Spartina France : annotated imagery, processing code and trained models for invasive Spartina mapping

Résumé fourni par la source

SUMMARY This deposit provides a complete, reproducible package for mapping invasive Spartina (Spartina spp.) along the French Channel and Atlantic coasts using 20 cm Color Near-Infrared (CIR) aerial imagery. It combines an annotated multi-site aerial dataset, a 31-script Python processing pipeline, trained SegFormer semantic segmentation model weights, and spatial validation protocols. REPOSITORY STRUCTURE ├── folds.json # Spatial partitioning protocols for validation├── data_dictionary.csv # Complete attribute definitions for all tile indices├── index.zip # Spatial indices and tile metadata├── labels.zip # Primary photo-interpretation vector polygons (AOI & labels)├── spartina-france-dataset-v1.zip # Georeferenced GeoTIFF tiles, labels, and metadata (17 sites)├── spartina-france-toolbox-v1.zip # Complete processing codebase, scripts, and environment files└── spartina-france-models-v1.zip # Deployable models, CV folds, benchmarks, and MODEL_CARD.md ARCHIVE BREAKDOWN index.zip: Contains spatial index files: * index_tuiles.csv: 7,221 annotated tiles across 12 sites. * index_negatif_prepare_exported.csv: 10,072 exported negative-pool tiles. * index_negatif_full_pool.csv: 49,361 tiles (full negative pool including non-distributed imagery, with false-detection scores, spatial blocks, and selection status). labels.zip: Digitised area-of-interest (AOI) and label polygons representing primary photo-interpretation data. spartina-france-dataset-v1.zip: 17 study sites — 12 annotated sites with Spartina present (photo-interpreted into three cover classes) and 5 verified Spartina-free control sites selected for spectral confounders (such as algal mats and vegetated rocky shores). Includes georeferenced GeoTIFF image and label tiles plus acquisition metadata. spartina-france-toolbox-v1.zip: Complete processing chain consisting of 31 scripts (acquisition, index/label creation, training, cross-validation, large-mosaic inference, and diagnostics) documented in SCRIPTS.md, alongside environment.yml and requirements.txt. spartina-france-models-v1.zip: Deployable production model, five cross-validation fold models, evaluation benchmarks, and MODEL_CARD.md (which maps excluded spatial zones per fold and documents the full benchmark performance table). TECHNICAL SPECIFICATIONS & BASELINE PERFORMANCE Source Imagery: BD ORTHO, Institut national de l'information géographique et forestière (IGN) Licence: Licence Ouverte Etalab 2.0 Spatial Resolution: 20 cm Ground Sampling Distance (GSD) Spectral Composition: 3-band Colour Near-Infrared (CIR: Near-Infrared, Red, Green) Problem Formulation: Binary semantic segmentation (Dense Spartina vs. Background) Model Architecture: SegFormer Cross-Validated Mean mIoU: 0.6052 False Detection Rate (Held-out block): 1.38% (Cross-validation mean) / 0.80% (Production model) CRITICAL REPRODUCIBILITY NOTES Spatial Partitioning Protocol (folds.json): Since individual tile index files do not carry a fold-assignment column, folds.json explicitly defines all spatial partitioning protocols (5-fold cross-validation, negative-pool spatial blocks, and the production model's internal validation split). Deterministic Inference Requirements: PyTorch and cuDNN are configured for deterministic algorithm selection within the inference loader (spartina-france-toolbox-v1.zip). This flag must remain enabled; non-deterministic GPU execution was observed to produce materially divergent segmentation outputs on identical imagery. KNOWN LIMITATIONS 1. Annotation Confidence: Heterogeneous confidence levels across sites, recorded at the individual tile level.2. AOI Framing Bias: Digitised areas of interest were framed around known Spartina stands, leading to an under-representation of Spartina-free salt marsh habitats in the annotated set.3. Negative-Pool Selection: Composed of a targeted selection of spectral confounders rather than a uniform random spatial sample.4. Radiometric Sensitivity: Delineation performance shows a demonstrated dependence on absolute image radiometric levels. See README.md inside this deposit for full detail, and the associated Data in Brief article for the complete methodology.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.