"
+_HEADERS = {"User-Agent": "Mozilla/5.0 (lidar-archeo-pipeline)"}
+
+
+def parse_tile_specs(args):
+    """Convertit des spécifications "col,row" ou "col:row" en liste de tuples.
+
+    Args:
+        args: liste de chaînes (ex: ["1055,6882", "1056:6883"]).
+
+    Returns:
+        Liste de tuples (col, row).
+
+    Raises:
+        ValueError: si une spécification est mal formée.
+    """
+    specs = []
+    for raw in args:
+        text = raw.strip().replace(":", ",").replace(";", ",")
+        parts = [p.strip() for p in text.split(",") if p.strip()]
+        if len(parts) != 2:
+            raise ValueError(f"Spécification de tuile invalide: {raw!r} (attendu: col,row)")
+        try:
+            col, row = int(parts[0]), int(parts[1])
+        except ValueError:
+            raise ValueError(f"Spécification de tuile invalide: {raw!r} (col et row doivent être des entiers)")
+        specs.append((col, row))
+    return specs
+
+
+def tile_filename(col, row):
+    """Nom de fichier LAZ standard d'une dalle (col, row)."""
+    return f"LHD_FXX_{col:04d}_{row:04d}_PTS_LAMB93_IGN69.copc.laz"
+
+
+def _bbox_wgs84(col, row):
+    """Bbox WGS84 de la dalle (col,row) pour la requête STAC (peut être élargie)."""
+    try:
+        from rasterio.warp import transform as warp_transform
+        xs = [col * 1000, (col + 1) * 1000, col * 1000, (col + 1) * 1000]
+        ys = [(row - 1) * 1000] * 2 + [row * 1000] * 2
+        lons, lats = warp_transform('EPSG:2154', 'EPSG:4326', xs, ys)
+    except Exception:
+        from .index import _approx_l93_to_wgs84
+        pts = [_approx_l93_to_wgs84(x, y)
+               for x in (col * 1000, (col + 1) * 1000)
+               for y in ((row - 1) * 1000, row * 1000)]
+        lons = [p[0] for p in pts]
+        lats = [p[1] for p in pts]
+    pad = 0.005  # ~500 m de marge pour éviter les erreurs d'arrondi aux bords
+    return (min(lons) - pad, min(lats) - pad, max(lons) + pad, max(lats) + pad)
+
+
+def match_feature(features, col, row):
+    """Retourne l'item STAC correspondant à la dalle (col,row), sinon None."""
+    want = f"{col:04d}-{row:04d}"
+    for feature in features:
+        props = feature.get("properties", {})
+        if props.get("lidarhd:coordonnees_NW") == want:
+            return feature
+    return None
+
+
+def find_tile_url(col, row, timeout=20):
+    """Cherche l'URL de téléchargement de la dalle (col,row) dans le catalogue STAC.
+
+    Returns:
+        URL (str) ou None si la dalle n'est pas (encore) publiée par l'IGN.
+    """
+    w, s, e, n = _bbox_wgs84(col, row)
+    query = urllib.parse.urlencode({"bbox": f"{w:.6f},{s:.6f},{e:.6f},{n:.6f}", "limit": 50})
+    req = urllib.request.Request(f"{_STAC_ITEMS_URL}?{query}", headers=_HEADERS)
+    with urllib.request.urlopen(req, timeout=timeout) as response:
+        data = json.loads(response.read().decode("utf-8"))
+    feature = match_feature(data.get("features", []), col, row)
+    if not feature:
+        return None
+    return feature.get("assets", {}).get("data", {}).get("href")
+
+
+def download_file(url, dest_path, timeout=120, chunk=1024 * 1024):
+    """Télécharge url vers dest_path en streaming. Retourne la taille en octets."""
+    req = urllib.request.Request(url, headers=_HEADERS)
+    t0 = time.time()
+    with urllib.request.urlopen(req, timeout=timeout) as response, open(dest_path, "wb") as out:
+        done = 0
+        while True:
+            block = response.read(chunk)
+            if not block:
+                break
+            out.write(block)
+            done += len(block)
+        elapsed = time.time() - t0
+        logger.info(f"    {done / 1e6:.0f} Mo en {elapsed:.0f}s"
+                    f" ({done / 1e6 / max(elapsed, 0.1):.1f} Mo/s)")
+    return done
+
+
+def fetch_tiles(input_dir, specs, output_dir=None):
+    """Télécharge les dalles IGN spécifiées, sauf celles déjà présentes/générées.
+
+    Args:
+        input_dir: dossier des fichiers LAZ (écriture autorisée requise).
+        specs: liste de tuples (col, row).
+        output_dir: dossier de sortie (optionnel) — permet d'ignorer les
+                    tuiles dont les visualisations existent déjà.
+
+    Returns:
+        Liste des chemins téléchargés.
+    """
+    input_dir = Path(input_dir)
+    downloaded = []
+    for col, row in specs:
+        name = tile_filename(col, row)
+        dest = input_dir / name
+        if dest.exists():
+            logger.info(f"  {name} : déjà présent dans input/ — aucun téléchargement")
+            continue
+        if output_dir is not None:
+            vis_dir = Path(output_dir) / "visualisations"
+            if list(vis_dir.glob(f"LHD_FXX_{col:04d}_{row:04d}_PTS*")):
+                logger.info(f"  {name} : visualisations déjà générées — ignorée")
+                continue
+        logger.info(f"  {name} : recherche dans le catalogue IGN...")
+        try:
+            url = find_tile_url(col, row)
+        except Exception as e:
+            logger.warning(f"  ✗ {name} : erreur catalogue ({e})")
+            continue
+        if not url:
+            logger.warning(f"  ✗ {name} : introuvable dans le catalogue IGN (zone non publiée ?)")
+            continue
+        logger.info(f"  {name} : téléchargement depuis la géoplateforme...")
+        try:
+            download_file(url, dest)
+            logger.info(f"  ✓ {name} téléchargée")
+            downloaded.append(dest)
+        except Exception as e:
+            dest.unlink(missing_ok=True)
+            logger.warning(f"  ✗ {name} : échec du téléchargement ({e})")
+    return downloaded

lidar_pipeline/pipeline.py

@@ -109,7 +109,7 @@ VIZ_STEPS = [
 class LidarArchaeoPipeline:
     """Orchestrates the LiDAR archaeological analysis pipeline."""
 
-    def __init__(self, input_dir, output_dir, resolution=0.5, workers=1, force=False, ground_method='auto', force_classify=False, keep_tif=False, quality=98, only_viz=None, skip_viz=None, output_format='avif', gpu_ids=None, no_index=False):
+    def __init__(self, input_dir, output_dir, resolution=0.5, workers=1, force=False, ground_method='auto', ign_classes="sol", force_classify=False, keep_tif=False, bare_earth=False, quality=98, only_viz=None, skip_viz=None, output_format='avif', gpu_ids=None, no_index=False):
         self.input_dir = Path(input_dir)
         self.output_dir = Path(output_dir)
         # Accept single float or comma-separated string for multi-resolution
@@ -123,7 +123,9 @@ class LidarArchaeoPipeline:
         self.workers = workers
         self.force = force
         self.ground_method = ground_method
+        self.ign_classes = ign_classes
         self.force_classify = force_classify
+        self.bare_earth = bare_earth
         self.keep_tif = keep_tif
         self.quality = quality
         self.only_viz = only_viz
@@ -212,7 +214,7 @@ class LidarArchaeoPipeline:
         else:
             return file_vis_dir / f"{basename}_{name}.{ext}"
 
-    def generate_all_visualizations(self, dtm_file, basename, resolution=None, vis_dir=None):
+    def generate_all_visualizations(self, dtm_file, basename, resolution=None, vis_dir=None, force_images=None):
         """Generate all archaeological visualizations for one DTM file.
 
         Optimisation: SharedDEM is only computed if at least one visualization
@@ -229,9 +231,10 @@ class LidarArchaeoPipeline:
         total = len(self.viz_steps)
 
         # Phase 1: determine which visualizations need generation
+        force_viz = self.force if force_images is None else force_images
         needs_generation = {}  # name -> True/False
         for name, func in self.viz_steps:
-            if self.force:
+            if force_viz:
                 needs_generation[name] = True
             else:
                 expected_webp = self._expected_output_path(name, basename, file_vis_dir, self.output_format)
@@ -265,8 +268,8 @@ class LidarArchaeoPipeline:
                 vis_results[name] = self._expected_output_path(name, basename, file_vis_dir, self.output_format)
                 continue
 
-            # When --force, delete existing TIF to ensure clean regeneration
+            # When regenerating, delete existing TIF to ensure clean regeneration
-            if self.force:
+            if force_viz:
                 for tif in file_vis_dir.glob(f"{basename}_{name}.tif"):
                     tif.unlink(missing_ok=True)
                 if name == 'pos_open':
@@ -325,6 +328,48 @@ class LidarArchaeoPipeline:
         res_str = f"{resolution}".replace('.', 'p')
         return f"_r{res_str}"
 
+    def _dtm_method_path(self, basename, res_suffix):
+        """Sidecar path storing which ground method produced a DTM."""
+        return self.dtm_dir / f"{basename}_dtm{res_suffix}_method.txt"
+
+    def _dtm_method_name(self, basename, res_suffix):
+        """Read the recorded ground method for a DTM, or None if unknown."""
+        p = self._dtm_method_path(basename, res_suffix)
+        if p.exists():
+            try:
+                return p.read_text(encoding="utf-8").strip() or None
+            except Exception:
+                return None
+        return None
+
+    def _effective_ground_method(self):
+        """Méthode effective pour le suivi de cache, classes IGN incluses.
+
+        La méthode 'ign' est étiquetée avec les classes choisies (ex. 'ign_1_2')
+        pour qu'un changement de --ign-classes déclenche la reclassification.
+        """
+        if self.ground_method == 'ign':
+            from .dtm import parse_ign_classes, ign_method_label
+            return ign_method_label(parse_ign_classes(self.ign_classes))
+        return self.ground_method
+
+    def _dtm_method_matches(self, basename, res_suffix):
+        """True if the recorded ground method matches the requested one.
+
+        A DTM without a recorded method is treated as matching so the existing
+        cache is preserved; the method is adopted on its next reclassification.
+        """
+        recorded = self._dtm_method_name(basename, res_suffix)
+        return recorded is None or recorded == self._effective_ground_method()
+
+    def _write_dtm_method(self, basename, res_suffix):
+        """Record the ground classification method used to build a DTM."""
+        try:
+            self._dtm_method_path(basename, res_suffix).write_text(
+                self._effective_ground_method(), encoding="utf-8")
+        except Exception:
+            pass
+
     def process_file(self, laz_file):
         """Process a single LAZ file through the full pipeline.
 
@@ -349,10 +394,18 @@ class LidarArchaeoPipeline:
         # Step 1: Ground classification (shared across all resolutions)
         las_file = None
         t_classif = 0
+        dtm_rebuilt = False
+        # The ground method is shared across resolutions. It is recorded per DTM
+        # in a sidecar so that changing --ground-classification invalidates the
+        # cache; otherwise the cached DTM would be reused and the new method
+        # never applied (nothing would change visually).
+        primary_suffix = self._res_suffix(self.resolutions[0])
+        method_matches = self._dtm_method_matches(basename, primary_suffix)
         for i, res in enumerate(self.resolutions):
             res_suffix = self._res_suffix(res)
             dtm_path = self.dtm_dir / f"{basename}_dtm{res_suffix}.tif"
-            if dtm_path.exists():
+            if dtm_path.exists() and not self.force_classify:
+                if method_matches:
                     import rasterio
                     try:
                         with rasterio.open(dtm_path) as src:
@@ -370,13 +423,16 @@ class LidarArchaeoPipeline:
                     except Exception:
                         logger.warning(f"Impossible de lire le DTM existant — régénération")
                         dtm_path.unlink()
+                else:
+                    logger.info(f"  DTM{res_suffix} produit par {self._dtm_method_name(basename, primary_suffix) or '?'} ≠ {self._effective_ground_method()} → reclassification")
+                    dtm_path.unlink()
 
             # Need to classify/generate DTM for this resolution
             if las_file is None:
                 # First time: do ground classification
                 logger.info("[1/5] Classification du sol...")
                 t1 = time.time()
-                las_file = classify_ground(laz_file, self.temp_dir, method=self.ground_method, force=self.force_classify)
+                las_file = classify_ground(laz_file, self.temp_dir, method=self.ground_method, force=self.force_classify, ign_classes=self.ign_classes)
                 t_classif = time.time() - t1
                 if not las_file:
                     logger.error(f"  ✗ Échec classification ({t_classif:.1f}s)")
@@ -386,7 +442,15 @@ class LidarArchaeoPipeline:
             # Generate DTM at this resolution
             logger.info(f"{'[2/5]' if i == 0 else '  '} Génération DTM {res}m/px...")
             t2 = time.time()
-            dtm_file = create_dtm_fast(las_file, basename, self.dtm_dir, res, force=self.force, output_suffix=res_suffix)
+            # Classification IGN → mode pur : DTM = rasterisation brute des
+            # classes choisies, sans plancher ni comblement (sauf --bare-earth)
+            pure_ign = "_ground_ign" in Path(las_file).name
+            dtm_file = create_dtm_fast(las_file, basename, self.dtm_dir, res,
+                                       force=self.force or self.force_classify,
+                                       output_suffix=res_suffix,
+                                       source_laz=laz_file,
+                                       bare_earth=self.bare_earth,
+                                       pure=pure_ign)
             t_dtm = time.time() - t2
             if not dtm_file:
                 logger.error(f"  ✗ Échec DTM {res}m/px ({t_dtm:.1f}s)")
@@ -394,6 +458,8 @@ class LidarArchaeoPipeline:
                     return False  # Primary resolution failure is fatal
                 continue  # Additional resolution failure is non-fatal
             logger.info(f"  ✓ DTM {res}m/px terminé ({t_dtm:.1f}s)")
+            dtm_rebuilt = True
+            self._write_dtm_method(basename, res_suffix)
 
         # Process each resolution: visualizations + PDF
         all_vis_results = {}
@@ -420,16 +486,18 @@ class LidarArchaeoPipeline:
 
             vis_dir.mkdir(exist_ok=True)
 
-            self.generate_all_visualizations(dtm_path, basename, actual_res, vis_dir=vis_dir)
+            self.generate_all_visualizations(
+                dtm_path, basename, actual_res, vis_dir=vis_dir,
+                force_images=self.force or self.force_classify or dtm_rebuilt)
 
         t_total = time.time() - t_start
         logger.info(f"✓ {basename} terminé en {t_total:.1f}s")
         _file_filter.basename = None
         return True
 
-    def process_all(self):
+    def process_all(self, files=None):
-        """Process all LAZ files in input directory."""
+        """Process all LAZ files in input directory (or an explicit list)."""
-        files = self.find_laz_files()
+        files = files if files is not None else self.find_laz_files()
 
         if not files:
             logger.error("Aucun fichier LAZ/LAS trouvé !")
@@ -466,7 +534,7 @@ class LidarArchaeoPipeline:
                 active_ids = self.gpu_ids if self.gpu_ids else available_gpu_ids()
                 resolutions_str = ','.join(str(r) for r in self.resolutions)
                 future_to_file = {
-                    executor.submit(_process_file_standalone, str(laz_file), str(self.input_dir), str(self.output_dir), resolutions_str, self.force, self.ground_method, self.force_classify, self.keep_tif, self.quality, self.only_viz, self.skip_viz, self.output_format, active_ids[file_idx % len(active_ids)] if active_ids else None): laz_file
+                    executor.submit(_process_file_standalone, str(laz_file), str(self.input_dir), str(self.output_dir), resolutions_str, self.force, self.ground_method, self.ign_classes, self.force_classify, self.keep_tif, self.bare_earth, self.quality, self.only_viz, self.skip_viz, self.output_format, active_ids[file_idx % len(active_ids)] if active_ids else None): laz_file
                     for file_idx, laz_file in enumerate(files)
                 }
                 done = 0
@@ -545,7 +613,7 @@ class LidarArchaeoPipeline:
             logger.warning(f"  Note: Impossible de supprimer les fichiers temporaires: {e}")
 
 
-def _process_file_standalone(laz_file_str, input_dir, output_dir, resolution, force=False, ground_method='auto', force_classify=False, keep_tif=False, quality=98, only_viz=None, skip_viz=None, output_format='avif', gpu_id=None):
+def _process_file_standalone(laz_file_str, input_dir, output_dir, resolution, force=False, ground_method='auto', ign_classes="sol", force_classify=False, keep_tif=False, bare_earth=False, quality=98, only_viz=None, skip_viz=None, output_format='avif', gpu_id=None):
     """Standalone function for multiprocessing — creates its own pipeline instance.
 
     Each worker gets its own temp directory to avoid file conflicts.
@@ -572,7 +640,7 @@ def _process_file_standalone(laz_file_str, input_dir, output_dir, resolution, fo
         worker_logger.addHandler(handler)
     worker_logger.addFilter(_file_filter)
 
-    pipeline = LidarArchaeoPipeline(input_dir, output_dir, resolution=resolution, workers=1, force=force, ground_method=ground_method, force_classify=force_classify, keep_tif=keep_tif, quality=quality, only_viz=only_viz, skip_viz=skip_viz, output_format=output_format)
+    pipeline = LidarArchaeoPipeline(input_dir, output_dir, resolution=resolution, workers=1, force=force, ground_method=ground_method, ign_classes=ign_classes, force_classify=force_classify, keep_tif=keep_tif, bare_earth=bare_earth, quality=quality, only_viz=only_viz, skip_viz=skip_viz, output_format=output_format)
     basename = _file_basename(laz_file_str)
     pipeline.temp_dir = pipeline.output_dir / "temp" / basename
     pipeline.temp_dir.mkdir(exist_ok=True)

lidar_pipeline/rendering.py

@@ -818,8 +818,11 @@ def tif_to_crop(tif_file, vis_dir, resolution, keep_tif=False, quality=98, outpu
 
         # Convert to RGB using colormap
         if is_rgb_result:
-            # RGB images are already in RGB
+            # RGB images are already in RGB (uint8 depuis le TIF IGN, ou float 0-1)
-            rgb_data = (data * 255).astype(np.uint8)
+            if data.dtype == np.uint8:
+                rgb_data = data
+            else:
+                rgb_data = (np.clip(data, 0, 1) * 255).astype(np.uint8)
         else:
             # Normalize data to 0-1 range for colormap
             cmap = plt.get_cmap(cmap_name)

lidar_pipeline/tests/test_cli.py

@@ -86,3 +86,20 @@ class TestSetupLogging:
         fmt = logger.handlers[0].formatter._fmt
         assert "%(filename)s" in fmt
         assert "%(lineno)d" in fmt
+
+
+def test_rebuild_index_without_input_arg(tmp_path):
+    """--rebuild-index fonctionne sans l'argument positionnel input.
+
+    Régression : input était obligatoire alors que --rebuild-index ne
+    l'utilise pas (erreur argparse « the following arguments are required »).
+    """
+    import subprocess
+
+    r = subprocess.run(
+        [sys.executable, "-m", "lidar_pipeline", "--rebuild-index",
+         "-o", str(tmp_path)],
+        capture_output=True, text=True, timeout=180,
+    )
+    assert r.returncode == 0, r.stderr
+    assert "the following arguments are required" not in r.stderr

lidar_pipeline/tests/test_dtm.py

@@ -94,12 +94,127 @@ class TestCSFPipeline:
         pipeline = json.loads(result)
 
         csf_stage = [s for s in pipeline["pipeline"] if isinstance(s, dict) and s.get("type") == "filters.csf"][0]
-        assert csf_stage["resolution"] == 0.5
+        assert csf_stage["resolution"] == 1.0  # cloth 1 m : ~4× plus rapide, MNT inchangé
         assert csf_stage["rigidness"] == 3
         assert csf_stage["smooth"] is True
         assert "hdiff" not in csf_stage  # hdiff is not a valid PDAL CSF parameter
 
 
+class TestInterpolateHoles:
+    def test_fills_interior_hole_with_surface(self):
+        """Large interior NaN hole is filled (no NaN left, value is plausible)."""
+        from lidar_pipeline.dtm import _interpolate_holes
+        # Linear-in-column surface z = 0.02 * x, with a large square hole in the middle.
+        x = np.arange(40, dtype=float) * 0.02
+        dtm = np.tile(x, (40, 1))
+        dtm[16:24, 16:24] = np.nan
+        filled, count = _interpolate_holes(dtm)
+        assert count == 64
+        assert not np.isnan(filled).any()
+        # Filled values stay within the surrounding z range (no wild extrapolation).
+        zmin, zmax = np.nanmin(dtm), np.nanmax(dtm)
+        hole_vals = filled[16:24, 16:24]
+        assert np.all(hole_vals >= zmin - 1e-6)
+        assert np.all(hole_vals <= zmax + 1e-6)
+        # A linear surface is interpolated near-exactly in the interior.
+        expected = np.tile(x[16:24], (8, 1))
+        assert np.allclose(hole_vals, expected, atol=0.02)
+        # Original valid cells are untouched.
+        valid = ~np.isnan(dtm)
+        assert np.allclose(filled[valid], dtm[valid])
+
+    def test_no_holes_returns_unchanged(self):
+        """No NaN → returns same array and zero count."""
+        from lidar_pipeline.dtm import _interpolate_holes
+        dtm = np.arange(64, dtype=float).reshape(8, 8)
+        filled, count = _interpolate_holes(dtm)
+        assert count == 0
+        assert np.shares_memory(filled, dtm)
+
+    def test_all_nan_returns_unchanged(self):
+        """No valid data → cannot interpolate, returns zeros-free NaN array."""
+        from lidar_pipeline.dtm import _interpolate_holes
+        dtm = np.full((8, 8), np.nan)
+        filled, count = _interpolate_holes(dtm)
+        assert count == 0
+        assert np.isnan(filled).all()
+
+
+class TestMinReturnGrid:
+    def test_takes_lowest_return_per_cell(self, tmp_output_dir):
+        """_min_return_grid rasterise le point le plus bas par cellule (pas la moyenne)."""
+        import laspy
+        from lidar_pipeline.dtm import _min_return_grid
+        out = tmp_output_dir / "pts.las"
+        hdr = laspy.LasHeader(version='1.2', point_format=0)
+        las = laspy.LasData(hdr)
+        # Grille 2x2 sur [0,2]x[0,2]. La cellule (0,0) porte deux points z=5 et
+        # z=2 (min=2, moyenne=3.5) ; (1,0) z=3 ; (0,1) z=4 ; (1,1) vide.
+        las.x = [0.2, 0.5, 1.2, 0.3]
+        las.y = [0.2, 0.3, 0.4, 1.5]
+        las.z = [5.0, 2.0, 3.0, 4.0]
+        las.write(str(out))
+
+        grid = _min_return_grid(out, 2, 2, (0.0, 0.0, 2.0, 2.0))
+        assert grid.shape == (2, 2)
+        assert int(np.isnan(grid).sum()) == 1
+        # Une seule valeur par cellule, et la cellule (0,0) vaut le MIN (2.0).
+        vals = sorted(float(v) for v in grid[~np.isnan(grid)])
+        assert vals == [2.0, 3.0, 4.0]
+        assert 3.5 not in vals
+
+
+class TestBareEarth:
+    """Le plancher « sol nu » ramène le DTM au retour le plus bas de chaque cellule."""
+
+    def _write_las(self, path, points):
+        """points: list of (x, y, z). Écrit un LAS 1.2 format 0 aux bornes [0,2]x[0,2]."""
+        import laspy
+        hdr = laspy.LasHeader(version='1.2', point_format=0)
+        las = laspy.LasData(hdr)
+        las.x = [p[0] for p in points]
+        las.y = [p[1] for p in points]
+        las.z = [p[2] for p in points]
+        las.write(str(path))
+        return path
+
+    def _make_clouds(self, tmp_output_dir):
+        """Grille 2x2 (res=1.0). Les points « coin » à 0.05/1.95 imposent l'étendue
+        [0.05,1.95] (laspy re-déduit les bornes de l'en-tête depuis les points).
+        Le sol (las_file) vaut z=10 partout. Le nuage complet (source_laz) a un
+        retour plus bas dans les cellules (0,0) -> 2 et (1,0) -> 5 ; les deux autres
+        cellules n'ont que z=10."""
+        ground = [(0.5, 0.5, 10.0), (1.5, 0.5, 10.0), (0.5, 1.5, 10.0), (1.5, 1.5, 10.0),
+                  (0.05, 0.05, 10.0), (1.95, 1.95, 10.0)]
+        source = list(ground) + [(0.3, 0.3, 2.0), (1.3, 0.3, 5.0)]
+        las_file = self._write_las(tmp_output_dir / "ground.las", ground)
+        source_laz = self._write_las(tmp_output_dir / "source.las", source)
+        return las_file, source_laz
+
+    def _dtm_values(self, tmp_output_dir, bare_earth):
+        from lidar_pipeline.dtm import create_dtm_fast
+        import rasterio
+        las_file, source_laz = self._make_clouds(tmp_output_dir)
+        out = create_dtm_fast(las_file, "tile", tmp_output_dir, 1.0,
+                              force=True, source_laz=source_laz, bare_earth=bare_earth)
+        assert out is not None
+        with rasterio.open(str(out)) as src:
+            arr = src.read(1).astype("float64")
+        return arr[~np.isnan(arr)]
+
+    def test_bare_earth_pulls_dtm_to_lowest_return(self, tmp_output_dir):
+        """Avec bare_earth, le DTM descend aux retours les plus bas (2 et 5)."""
+        vals = sorted(float(v) for v in self._dtm_values(tmp_output_dir, bare_earth=True))
+        # Les cellules sans retour plus bas restent à 10 ; les deux autres descendent.
+        assert vals == [2.0, 5.0, 10.0, 10.0]
+        assert vals[0] == 2.0 and vals[1] == 5.0
+
+    def test_no_bare_earth_keeps_mean(self, tmp_output_dir):
+        """Sans bare_earth, le DTM garde la moyenne des points sol (10 partout)."""
+        vals = sorted(float(v) for v in self._dtm_values(tmp_output_dir, bare_earth=False))
+        assert all(v == 10.0 for v in vals)
+
+
 class TestDetectGroundMethod:
     def _make_mock_las(self, num_returns, z_values):
         """Create a mock laspy object with specified NumberOfReturns and z."""
@@ -162,6 +277,73 @@ class TestDetectGroundMethod:
         assert result == 'csf'
 
 
+class TestIGNPipeline:
+    def test_pipeline_keeps_supplier_classification(self):
+        """create_ign_pipeline réutilise la pré-classification (classe 2) sans refiltrer."""
+        from lidar_pipeline.dtm import create_ign_pipeline
+        result = create_ign_pipeline("/input/a.laz", "/output/a_ground.las")
+        pipeline = json.loads(result)
+
+        stages = pipeline["pipeline"]
+        stage_types = [s.get("type") if isinstance(s, dict) else None for s in stages]
+
+        # Aucun algorithme de classification, pas de remise à zéro, pas de filtres de bruit
+        assert "filters.smrf" not in stage_types
+        assert "filters.csf" not in stage_types
+        assert "filters.assign" not in stage_types
+        assert "filters.elm" not in stage_types
+        assert "filters.outlier" not in stage_types
+
+        # Filtre ReturnNumber conservé + extraction des points classe 2
+        range_stages = [s for s in stages if isinstance(s, dict) and s.get("type") == "filters.range"]
+        assert any("ReturnNumber" in str(s.get("limits", "")) for s in range_stages)
+        assert any(s.get("limits") == "Classification[2:2]" for s in range_stages)
+
+        writer = [s for s in stages if isinstance(s, dict) and s.get("type") == "writers.las"][0]
+        assert writer["filename"] == "/output/a_ground.las"
+
+
+class TestDetectIGN:
+    def _make_mock_las(self, classification, num_returns, z_values):
+        mock_las = MagicMock()
+        mock_las.classification = classification
+        mock_las.NumberOfReturns = np.array(num_returns)
+        mock_las.z = np.array(z_values)
+        mock_las.points = MagicMock()
+        mock_las.points.__len__ = lambda self: len(num_returns)
+        return mock_las
+
+    @patch('lidar_pipeline.dtm._read_with_pdal')
+    @patch('laspy.read')
+    def test_preclassified_returns_ign(self, mock_read, mock_pdal):
+        """Fichier pré-classifié (majorité classe 2) → méthode IGN."""
+        from lidar_pipeline.dtm import detect_ground_method
+
+        n = 10000
+        num_returns = np.ones(n, dtype=int)
+        cls = np.zeros(n, dtype=np.uint8)
+        cls[int(n * 0.15):] = 2  # 85 % de points classe 2
+        z_values = np.random.normal(100, 5, n)
+
+        mock_read.return_value = self._make_mock_las(cls, num_returns, z_values)
+        assert detect_ground_method(Path("/data/input/test.laz")) == 'ign'
+
+    @patch('lidar_pipeline.dtm._read_with_pdal')
+    @patch('laspy.read')
+    def test_unclassified_falls_back_to_smrf_or_csf(self, mock_read, mock_pdal):
+        """Sans classification exploitable → détection SMRF/CSF classique."""
+        from lidar_pipeline.dtm import detect_ground_method
+
+        n = 10000
+        num_returns = np.ones(n, dtype=int)
+        num_returns[:int(n * 0.6)] = 2  # 60 % multi-retours (forêt) → non urbain
+        cls = np.zeros(n, dtype=np.uint8)  # aucun point classe 2
+        z_values = np.random.normal(100, 5, n)
+
+        mock_read.return_value = self._make_mock_las(cls, num_returns, z_values)
+        assert detect_ground_method(Path("/data/input/test.laz")) == 'smrf'
+
+
 class TestClassifyGroundMethod:
     @patch('lidar_pipeline.dtm.subprocess')
     def test_classify_ground_auto_calls_detect(self, mock_subprocess):
@@ -212,3 +394,157 @@ class TestClassifyGroundMethod:
                     pipeline = json.loads(pipeline_file.read_text())
                     stage_types = [s.get("type") if isinstance(s, dict) else None for s in pipeline["pipeline"]]
                     assert "filters.csf" in stage_types
+
+class TestParseIgnClasses:
+    def test_default_sol(self):
+        """'sol' → code 2 seul."""
+        from lidar_pipeline.dtm import parse_ign_classes
+        assert parse_ign_classes("sol") == [2]
+
+    def test_names_sorted_dedup(self):
+        """Noms acceptés (EN/FR), triés et dédupliqués."""
+        from lidar_pipeline.dtm import parse_ign_classes
+        assert parse_ign_classes("sol,unclassified") == [1, 2]
+        assert parse_ign_classes("unclassified,sol") == [1, 2]
+        assert parse_ign_classes("non-classe") == [1]
+        assert parse_ign_classes("sol,2") == [2]
+
+    def test_numeric_codes(self):
+        """Codes LAS directs, triés."""
+        from lidar_pipeline.dtm import parse_ign_classes
+        assert parse_ign_classes("2,1") == [1, 2]
+        assert parse_ign_classes("66") == [66]
+
+    def test_invalid_raises(self):
+        """Nom inconnu, code hors bornes ou liste vide → ValueError."""
+        from lidar_pipeline.dtm import parse_ign_classes
+        with pytest.raises(ValueError):
+            parse_ign_classes("foo")
+        with pytest.raises(ValueError):
+            parse_ign_classes("300")
+        with pytest.raises(ValueError):
+            parse_ign_classes("")
+
+    def test_method_label(self):
+        """'ign' seul pour le sol, combinaison encodée sinon (invalidation cache)."""
+        from lidar_pipeline.dtm import ign_method_label
+        assert ign_method_label([2]) == "ign"
+        assert ign_method_label([1, 2]) == "ign_1_2"
+        assert ign_method_label([2, 1]) == "ign_1_2"
+
+
+class TestIGNPipelineMultiClasses:
+    def test_multi_class_limits(self):
+        """Plusieurs classes → plages OU logiques sur Classification."""
+        from lidar_pipeline.dtm import _create_ground_pipeline
+        result = _create_ground_pipeline("/input/a.laz", "/output/a_ground.las",
+                                         'ign', ign_codes=[1, 2])
+        pipeline = json.loads(result)
+        range_stages = [s for s in pipeline["pipeline"]
+                        if isinstance(s, dict) and s.get("type") == "filters.range"]
+        limits = [str(s.get("limits", "")) for s in range_stages]
+        assert any("Classification[1:1]" in l and "Classification[2:2]" in l
+                   for l in limits)
+
+    def test_default_sol_only(self):
+        """Sans ign_codes, la voie IGN reste sol seul (2) — rétrocompatible."""
+        from lidar_pipeline.dtm import _create_ground_pipeline
+        result = _create_ground_pipeline("/input/a.laz", "/output/a_ground.las", 'ign')
+        pipeline = json.loads(result)
+        range_stages = [s for s in pipeline["pipeline"]
+                        if isinstance(s, dict) and s.get("type") == "filters.range"]
+        limits = [str(s.get("limits", "")) for s in range_stages]
+        assert any("Classification[2:2]" in l and "Classification[1:1]" not in l
+                   for l in limits)
+
+
+class TestClassifyGroundIgnClasses:
+    @patch('lidar_pipeline.dtm.subprocess')
+    def test_ign_classes_encoded_in_filenames(self, mock_subprocess):
+        """--ign-classes sol,unclassified → fichiers ign_1_2 + filtre multi-classes."""
+        import tempfile
+        from lidar_pipeline.dtm import classify_ground
+
+        mock_subprocess.run.return_value = MagicMock(returncode=0)
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tmpdir = Path(tmpdir)
+            classify_ground(Path("/data/input/test.laz"), tmpdir,
+                            method='ign', ign_classes="sol,unclassified")
+
+            pipeline_file = tmpdir / "pipeline_ign_1_2.json"
+            assert pipeline_file.exists()
+            pipeline = json.loads(pipeline_file.read_text())
+            limits = [str(s.get("limits", "")) for s in pipeline["pipeline"]
+                      if isinstance(s, dict) and s.get("type") == "filters.range"]
+            assert any("Classification[1:1]" in l and "Classification[2:2]" in l
+                       for l in limits)
+
+    @patch('lidar_pipeline.dtm.subprocess')
+    def test_ign_default_label_unchanged(self, mock_subprocess):
+        """--ign-classes sol (défaut) → noms 'ign' inchangés (cache préservé)."""
+        import tempfile
+        from lidar_pipeline.dtm import classify_ground
+
+        mock_subprocess.run.return_value = MagicMock(returncode=0)
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tmpdir = Path(tmpdir)
+            classify_ground(Path("/data/input/test.laz"), tmpdir, method='ign')
+
+            assert (tmpdir / "pipeline_ign.json").exists()
+            assert not (tmpdir / "pipeline_ign_1_2.json").exists()
+
+
+class TestPureDtm:
+    """Mode pur (classification IGN) : aucune retouche, trous en nodata."""
+
+    def _write_las(self, path, points):
+        import laspy
+        hdr = laspy.LasHeader(version='1.2', point_format=0)
+        las = laspy.LasData(hdr)
+        las.x = [p[0] for p in points]
+        las.y = [p[1] for p in points]
+        las.z = [p[2] for p in points]
+        las.write(str(path))
+        return path
+
+    def _make_clouds(self, tmp_output_dir):
+        """Grille 2x2 (res=1.0). Sol sur 3 cellules (z=10), trou en (1,1).
+        Le nuage complet a un retour plus bas (z=7) dans le trou."""
+        corners = [(0.05, 0.05, 10.0), (1.95, 0.05, 10.0), (0.05, 1.95, 10.0)]
+        ground = [(0.5, 0.5, 10.0), (1.5, 0.5, 10.0), (0.5, 1.5, 10.0)] + corners
+        source = list(ground) + [(1.5, 1.5, 7.0)]
+        las_file = self._write_las(tmp_output_dir / "ground_pure.las", ground)
+        source_laz = self._write_las(tmp_output_dir / "source_pure.las", source)
+        return las_file, source_laz
+
+    def _dtm_array(self, tmp_output_dir, pure):
+        from lidar_pipeline.dtm import create_dtm_fast
+        import rasterio
+        las_file, source_laz = self._make_clouds(tmp_output_dir)
+        out = create_dtm_fast(las_file, "tile_pure", tmp_output_dir, 1.0,
+                              force=True, source_laz=source_laz, pure=pure)
+        assert out is not None
+        with rasterio.open(str(out)) as src:
+            return src.read(1).astype("float64")
+
+    def test_pure_fills_holes_without_floor(self, tmp_output_dir):
+        """pur=True : trous comblés par interpolation, sans plancher à 7.
+
+        Le comblement est actif dans tous les modes (comportement
+        historique) ; « pur » ne désactive que l'abaissement au retour
+        le plus bas.
+        """
+        arr = self._dtm_array(tmp_output_dir, pure=True)
+        assert int(np.isnan(arr).sum()) == 0
+        vals = sorted(float(v) for v in arr.flatten())
+        assert vals == [10.0, 10.0, 10.0, 10.0]
+
+    def test_not_pure_fills_holes(self, tmp_output_dir):
+        """pur=False : le trou est comblé (comportement historique conservé)."""
+        arr = self._dtm_array(tmp_output_dir, pure=False)
+        assert int(np.isnan(arr).sum()) == 0
+        vals = sorted(float(v) for v in arr.flatten())
+        assert len(vals) == 4
+        assert vals[-1] == 10.0

lidar_pipeline/tests/test_fetch_ign.py

@@ -0,0 +1,98 @@
+"""Tests du téléchargement des dalles LiDAR HD de l'IGN (fetch_ign)."""
+
+
+def test_parse_tile_specs():
+    """Accepte 'col,row', 'col:row' et ignore les espaces."""
+    from lidar_pipeline.fetch_ign import parse_tile_specs
+    assert parse_tile_specs(["1055,6882"]) == [(1055, 6882)]
+    assert parse_tile_specs(["1055:6883", " 651 , 6630 "]) == [(1055, 6883), (651, 6630)]
+
+
+def test_parse_tile_specs_rejects_invalid():
+    """Les spécifications mal formées lèvent une erreur explicite."""
+    import pytest
+    from lidar_pipeline.fetch_ign import parse_tile_specs
+    with pytest.raises(ValueError):
+        parse_tile_specs(["1055"])
+    with pytest.raises(ValueError):
+        parse_tile_specs(["abc,def"])
+    with pytest.raises(ValueError):
+        parse_tile_specs(["1055,6882,9999"])
+
+
+def test_tile_filename_pads_coordinates():
+    """Le nom de fichier DALLE utilise des coordonnées à 4 chiffres."""
+    from lidar_pipeline.fetch_ign import tile_filename
+    assert tile_filename(1055, 6882) == "LHD_FXX_1055_6882_PTS_LAMB93_IGN69.copc.laz"
+    assert tile_filename(651, 6630) == "LHD_FXX_0651_6630_PTS_LAMB93_IGN69.copc.laz"
+
+
+def test_match_feature_uses_coordonnees_nw():
+    """La correspondance se fait sur lidarhd:coordonnees_NW (format 0816-6847)."""
+    from lidar_pipeline.fetch_ign import match_feature
+    features = [
+        {"id": "LHD_FXX_1054_6882_PTS_LAMB93_IGN69_NE",
+         "properties": {"lidarhd:coordonnees_NW": "1054-6882"},
+         "assets": {"data": {"href": "https://example.org/a.copc.laz"}}},
+        {"id": "LHD_FXX_1055_6882_PTS_LAMB93_IGN69_NE",
+         "properties": {"lidarhd:coordonnees_NW": "1055-6882"},
+         "assets": {"data": {"href": "https://example.org/b.copc.laz"}}},
+    ]
+    assert match_feature(features, 1055, 6882) is features[1]
+    assert match_feature(features, 9999, 9999) is None
+
+
+def test_find_tile_url_matches_and_returns_href(monkeypatch):
+    """find_tile_url interroge le STAC et retourne l'href de l'asset data."""
+    from lidar_pipeline import fetch_ign
+
+    class FakeResponse:
+        def __init__(self, payload):
+            self._payload = payload.encode("utf-8")
+
+        def read(self):
+            return self._payload
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *args):
+            return False
+
+    payload = ('{"features": [{"properties": {"lidarhd:coordonnees_NW": "1055-6882"},'
+               '"assets": {"data": {"href": "https://data.geopf.fr/x.copc.laz"}}}]}')
+    captured = {}
+
+    def fake_urlopen(req, timeout=None):
+        captured["url"] = req.full_url
+        return FakeResponse(payload)
+
+    monkeypatch.setattr(fetch_ign.urllib.request, "urlopen", fake_urlopen)
+    url = fetch_ign.find_tile_url(1055, 6882)
+    assert url == "https://data.geopf.fr/x.copc.laz"
+    assert "api.stac.teledetection.fr" in captured["url"]
+    assert "bbox=" in captured["url"]
+
+
+def test_fetch_tiles_skips_existing_and_generated(tmp_path, monkeypatch):
+    """Pas de téléchargement si le LAZ existe ou si les visualisations existent."""
+    from lidar_pipeline import fetch_ign
+
+    input_dir = tmp_path / "input"
+    input_dir.mkdir()
+    output_dir = tmp_path / "output"
+    vis_dir = output_dir / "visualisations" / "LHD_FXX_1054_6882_PTS_LAMB93_IGN69"
+    vis_dir.mkdir(parents=True)
+
+    # 1054,6882 : visualisations déjà générées
+    # 1055,6882 : LAZ déjà présent
+    existing = input_dir / fetch_ign.tile_filename(1055, 6882)
+    existing.write_bytes(b"naze")
+
+    def fail_download(*args, **kwargs):
+        raise AssertionError("ne doit pas être appelé")
+
+    monkeypatch.setattr(fetch_ign, "find_tile_url", fail_download)
+    result = fetch_ign.fetch_tiles(input_dir, [(1054, 6882), (1055, 6882)],
+                                   output_dir=output_dir)
+    assert result == []

lidar_pipeline/tests/test_index.py

@@ -60,7 +60,11 @@ def test_compute_bbox_empty():
 
 
 def _make_fake_viz_dir(vis_dir, basename, col, row, viz_keys=('hillshade_multi', 'svf'), ext='webp', res_suffix=''):
-    """Crée un faux dossier de visualisations avec de petites images."""
+    """Crée un faux dossier de visualisations avec de petites images.
+
+    Le suffixe de résolution apparaît seulement dans le nom du dossier (miroir
+    du pipeline : les fichiers restent préfixés par le basename nu).
+    """
     from PIL import Image as PILImage
     import numpy as np
 
@@ -70,7 +74,7 @@ def _make_fake_viz_dir(vis_dir, basename, col, row, viz_keys=('hillshade_multi',
     for v in viz_keys:
         arr = np.random.randint(0, 255, (50, 50, 3), dtype=np.uint8)
         img = PILImage.fromarray(arr)
-        fname = f"LHD_FXX_{col}_{row}_PTS_LAMB93_IGN69{res_suffix}_{v}.{ext}"
+        fname = f"LHD_FXX_{col}_{row}_PTS_LAMB93_IGN69_{v}.{ext}"
         img.save(str(tile_dir / fname), format='WEBP', quality=80)
     return tile_dir
 
@@ -124,6 +128,85 @@ def test_scan_tiles_multi_resolution(tmp_path):
     assert resolutions == [0.2, 0.5]
 
 
+def test_res_suffix_str():
+    """Le suffixe de résolution reflète le nommage du pipeline (miroir)."""
+    from lidar_pipeline.index import _res_suffix_str
+    assert _res_suffix_str(0.5) == ''
+    assert _res_suffix_str(0.2) == '_r0p2'
+
+
+def test_collect_tile_metadata(tmp_path):
+    """Les métadonnées lisent la méthode DTM et les dates/tailles des viz."""
+    import os
+    from datetime import datetime
+    from lidar_pipeline.index import _collect_tile_metadata
+
+    basename = "LHD_FXX_1000_6881_PTS_LAMB93_IGN69"
+    tile_dir = tmp_path / "visualisations" / basename
+    tile_dir.mkdir(parents=True)
+    viz_file = tile_dir / f"{basename}_hillshade_multi.webp"
+    viz_file.write_bytes(b"fake")
+
+    dtm_dir = tmp_path / "DTM"
+    dtm_dir.mkdir()
+    method_file = dtm_dir / f"{basename}_dtm_method.txt"
+    method_file.write_text("ign", encoding="utf-8")
+    # Dates déterministes : method.txt plus ancien que la viz
+    os.utime(method_file, (1600000000, 1600000000))
+    os.utime(viz_file, (1700000000, 1700000000))
+    fmt = lambda ts: datetime.fromtimestamp(ts).strftime('%Y-%m-%d %H:%M')
+
+    tile = {
+        'basename': basename, 'resolution': 0.5,
+        'dir_path': str(tile_dir),
+        'viz': {'hillshade_multi': {'filename': viz_file.name, 'ext': 'webp'}},
+    }
+    meta = _collect_tile_metadata(tile, dtm_dir)
+    assert meta['method'] == 'ign'
+    assert meta['generated'] == fmt(1600000000)
+    assert meta['viz']['hillshade_multi']['size'] == 4
+    assert meta['viz']['hillshade_multi']['date'] == fmt(1700000000)
+
+
+def test_collect_tile_metadata_resolution_suffix(tmp_path):
+    """Une tuile 0,2 m lit son sidecar _dtm_r0p2_method.txt dédié."""
+    from lidar_pipeline.index import _collect_tile_metadata
+
+    basename = "LHD_FXX_1000_6881_PTS_LAMB93_IGN69"
+    tile_dir = tmp_path / "visualisations" / (basename + "_r0p2")
+    tile_dir.mkdir(parents=True)
+    dtm_dir = tmp_path / "DTM"
+    dtm_dir.mkdir()
+    (dtm_dir / f"{basename}_dtm_r0p2_method.txt").write_text("smrf", encoding="utf-8")
+
+    tile = {'basename': basename, 'resolution': 0.2,
+            'dir_path': str(tile_dir),
+            'viz': {}}
+    meta = _collect_tile_metadata(tile, dtm_dir)
+    assert meta['method'] == 'smrf'
+    # La date vient du sidecar (écrit juste après la création du DTM)
+    assert meta['generated'] is not None
+    assert meta['viz'] == {}
+
+
+def test_collect_tile_metadata_fallback_date(tmp_path):
+    """Sans sidecar DTM, la date de génération remonte au plus ancien fichier viz."""
+    from lidar_pipeline.index import _collect_tile_metadata
+
+    basename = "LHD_FXX_1000_6881_PTS_LAMB93_IGN69"
+    tile_dir = tmp_path / "visualisations" / basename
+    tile_dir.mkdir(parents=True)
+    f = tile_dir / f"{basename}_svf.webp"
+    f.write_bytes(b"x")
+
+    tile = {'basename': basename, 'resolution': 0.5,
+            'dir_path': str(tile_dir),
+            'viz': {'svf': {'filename': f.name, 'ext': 'webp'}}}
+    meta = _collect_tile_metadata(tile, tmp_path / "DTM")
+    assert meta['method'] is None
+    assert meta['generated'] is not None
+
+
 def test_build_index_generates_html(tmp_path):
     """build_index génère index.html et les vignettes."""
     from lidar_pipeline.index import build_index
@@ -142,11 +225,18 @@ def test_build_index_generates_html(tmp_path):
 
     content = html_path.read_text(encoding='utf-8')
     # Vérifie la présence des éléments clés
-    assert "Carte continue LiDAR" in content
+    assert "Carte LiDAR" in content
     assert "LHD_FXX_1000_6881" in content
     assert "LHD_FXX_1001_6881" in content
     # Vérifie que le JSON intégré est valide
-    assert "const DATA" in content
+    assert "const TILES" in content
+    # Vérifie les assets de l'interface (CSS/JS séparés)
+    assets = output_dir / "assets"
+    assert (assets / "app.css").read_text(encoding='utf-8').startswith('/*')
+    app_js = (assets / "app.js").read_text(encoding='utf-8')
+    assert "Couches" in app_js or "layers" in app_js
+    assert 'assets/app.css' in content
+    assert 'assets/app.js' in content
     # Vérifie les vignettes générées
     thumb_dir = output_dir / "index_thumbs"
     assert thumb_dir.is_dir()
@@ -154,6 +244,68 @@ def test_build_index_generates_html(tmp_path):
     assert len(thumbs) >= 2  # au moins hillshade pour chaque tuile
 
 
+def test_build_index_regenerates_stale_thumbnails(tmp_path):
+    """Une tuile recalculée (source plus récente) régénère sa vignette."""
+    import os
+    import time
+    import numpy as np
+    from PIL import Image as PILImage
+    from lidar_pipeline.index import build_index
+
+    output_dir = tmp_path / "output"
+    vis_dir = output_dir / "visualisations"
+    vis_dir.mkdir(parents=True)
+    tile_dir = _make_fake_viz_dir(vis_dir, "a", 1000, 6881, ('hillshade_multi',))
+
+    assert build_index(output_dir) is not None
+    thumb_path = output_dir / "index_thumbs" / "LHD_FXX_1000_6881_PTS_LAMB93_IGN69_hillshade_multi.jpg"
+    assert thumb_path.exists()
+    m1 = thumb_path.stat().st_mtime
+
+    # Recalcul de la tuile : source réécrite avec une mtime plus récente
+    src = tile_dir / "LHD_FXX_1000_6881_PTS_LAMB93_IGN69_hillshade_multi.webp"
+    arr = np.random.randint(0, 255, (50, 50, 3), dtype=np.uint8)
+    PILImage.fromarray(arr).save(str(src), format='WEBP', quality=80)
+    os.utime(src, (m1 + 5, m1 + 5))
+
+    assert build_index(output_dir) is not None
+    m2 = thumb_path.stat().st_mtime
+    assert m2 > m1  # vignette régénérée
+
+    # Source non modifiée depuis → pas de régénération inutile
+    os.utime(src, (time.time() - 10, time.time() - 10))
+    assert build_index(output_dir) is not None
+    assert thumb_path.stat().st_mtime == m2
+
+
+def test_build_subtiles_regenerates_stale_crops(tmp_path):
+    """Une dalle 0,2 m recalculée régénère ses sous-tuiles (par visualisation)."""
+    import os
+    from lidar_pipeline.index import build_index
+
+    output_dir = tmp_path / "output"
+    vis_dir = output_dir / "visualisations"
+    vis_dir.mkdir(parents=True)
+    tile_dir = _make_fake_viz_dir(vis_dir, "a", 1000, 6881,
+                                  ('hillshade_multi', 'aspect'), res_suffix='_r0p2')
+
+    assert build_index(output_dir) is not None
+    sub_dir = output_dir / "index_subtiles"
+    hill_avif = sub_dir / "LHD_FXX_1000_6881_PTS_LAMB93_IGN69_r0p2_hillshade_multi_0_0.avif"
+    aspect_avif = sub_dir / "LHD_FXX_1000_6881_PTS_LAMB93_IGN69_r0p2_aspect_0_0.avif"
+    assert hill_avif.exists() and aspect_avif.exists()
+    m_hill_1 = hill_avif.stat().st_mtime
+    m_aspect_1 = aspect_avif.stat().st_mtime
+
+    # Recalcul : seule la source hillshade est plus récente
+    src = tile_dir / "LHD_FXX_1000_6881_PTS_LAMB93_IGN69_hillshade_multi.webp"
+    os.utime(src, (m_hill_1 + 5, m_hill_1 + 5))
+
+    assert build_index(output_dir) is not None
+    assert hill_avif.stat().st_mtime > m_hill_1   # sous-tuiles hillshade régénérées
+    assert aspect_avif.stat().st_mtime == m_aspect_1  # aspect intact
+
+
 def test_build_index_empty_returns_none(tmp_path):
     """Aucune tuile → build_index retourne None sans crash."""
     from lidar_pipeline.index import build_index
@@ -175,26 +327,66 @@ def test_build_index_embeds_valid_json(tmp_path):
 
     build_index(output_dir)
     content = (output_dir / "index.html").read_text(encoding='utf-8')
-    # Extrait le JSON entre "const DATA = " et ";"
+    # Extrait le JSON entre "const TILES = " et la fin de déclaration
-    start = content.index("const DATA = ") + len("const DATA = ")
+    start = content.index("const TILES = ") + len("const TILES = ")
     # Trouve le ; de fin de déclaration
     depth = 0
     end = start
     for i, ch in enumerate(content[start:], start):
-        if ch == '{':
+        if ch in ('{', '['):
             depth += 1
-        elif ch == '}':
+        elif ch in ('}', ']'):
             depth -= 1
             if depth == 0:
                 end = i + 1
                 break
     data = json.loads(content[start:end])
-    assert 'tiles' in data
+    assert len(data) > 0
-    assert 'bbox' in data
+    assert data[0]['col'] == 1000
-    assert 'vizList' in data
+    assert data[0]['row'] == 6881
-    assert len(data['tiles']) == 1
+
-    assert data['tiles'][0]['col'] == 1000
+
-    assert data['tiles'][0]['row'] == 6881
+def test_attach_gps_bounds():
+    """attach_gps_bounds ajoute des bounds GPS ordonnées (France métropolitaine)."""
+    from lidar_pipeline.index import attach_gps_bounds
+    tiles = [{'col': 1000, 'row': 6881}, {'col': 1042, 'row': 6900}]
+    attach_gps_bounds(tiles)
+    for t in tiles:
+        assert 'bounds' in t
+        (lat_s, lon_w), (lat_n, lon_e) = t['bounds']
+        assert lat_n > lat_s
+        assert lon_e > lon_w
+        # France métropolitaine
+        assert 41 < lat_s < 51
+        assert -5 < lon_w < 10
+
+
+def test_attach_gps_bounds_row_is_north_edge():
+    """Le numéro de ligne du fichier = bord NORD (convention LiDAR HD IGN).
+
+    Vérifié sur les bounds des DTM : X ∈ [col, col+1] km, Y ∈ [row-1, row] km.
+    La régression historique plaçait Y ∈ [row, row+1] (1 km trop au nord).
+    """
+    from rasterio.warp import transform as warp_transform
+    from lidar_pipeline.index import attach_gps_bounds
+
+    col, row = 1054, 6882
+    tiles = [{'col': col, 'row': row}]
+    attach_gps_bounds(tiles)
+    corners = tiles[0]['corners']
+
+    # Référence exacte de la vraie cellule : SW, SE, NE, NW
+    xs = [col * 1000, (col + 1) * 1000, (col + 1) * 1000, col * 1000]
+    ys = [(row - 1) * 1000, (row - 1) * 1000, row * 1000, row * 1000]
+    lons, lats = warp_transform('EPSG:2154', 'EPSG:4326', xs, ys)
+    for k in range(4):
+        assert abs(corners[k][0] - lats[k]) < 1e-9
+        assert abs(corners[k][1] - lons[k]) < 1e-9
+
+    # L'ancienne convention (row = bord sud) serait décalée d'environ 1 km
+    lat_n = max(c[0] for c in corners)
+    assert abs(lat_n - max(lats)) < 1e-9  # bord nord = Y = row×1000
+
 
 
 def test_pick_display_viz_prefers_hillshade():
@@ -203,3 +395,30 @@ def test_pick_display_viz_prefers_hillshade():
     assert _pick_display_viz(['svf', 'hillshade_multi', 'slope']) == 'hillshade_multi'
     assert _pick_display_viz(['svf', 'slope']) == 'svf'
     assert _pick_display_viz(['topo']) == 'topo'
+
+
+def test_subdivision_k():
+    """0,5 m/px (2000 px) reste entier ; 0,2 m/px (5000 px) est découpé en 2×2."""
+    from lidar_pipeline.index import _subdivision_k
+    assert _subdivision_k(0.5) == 1
+    assert _subdivision_k(0.2) == 2
+    assert _subdivision_k(1.0) == 1
+
+
+def test_subtile_corners_grid():
+    """Les sous-tuiles reconstruisent exactement la grille de la dalle."""
+    from lidar_pipeline.index import _subtile_corners
+    corners = [[10.0, 2.0], [10.0, 3.0], [11.0, 3.0], [11.0, 2.0]]  # SW SE NE NW
+    k = 2
+    sw_quad = _subtile_corners(corners, 0, 0, k)   # quadrant sud-ouest
+    ne_quad = _subtile_corners(corners, 1, 1, k)   # quadrant nord-est
+    # Le quadrant SW partage le coin SW de la dalle
+    assert sw_quad[0] == corners[0]
+    # Le quadrant NE partage le coin NE de la dalle
+    assert ne_quad[2] == corners[2]
+    # Le quadrant SW a son coin NE au centre de la dalle
+    assert sw_quad[2] == [10.5, 2.5]
+    # Adjacence : bord est du SW = bord ouest du SE (0,0)-(1,0)
+    se_quad = _subtile_corners(corners, 1, 0, k)
+    assert sw_quad[1] == se_quad[0]
+    assert sw_quad[2] == se_quad[3]

lidar_pipeline/tests/test_pipeline.py

@@ -71,3 +71,98 @@ class TestLidarArchaeoPipeline:
         assert "test.laz" in names
         assert "other.las" in names
         assert "readme.txt" not in names
+
+
+class TestDtmMethodSidecar:
+    """Méthode de classification enregistrée à côté du DTM (invalidation du cache)."""
+
+    def test_missing_sidecar_matches(self, tmp_path):
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        input_dir = tmp_path / "input"
+        input_dir.mkdir()
+        pipeline = LidarArchaeoPipeline(str(input_dir), str(tmp_path / "output"), ground_method='csf')
+        # Aucun sidecar écrit → cache conservé (considéré compatible).
+        assert pipeline._dtm_method_matches("tileA", "") is True
+
+    def test_matching_method(self, tmp_path):
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        input_dir = tmp_path / "input"
+        input_dir.mkdir()
+        pipeline = LidarArchaeoPipeline(str(input_dir), str(tmp_path / "output"), ground_method='csf')
+        pipeline._write_dtm_method("tileA", "")
+        assert pipeline._dtm_method_matches("tileA", "") is True
+
+    def test_different_method_invalidates_cache(self, tmp_path):
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        input_dir = tmp_path / "input"
+        input_dir.mkdir()
+        out = str(tmp_path / "output")
+        LidarArchaeoPipeline(str(input_dir), out, ground_method='ign')._write_dtm_method("tileA", "")
+        csf = LidarArchaeoPipeline(str(input_dir), out, ground_method='csf')
+        assert csf._dtm_method_matches("tileA", "") is False
+
+    def test_write_dtm_method(self, tmp_path):
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        input_dir = tmp_path / "input"
+        input_dir.mkdir()
+        pipeline = LidarArchaeoPipeline(str(input_dir), str(tmp_path / "output"), ground_method='smrf')
+        pipeline._write_dtm_method("tileA", "_r0p2")
+        sidecar = tmp_path / "output" / "DTM" / "tileA_dtm_r0p2_method.txt"
+        assert sidecar.exists()
+        assert sidecar.read_text(encoding="utf-8").strip() == "smrf"
+        assert pipeline._dtm_method_name("tileA", "_r0p2") == "smrf"
+        # Le sidecar est un fichier .txt : il ne gêne pas la recherche des DTM .tif.
+        dtm = tmp_path / "output" / "DTM" / "tileA_dtm_r0p2.tif"
+        dtm.touch()
+        assert [p.name for p in (tmp_path / "output" / "DTM").glob("*.tif")] == ["tileA_dtm_r0p2.tif"]
+
+    def test_force_images_regenerates_existing(self, tmp_path):
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        input_dir = tmp_path / "input"
+        input_dir.mkdir()
+        pipeline = LidarArchaeoPipeline(str(input_dir), str(tmp_path / "output"), output_format='avif')
+        calls = []
+
+        def fake_ortho(dem_file, basename, vis_dir, resolution):
+            calls.append(basename)
+            return vis_dir / f"{basename}_ortho.avif"
+
+        pipeline.viz_steps = [('ortho', fake_ortho)]
+        vis_dir = tmp_path / "output" / "visualisations" / "tileA"
+        vis_dir.mkdir(parents=True)
+        (vis_dir / "tileA_ortho.avif").touch()
+        dtm = tmp_path / "dtm.tif"
+
+        # Image existante, pas de force → ignorée (pas de régénération).
+        pipeline.generate_all_visualizations(dtm, "tileA", resolution=0.5, vis_dir=vis_dir, force_images=False)
+        assert calls == []
+
+        # Image existante, force_images=True → régénérée.
+        calls.clear()
+        pipeline.generate_all_visualizations(dtm, "tileA", resolution=0.5, vis_dir=vis_dir, force_images=True)
+        assert calls == ["tileA"]
+
+class TestEffectiveGroundMethod:
+    def test_ign_label_encodes_classes(self):
+        """Les classes IGN sont encodées dans l'étiquette de cache (reclassification)."""
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        import tempfile
+        with tempfile.TemporaryDirectory() as tmpdir:
+            p = LidarArchaeoPipeline(tmpdir, tmpdir, ground_method='ign',
+                                     ign_classes="sol,unclassified")
+            assert p._effective_ground_method() == "ign_1_2"
+
+    def test_ign_default_label(self):
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        import tempfile
+        with tempfile.TemporaryDirectory() as tmpdir:
+            p = LidarArchaeoPipeline(tmpdir, tmpdir, ground_method='ign')
+            assert p._effective_ground_method() == "ign"
+
+    def test_other_methods_unchanged(self):
+        from lidar_pipeline.pipeline import LidarArchaeoPipeline
+        import tempfile
+        with tempfile.TemporaryDirectory() as tmpdir:
+            p = LidarArchaeoPipeline(tmpdir, tmpdir, ground_method='smrf',
+                                     ign_classes="sol,unclassified")
+            assert p._effective_ground_method() == "smrf"

lidar_pipeline/tests/test_rendering.py

@@ -96,3 +96,57 @@ class TestApplyColormap:
         result = tif_to_png(tif_file, tmp_path, 5.0)
         assert result is not None
         assert result.exists()
+
+
+class TestTifToCrop:
+    """Conversion TIF → dalle cartographique (tif_to_crop)."""
+
+    @staticmethod
+    def _write_named_tif(tmp_path, name, arr):
+        transform = from_bounds(660000, 6700000, 661000, 6701000, arr.shape[1], arr.shape[0])
+        tif_file = tmp_path / name
+        with rasterio.open(
+            tif_file, 'w', driver='GTiff', height=arr.shape[0], width=arr.shape[1],
+            count=1, dtype='float32', crs='EPSG:2154', transform=transform,
+            nodata=float('nan'), compress='lzw'
+        ) as dst:
+            dst.write(arr.astype('float32'), 1)
+        return tif_file
+
+    def test_nodata_renders_black(self, tmp_path):
+        """Le nodata restant est rendu en noir (comportement historique).
+
+        Les trous du MNT sont comblés en amont (interpolation dans
+        create_dtm_fast, tous modes) ; ce qui reste en nodata doit rester
+        visible en noir sur la dalle plutôt qu'inventé au rendu.
+        """
+        from PIL import Image as PILImage
+        from lidar_pipeline.rendering import tif_to_crop
+
+        data = np.random.default_rng(7).normal(50, 10, (40, 40)).astype(np.float32)
+        data[15:25, 15:25] = np.nan
+        tif_file = self._write_named_tif(tmp_path, "LHD_test_slope.tif", data)
+
+        # WebP lossless : l'encodeur AVIF de l'image « saigne » légèrement les
+        # bords du noir même en lossless — on teste la logique nodata→noir,
+        # pas les artefacts du codec.
+        out = tif_to_crop(tif_file, tmp_path, 5.0, keep_tif=True,
+                          quality=100, output_format='webp')
+        assert out is not None and out.exists()
+
+        rgb = np.asarray(PILImage.open(str(out)).convert('RGB'))
+        hole = rgb[15:25, 15:25, :]
+        assert np.all(hole == 0), "le nodata doit être rendu en noir"
+
+    def test_without_nodata(self, tmp_path):
+        """Un TIF sans nodata est converti sans crash, taille préservée."""
+        from PIL import Image as PILImage
+        from lidar_pipeline.rendering import tif_to_crop
+
+        data = np.random.default_rng(7).normal(50, 10, (40, 40)).astype(np.float32)
+        tif_file = self._write_named_tif(tmp_path, "LHD_test_slope.tif", data)
+
+        out = tif_to_crop(tif_file, tmp_path, 5.0)
+        assert out is not None and out.exists()
+        img = PILImage.open(str(out))
+        assert img.size == (40, 40)

lidar_pipeline/tests/test_visualizations.py

@@ -157,3 +157,61 @@ class TestRayTrace:
         )
         assert pos.shape == (4, 2, rows, cols)
         assert neg.shape == (4, 2, rows, cols)
+
+
+class TestNodataPreserved:
+    """Nodata préservé dans les rendus (comportement historique).
+
+    Les trous du MNT sont comblés en amont (create_dtm_fast, tous modes) ;
+    si un nodata subsiste malgré tout, hillshade/slope/aspect le restituent
+    (rendu noir en carte) au lieu d'inventer des valeurs interpolées.
+    """
+
+    @staticmethod
+    def _dem_with_hole(synthetic_dem, tmp_path):
+        import rasterio
+        with rasterio.open(synthetic_dem) as src:
+            arr = src.read(1).copy()
+            profile = src.profile.copy()
+        arr[80:120, 80:120] = np.nan
+        dem_hole = tmp_path / "dem_hole.tif"
+        profile.update(dtype='float32', nodata=float('nan'))
+        with rasterio.open(dem_hole, 'w', **profile) as dst:
+            dst.write(arr.astype('float32'), 1)
+        return dem_hole
+
+    def test_aspect_solo_preserves_nodata(self, synthetic_dem, tmp_path):
+        from lidar_pipeline.visualizations import generate_aspect
+        dem_hole = self._dem_with_hole(synthetic_dem, tmp_path)
+        out = generate_aspect(dem_hole, "solo", tmp_path, 5.0)
+        assert out is not None and out.exists()
+        import rasterio
+        with rasterio.open(out) as src:
+            data = src.read(1)
+        assert np.isnan(data[80:120, 80:120]).all(), "le trou doit rester en nodata"
+        # Le gradient au bord du trou propage NaN sur un anneau de 1 px :
+        # on vérifie une zone éloignée du trou
+        assert not np.isnan(data[0:40, 0:40]).any(), "NaN loin du trou"
+
+    def test_aspect_shared_preserves_nodata(self, synthetic_dem, tmp_path):
+        from lidar_pipeline.visualizations import SharedDEM, generate_aspect
+        dem_hole = self._dem_with_hole(synthetic_dem, tmp_path)
+        shared = SharedDEM(dem_hole, 5.0)
+        out = generate_aspect(dem_hole, "partage", tmp_path, 5.0, shared=shared)
+        assert out is not None and out.exists()
+        import rasterio
+        with rasterio.open(out) as src:
+            data = src.read(1)
+        assert np.isnan(data[80:120, 80:120]).all(), "le trou doit rester en nodata"
+        assert not np.isnan(data[0:40, 0:40]).any(), "NaN loin du trou"
+
+    def test_slope_and_hillshade_preserve_nodata(self, synthetic_dem, tmp_path):
+        from lidar_pipeline.visualizations import generate_slope, generate_hillshade
+        dem_hole = self._dem_with_hole(synthetic_dem, tmp_path)
+        import rasterio
+        for gen, name in ((generate_slope, "p"), (generate_hillshade, "h")):
+            out = gen(dem_hole, name, tmp_path, 5.0)
+            assert out is not None and out.exists()
+            with rasterio.open(out) as src:
+                data = src.read(1)
+            assert np.isnan(data[80:120, 80:120]).any(), f"{out.name} : trou disparu"

lidar_pipeline/tests/test_webapp.py

@@ -0,0 +1,81 @@
+"""Tests du serveur web de génération de zones (webapp)."""
+
+
+def test_bbox_to_cells_single_km_cell():
+    """Une bbox couvrant ~1 km² retourne la cellule L93 correspondante."""
+    from lidar_pipeline.webapp import bbox_to_cells
+    # Cellule 1054,6882 : X∈[1054000,1055000], Y∈[6881000,6882000] (L93)
+    from rasterio.warp import transform as warp_transform
+    lons, lats = warp_transform('EPSG:2154', 'EPSG:4326',
+                                [1054100, 1054900], [6881100, 6881900])
+    cells = bbox_to_cells(min(lons), min(lats), max(lons), max(lats))
+    assert (1054, 6882) in cells
+    # La sélection reste locale : pas de cellule lointaine
+    for (c, r) in cells:
+        assert abs(c - 1054) <= 1 and abs(r - 6882) <= 1
+
+
+def test_bbox_to_cells_empty_for_tiny_bbox():
+    """Une bbox quasi ponctuelle ne sélectionne rien (rétrécie sous 1 m)."""
+    from lidar_pipeline.webapp import bbox_to_cells
+    assert bbox_to_cells(7.850000, 48.930000, 7.850001, 48.930001) == []
+
+
+def test_processed_cells(tmp_path):
+    """processed_cells lit les dossiers de visualisations."""
+    from lidar_pipeline.webapp import processed_cells
+    vis = tmp_path / "visualisations" / "LHD_FXX_1054_6882_PTS_LAMB93_IGN69_r0p2"
+    vis.mkdir(parents=True)
+    (vis / "LHD_FXX_1054_6882_PTS_LAMB93_IGN69_r0p2_aspect.avif").write_bytes(b"x")
+    assert processed_cells(tmp_path) == {(1054, 6882)}
+
+
+def test_missing_cells_filters_processed(tmp_path):
+    """Les cellules déjà traitées sont exclues, les autres gardent leurs coins."""
+    from lidar_pipeline.webapp import missing_cells_with_corners
+    vis = tmp_path / "visualisations" / "LHD_FXX_1054_6882_PTS_LAMB93_IGN69"
+    vis.mkdir(parents=True)
+    (vis / "LHD_FXX_1054_6882_PTS_LAMB93_IGN69_aspect.avif").write_bytes(b"x")
+    todo = missing_cells_with_corners([(1054, 6882), (1055, 6882)], tmp_path)
+    assert len(todo) == 1
+    assert todo[0]['col'] == 1055 and todo[0]['row'] == 6882
+    assert len(todo[0]['corners']) == 4  # SW, SE, NE, NW
+
+
+def test_missing_cells_include_done(tmp_path):
+    """include_done=True conserve les cellules déjà traitées (régénération)."""
+    from lidar_pipeline.webapp import missing_cells_with_corners
+    vis = tmp_path / "visualisations" / "LHD_FXX_1054_6882_PTS_LAMB93_IGN69"
+    vis.mkdir(parents=True)
+    (vis / "LHD_FXX_1054_6882_PTS_LAMB93_IGN69_aspect.avif").write_bytes(b"x")
+    todo = missing_cells_with_corners([(1054, 6882), (1055, 6882)], tmp_path,
+                                      include_done=True)
+    assert {(t['col'], t['row']) for t in todo} == {(1054, 6882), (1055, 6882)}
+
+
+def test_build_command_regenerate():
+    """regenerate=True ajoute --force --force-classification à la commande."""
+    from lidar_pipeline.webapp import _build_command
+    cmd = " ".join(_build_command([(1054, 6882)], regenerate=True))
+    assert "--force" in cmd
+    assert "--force-classification" in cmd
+    cmd = " ".join(_build_command([(1054, 6882)]))
+    assert "--force" not in cmd
+    assert "--force-classification" not in cmd
+
+
+def test_build_command_ground_classification():
+    """La commande utilise la méthode de classification demandée (défaut : ign)."""
+    from lidar_pipeline.webapp import _build_command, GROUND_CLASS_METHODS
+    # Défaut : ign (pré-classification)
+    cmd = _build_command([(1054, 6882)])
+    i = cmd.index("--ground-classification")
+    assert cmd[i + 1] == "ign"
+    # Chaque méthode valide est transmise telle quelle, avec ou sans régénération
+    for method in GROUND_CLASS_METHODS:
+        for regenerate in (False, True):
+            cmd = _build_command([(1054, 6882)], regenerate=regenerate, ground_class=method)
+            i = cmd.index("--ground-classification")
+            assert cmd[i + 1] == method
+            assert ("--force" in cmd) == regenerate
+            assert ("--force-classification" in cmd) == regenerate

lidar_pipeline/webapp.py

@@ -0,0 +1,254 @@
+"""Serveur web de la carte LiDAR : sert l'index et expose l'API de génération.
+
+Lancé via `./run.sh --serve [PORT]` (input/ monté en écriture pour permettre
+le téléchargement IGN). Endpoints :
+
+  GET  /               → carte interactive (output/index.html)
+  GET  /api/status     → état de la génération en cours (ou dernière terminée)
+  POST /api/preview    → cellules 1 km intersectant une bbox WGS84 (option
+                         regenerate=true pour inclure celles déjà générées)
+  POST /api/generate   → télécharge (géoplateforme IGN) puis traite des cellules
+                         (option regenerate=true ajoute --force --force-classification ;
+                         option ground_class choisit la méthode de classification du sol)
+
+Fichiers statiques : /assets (interface), /index_thumbs, /index_subtiles,
+/visualisations, /DTM.
+Un seul job à la fois : la génération lance `python -m lidar_pipeline` en
+sous-processus avec --fetch-tiles + --file, journalisé dans .generation.log.
+"""
+
+import math
+import os
+import subprocess
+import sys
+import threading
+import time
+from pathlib import Path
+
+from fastapi import FastAPI, HTTPException
+from fastapi.responses import FileResponse, JSONResponse
+from pydantic import BaseModel, Field
+
+INPUT_DIR = Path(os.environ.get("LIDAR_INPUT_DIR", "/data/input"))
+OUTPUT_DIR = Path(os.environ.get("LIDAR_OUTPUT_DIR", "/data/output"))
+PORT = int(os.environ.get("LIDAR_PORT", "8973"))
+LOG_FILE = OUTPUT_DIR / ".generation.log"
+MAX_CELLS = 400  # garde-fou : ~400 km² max par demande
+
+app = FastAPI(title="Carte LiDAR — génération de zones")
+
+# assets/ (CSS/JS de l'interface) est créé dès le démarrage pour que le monteur
+# statique soit actif même avant la première génération de l'index.
+_assets_dir = OUTPUT_DIR / "assets"
+_assets_dir.mkdir(parents=True, exist_ok=True)
+for name in ("index_thumbs", "index_subtiles", "visualisations", "DTM"):
+    _dir = OUTPUT_DIR / name
+    if _dir.exists():
+        from fastapi.staticfiles import StaticFiles
+        app.mount(f"/{name}", StaticFiles(directory=str(_dir)), name=name)
+
+
+@app.get("/assets/{file_path:path}")
+def assets(file_path: str):
+    """Sert les fichiers de l'interface sans cache (régénérés à chaque rebuild)."""
+    base = _assets_dir.resolve()
+    p = (_assets_dir / file_path).resolve()
+    if base not in p.parents or not p.is_file():
+        raise HTTPException(404, f"asset introuvable : {file_path}")
+    return FileResponse(str(p), headers={"Cache-Control": "no-cache, must-revalidate"})
+
+
+# Méthodes de classification du sol acceptées (mêmes valeurs que --ground-classification).
+GROUND_CLASS_METHODS = ("auto", "ign", "smrf", "csf")
+
+
+class PreviewRequest(BaseModel):
+    bbox: list = Field(..., description="[ouest, sud, est, nord] en WGS84")
+    regenerate: bool = Field(False, description="Inclure les tuiles déjà générées")
+
+
+class GenerateRequest(BaseModel):
+    tiles: list = Field(..., description="liste [col, row] (entiers km L93)")
+    regenerate: bool = Field(False, description="Régénérer les tuiles déjà générées")
+    ground_class: str = Field("ign",
+                              description="Méthode de classification du sol : "
+                                          "auto, ign, smrf, csf")
+    ign_classes: str = Field("sol",
+                             description="Classes LAS pour le MNT IGN : liste noms ou "
+                                         "codes séparés par virgules — sol(2), "
+                                         "unclassified(1), eau(9), virtuel(66), "
+                                         "pont(17), sursol(64). Mode pur, "
+                                         "aucune retouche. (défaut: sol)")
+    bare_earth: bool = Field(False,
+                             description="Sol nu : DTM au retour le plus bas de "
+                                         "chaque cellule (requalifie le point le plus "
+                                         "bas en terrain)")
+
+
+# --- État du job de génération -------------------------------------------
+_job = {"proc": None, "started": None, "returncode": None, "cmd": None}
+_job_lock = threading.Lock()
+
+
+def bbox_to_cells(w, s, e, n):
+    """Cellules L93 de 1 km (col, row) intersectant une bbox WGS84.
+
+    Une cellule (col, row) couvre X ∈ [col, col+1] km, Y ∈ [row-1, row] km.
+    """
+    from rasterio.warp import transform as warp_transform
+    lons, lats = warp_transform("EPSG:4326", "EPSG:2154", [w, e, w, e], [s, s, n, n])
+    # warp_transform renvoie (xs, ys) dans la CRS cible
+    min_x, max_x = min(lons) + 0.5, max(lons) - 0.5  # rétrécit d'1 m : bords exclus
+    min_y, max_y = min(lats) + 0.5, max(lats) - 0.5
+    if max_x <= min_x or max_y <= min_y:
+        return []
+    cols = range(int(math.floor(min_x / 1000)), int(math.floor(max_x / 1000)) + 1)
+    rows = range(int(math.floor(min_y / 1000)) + 1, int(math.floor(max_y / 1000)) + 2)
+    return [(c, r) for r in rows for c in cols]
+
+
+def processed_cells(output_dir):
+    """Ensemble des cellules (col, row) ayant déjà des visualisations."""
+    from .index import scan_tiles
+    tiles = scan_tiles(Path(output_dir) / "visualisations")
+    return {(t["col"], t["row"]) for t in tiles}
+
+
+def missing_cells_with_corners(cells, output_dir, include_done=False):
+    """Filtre les cellules déjà traitées et calcule leurs coins WGS84.
+
+    Retourne [{col, row, corners: [[lat, lon] × 4 SW,SE,NE,NW}].
+    """
+    from .index import attach_gps_bounds
+    done = processed_cells(output_dir)
+    todo = [{"col": c, "row": r} for (c, r) in cells
+            if include_done or (c, r) not in done]
+    if todo:
+        attach_gps_bounds(todo)
+    return todo
+
+
+@app.get("/")
+def root():
+    index = OUTPUT_DIR / "index.html"
+    if not index.exists():
+        return JSONResponse({"erreur": "index.html introuvable — lancez d'abord le pipeline"},
+                            status_code=404)
+    # index.html est régénéré à chaque passe du pipeline : on interdit le cache
+    # navigateur pour ne pas servir une version périmée (ex. menu de génération).
+    return FileResponse(
+        str(index), media_type="text/html",
+        headers={
+            "Cache-Control": "no-cache, no-store, must-revalidate",
+            "Pragma": "no-cache",
+            "Expires": "0",
+        })
+
+
+@app.get("/api/status")
+def status():
+    with _job_lock:
+        proc = _job["proc"]
+        running = proc is not None and proc.poll() is None
+        return {
+            "running": running,
+            "started": _job["started"],
+            "returncode": _job["returncode"],
+            "cmd": _job["cmd"],
+            "log": _tail_log(40),
+        }
+
+
+def _tail_log(n_lines):
+    try:
+        lines = LOG_FILE.read_text(encoding="utf-8", errors="replace").splitlines()
+        return lines[-n_lines:]
+    except Exception:
+        return []
+
+
+@app.post("/api/preview")
+def preview(req: PreviewRequest):
+    if len(req.bbox) != 4:
+        raise HTTPException(400, "bbox attendue : [ouest, sud, est, nord]")
+    w, s, e, n = (float(v) for v in req.bbox)
+    cells = bbox_to_cells(w, s, e, n)
+    capped = len(cells) > MAX_CELLS
+    todo = missing_cells_with_corners(cells[:MAX_CELLS], OUTPUT_DIR,
+                                      include_done=req.regenerate)
+    return {"count": len(todo), "capped": capped, "cells": todo}
+
+
+def _build_command(tiles, regenerate=False, ground_class="ign", bare_earth=False, ign_classes="sol"):
+    """Commande de génération : téléchargement IGN + traitement des fichiers.
+
+    La classification du sol est choisie via `ground_class` (défaut : "ign",
+    pré-classification IGN ; le pipeline bascule sur SMRF si un fichier ne la
+    contient pas). Avec ign_classes, on choisit les classes LAS extraites
+    pour le MNT (mode pur, ex. "sol,unclassified" pour combler les trous
+    sans retouche). Avec regenerate=True, force la reclassification et la
+    régénération des visualisations des tuiles déjà présentes. Avec
+    bare_earth=True, le DTM est ramené au retour le plus bas de chaque
+    cellule (sol nu).
+    """
+    from .fetch_ign import tile_filename
+    # -u : sortie non bufferisée — le journal .generation.log doit être
+    # lu en temps réel par /api/status (progression affichée dans l'UI).
+    cmd = [sys.executable, "-u", "-m", "lidar_pipeline", str(INPUT_DIR),
+           "-o", str(OUTPUT_DIR), "-r", "0.5,0.2", "--only", "aspect",
+           "--ground-classification", ground_class,
+           "--ign-classes", ign_classes]
+    if bare_earth:
+        cmd += ["--bare-earth"]
+    if regenerate:
+        cmd += ["--force", "--force-classification"]
+    if os.environ.get("LIDAR_GPU", "") == "1":
+        cmd += ["-g", "all", "-w", os.environ.get("LIDAR_WORKERS", "2")]
+    cmd += ["--fetch-tiles"]
+    cmd += [f"{c},{r}" for (c, r) in tiles]
+    cmd += ["--file"]
+    cmd += [tile_filename(c, r) for (c, r) in tiles]
+    return cmd
+
+
+@app.post("/api/generate")
+def generate(req: GenerateRequest):
+    tiles = []
+    for pair in req.tiles:
+        if not (isinstance(pair, list) and len(pair) == 2):
+            raise HTTPException(400, f"tuile invalide : {pair!r} (attendu [col, row])")
+        tiles.append((int(pair[0]), int(pair[1])))
+    if not tiles:
+        raise HTTPException(400, "aucune tuile fournie")
+    if len(tiles) > MAX_CELLS:
+        raise HTTPException(400, f"trop de tuiles ({len(tiles)}) — max {MAX_CELLS}")
+    if req.ground_class not in GROUND_CLASS_METHODS:
+        raise HTTPException(
+            400, f"méthode de classification invalide : {req.ground_class!r} "
+                 f"(attendu : {', '.join(GROUND_CLASS_METHODS)})")
+
+    with _job_lock:
+        proc = _job["proc"]
+        if proc is not None and proc.poll() is None:
+            raise HTTPException(409, "une génération est déjà en cours")
+        cmd = _build_command(tiles, regenerate=req.regenerate, ground_class=req.ground_class, bare_earth=req.bare_earth, ign_classes=req.ign_classes)
+        LOG_FILE.parent.mkdir(parents=True, exist_ok=True)
+        log_fh = open(LOG_FILE, "w", encoding="utf-8")
+        _job.update({"proc": None, "started": time.time(), "returncode": None, "cmd": cmd})
+        p = subprocess.Popen(cmd, stdout=log_fh, stderr=subprocess.STDOUT,
+                             cwd="/app" if Path("/app").exists() else None)
+
+        def _watch():
+            rc = p.wait()
+            log_fh.close()
+            with _job_lock:
+                _job["returncode"] = rc
+
+        threading.Thread(target=_watch, daemon=True).start()
+        _job["proc"] = p
+    return {"demarré": True, "tuiles": len(tiles), "commande": " ".join(cmd)}
+
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=PORT)

run.sh

@@ -11,7 +11,9 @@
 #   -f / --force    Régénérer tous les fichiers
 #   --keep-tif      Conserver les fichiers TIFF
 #   --force-classification
-#   --ground-classification {auto,smrf,csf}
+#   --ground-classification {auto,ign,smrf,csf}
+#   --ign-classes CLASSES
+#                   Classes LAS pour le MNT IGN (ex: sol,unclassified ; défaut: sol)
 #   --quality N     Qualité image 1-100 (défaut: 98)
 #   --lossless      Compression lossless
 #   --format FMT    Format de sortie : avif (défaut) ou webp
@@ -44,8 +46,11 @@ if [ $# -eq 0 ]; then
     echo "  --force-classification"
     echo "                  Reclassifier le sol même si le fichier .las existe"
     echo "  --keep-tif      Conserver les TIFF pour régénérer les AVIF"
-    echo "  --ground-classification {auto,smrf,csf}"
+    echo "  --ground-classification {auto,ign,smrf,csf}"
-    echo "                  Méthode de classification du sol (défaut: auto)"
+    echo "                  Méthode de classification du sol (défaut: auto = pré-classification IGN)"
+    echo "  --ign-classes CLASSES"
+    echo "                  Classes LAS pour le MNT IGN : sol,unclassified,eau,virtuel,pont,sursol"
+    echo "                  ou codes (ex: 2,1). Mode pur, aucune retouche. (défaut: sol)"
     echo "  --quality N     Qualité image 1-100 (défaut: 98, 100=lossless)"
     echo "  --lossless      Compression lossless (équivalent à --quality 100)"
     echo "  --format FMT    Format de sortie : avif (défaut) ou webp"
@@ -83,6 +88,7 @@ VERBOSE_FLAG=""
 FORCE_FLAG=""
 FILE_ARGS=""
 GROUND_METHOD=""
+IGN_CLASSES_FLAG=""
 FORCE_CLASSIFY_FLAG=""
 KEEP_TIF_FLAG=""
 QUALITY=""
@@ -92,6 +98,9 @@ SKIP_FLAG=""
 TEST_FLAG=0
 REBUILD_INDEX_FLAG=""
 NO_INDEX_FLAG=""
+FETCH_TILES_ARGS=""
+SERVE_FLAG=0
+SERVE_PORT=8973
 
 # Parse arguments manually (more robust than getopts for mixed short/long options)
 while [ $# -gt 0 ]; do
@@ -117,6 +126,8 @@ while [ $# -gt 0 ]; do
         --keep-tif) KEEP_TIF_FLAG="--keep-tif"; shift ;;
         --ground-classification) GROUND_METHOD="$2"; shift 2 ;;
         --ground-classification=*) GROUND_METHOD="${1#--ground-classification=}"; shift ;;
+        --ign-classes) IGN_CLASSES_FLAG="--ign-classes $2"; shift 2 ;;
+        --ign-classes=*) IGN_CLASSES_FLAG="--ign-classes=${1#--ign-classes=}"; shift ;;
         --quality) QUALITY="--quality $2"; shift 2 ;;
         --lossless) QUALITY="--lossless"; shift ;;
         --format) FORMAT_FLAG="--format $2"; shift 2 ;;
@@ -125,7 +136,12 @@ while [ $# -gt 0 ]; do
         --file) shift; while [ $# -gt 0 ] && [[ ! "$1" =~ ^- ]]; do FILE_ARGS="$FILE_ARGS $1"; shift; done ;;
         --rebuild-index) REBUILD_INDEX_FLAG="--rebuild-index"; shift ;;
         --no-index) NO_INDEX_FLAG="--no-index"; shift ;;
-        --test) TEST_FLAG=1 ;;
+        --fetch-tiles) shift; while [ $# -gt 0 ] && [[ ! "$1" =~ ^- ]]; do FETCH_TILES_ARGS="$FETCH_TILES_ARGS $1"; shift; done ;;
+        --test) TEST_FLAG=1; shift ;;
+        --serve)
+            SERVE_FLAG=1; shift
+            if [ $# -gt 0 ] && [[ ! "$1" =~ ^- ]]; then SERVE_PORT="$1"; shift; fi
+            ;;
         -h|--help|-help)
             echo "Pipeline LiDAR Archéologique"
             echo ""
@@ -141,14 +157,20 @@ while [ $# -gt 0 ]; do
             echo "  --force-classification"
             echo "                  Reclassifier le sol même si le fichier .las existe"
             echo "  --keep-tif      Conserver les TIFF pour régénérer les AVIF"
-            echo "  --ground-classification {auto,smrf,csf}"
+            echo "  --ground-classification {auto,ign,smrf,csf}"
-            echo "                  Méthode de classification du sol (défaut: auto)"
+            echo "                  Méthode de classification du sol (défaut: auto = pré-classification IGN)"
+            echo "  --ign-classes CLASSES"
+            echo "                  Classes LAS pour le MNT IGN : sol,unclassified,eau,virtuel,pont,sursol"
+            echo "                  ou codes (ex: 2,1). Mode pur, aucune retouche. (défaut: sol)"
             echo "  --quality N     Qualité image 1-100 (défaut: 98, 100=lossless)"
             echo "  --lossless      Compression lossless (équivalent à --quality 100)"
             echo "  --format FMT    Format de sortie : avif (défaut) ou webp"
             echo "  --only VIZ...   Générer uniquement ces visualisations"
             echo "  --skip VIZ...   Exclure ces visualisations"
             echo "  --file NOM...   Traiter un ou plusieurs fichiers LAZ"
+            echo "  --fetch-tiles COL,ROW..."
+            echo "                  Télécharger ces dalles depuis l'IGN puis les traiter"
+            echo "  --serve [PORT]  Servir la carte + API de génération de zones (défaut: 8973)"
             echo "  --test          Exécuter les tests unitaires"
             echo "  -h              Afficher cette aide"
             echo ""
@@ -181,6 +203,13 @@ done
 
 # Trim FILE_ARGS whitespace
 FILE_ARGS=$(echo "$FILE_ARGS" | xargs)
+FETCH_TILES_ARGS=$(echo "$FETCH_TILES_ARGS" | xargs)
+
+# input/ en lecture seule sauf si --fetch-tiles (téléchargement IGN dans input/)
+INPUT_RO=":ro"
+if [ -n "$FETCH_TILES_ARGS" ]; then
+    INPUT_RO=""
+fi
 
 # Check for --test flag first
 if [ "$TEST_FLAG" -eq 1 ]; then
@@ -207,6 +236,31 @@ fi
 # Créer les répertoires s'ils n'existent pas
 mkdir -p "$INPUT_DIR" "$OUTPUT_DIR"
 
+# Mode --serve : carte + API de génération (input/ en écriture)
+if [ "$SERVE_FLAG" -eq 1 ]; then
+    echo "============================================"
+    echo " Carte LiDAR + API de génération"
+    echo "============================================"
+    echo "  http://127.0.0.1:${SERVE_PORT}/"
+    echo "  GPU        : $([ -n "$GPU_FLAG" ] && echo 'OUI' || echo 'non')"
+    echo "  Bouton « + Zone » : dessiner un rectangle → téléchargement IGN + rendu"
+    echo "============================================"
+    GPU_ENV_FLAG=""
+    [ -n "$GPU_FLAG" ] && GPU_ENV_FLAG="-e LIDAR_GPU=1"
+    if [ -n "$GPU_ARG" ] && [ "$GPU_ARG" != "all" ]; then
+        GPU_ENV_FLAG="-e LIDAR_GPU=1 -e CUDA_VISIBLE_DEVICES=${GPU_ARG} --gpus '""device=${GPU_ARG}""'"
+    fi
+    exec docker run --rm --init $GPU_FLAG $CUDA_ENV_FLAG \
+        --user 1000:1000 \
+        -p "${SERVE_PORT}:8973" \
+        -v "${INPUT_DIR}:/data/input" \
+        -v "${OUTPUT_DIR}:/data/output" \
+        -e LIDAR_INPUT_DIR=/data/input -e LIDAR_OUTPUT_DIR=/data/output \
+        $GPU_ENV_FLAG \
+        "$IMAGE_NAME" \
+        python3 -m uvicorn lidar_pipeline.webapp:app --host 0.0.0.0 --port 8973
+fi
+
 # Lancer le pipeline
 echo "============================================"
 echo " Pipeline LiDAR Archéologique"
@@ -220,7 +274,7 @@ echo "  Force classif.: $([ -n "$FORCE_CLASSIFY_FLAG" ] && echo 'OUI' || echo 'n
 echo "  Keep TIFF    : $([ -n "$KEEP_TIF_FLAG" ] && echo 'OUI' || echo 'non')"
 echo "  Qualité image : $([ -n "$QUALITY" ] && echo "$QUALITY" || echo '98')"
 echo "  Format      : $([ -n "$FORMAT_FLAG" ] && echo "${FORMAT_FLAG#--format }" || echo 'avif')"
-echo "  Classification sol : $([ -n "$GROUND_METHOD" ] && echo "$GROUND_METHOD" || echo 'auto')"
+echo "  Classification sol : $([ -n "$GROUND_METHOD" ] && echo "$GROUND_METHOD" || echo 'auto')$([ -n "$IGN_CLASSES_FLAG" ] && echo " ${IGN_CLASSES_FLAG#--ign-classes }")"
 if [ -n "$ONLY_FLAG" ]; then
     echo "  Visualisations: uniquement${ONLY_FLAG#--only}"
 elif [ -n "$SKIP_FLAG" ]; then
@@ -229,6 +283,9 @@ fi
 if [ -n "$FILE_ARGS" ]; then
     echo "  Fichiers   :${FILE_ARGS}"
 fi
+if [ -n "$FETCH_TILES_ARGS" ]; then
+    echo "  Dalles IGN :${FETCH_TILES_ARGS} (input/ monté en écriture)"
+fi
 echo "============================================"
 
 CMD_ARGS="-o /data/output -r $RESOLUTION -w $WORKERS $VERBOSE_FLAG $FORCE_FLAG $FORCE_CLASSIFY_FLAG $KEEP_TIF_FLAG $QUALITY $FORMAT_FLAG"
@@ -241,6 +298,9 @@ fi
 if [ -n "$GROUND_METHOD" ]; then
     CMD_ARGS="$CMD_ARGS --ground-classification $GROUND_METHOD"
 fi
+if [ -n "$IGN_CLASSES_FLAG" ]; then
+    CMD_ARGS="$CMD_ARGS $IGN_CLASSES_FLAG"
+fi
 if [ -n "$ONLY_FLAG" ]; then
     CMD_ARGS="$CMD_ARGS $ONLY_FLAG"
 fi
@@ -256,6 +316,9 @@ fi
 if [ -n "$NO_INDEX_FLAG" ]; then
     CMD_ARGS="$CMD_ARGS $NO_INDEX_FLAG"
 fi
+if [ -n "$FETCH_TILES_ARGS" ]; then
+    CMD_ARGS="$CMD_ARGS --fetch-tiles $FETCH_TILES_ARGS"
+fi
 
 # Build CUDA_VISIBLE_DEVICES env var from GPU_ARG
 CUDA_ENV_FLAG=""
@@ -271,7 +334,7 @@ fi
 
 docker run --rm --init $GPU_FLAG $CUDA_ENV_FLAG \
     --user 1000:1000 \
-    -v "${INPUT_DIR}:/data/input:ro" \
+    -v "${INPUT_DIR}:/data/input${INPUT_RO}" \
     -v "${OUTPUT_DIR}:/data/output" \
     "$IMAGE_NAME" \
     python3 -m lidar_pipeline /data/input \