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

Spartina France : annotated imagery, processing code and trained models for invasive Spartina mapping Version 2 (October 2026) This record provides a reproducible package for mapping invasive Spartina along the French Channel and Atlantic coasts from 20 cm colour near-infrared aerial imagery (BD ORTHO IRC, EPSG:2154). It contains an annotated multi-site dataset, the processing pipeline, trained SegFormer segmentation models, the spatial validation protocol, and the results of every configuration tested during the study. What is new in version 2 The weights of all benchmark runs are now distributed, and seven training runs were added (seed replicates and loss composition ablation). The study comprises 43 training runs, listed and described in MODEL_CARD.md. New result tables in benchmark_results/: false detection scores on the negative pool test block, response of the false detection to a radiometric gain, evaluation by acquisition under three normalisation modes, and summary tables (results per site, by substrate and by confidence tier). Nine scripts were added to the toolbox for these experiments, and the default checkpoint of the inference configuration is now best_val_miou.pt. The labelling confidence of three sites was corrected in the index files (Pedel and Pont_Callec are high, Quiberon is moderate), to match the article. The sampling weights used for training were not recomputed, and no model was retrained. The tiles, the labels and the fold definitions are unchanged. The complete list of changes is in CHANGELOG.md. Contents File Content spartina-france-dataset-v2.zip Image and label tiles (256 by 256 pixels, 20 cm) of the annotated sites and of the negative pool, with the acquisition metadata of each site labels.zip Cleaned digitised label and area-of-interest vectors index.zip Tile indices: index_tuiles.csv (7,317 annotated tiles), index_negatif_full_pool.csv (the 49,361 tiles of the negative pool) and index_negatif_prepare_exported.csv (the 2,026 training tiles and the 8,046 test block tiles of the pool, with the paths of their files) folds.json Definition of the five spatial folds and of the spatial blocks of the negative pool spartina-france-models-v2.zip production/ (deployable model), cross_validation/ (the five models of the retained recipe, one per fold), benchmark_results/ (the other training runs with their weights, and the result tables), logs_experiments/ (logs of the revision experiments) and MODEL_CARD.md spartina-france-toolbox-v2.zip Pipeline scripts for data preparation, training, inference and the experiments, with SCRIPTS.md, environment.yml and requirements.txt flight_dates_BDORTHO.xlsx Acquisition dates of the imagery data_dictionary.csv Description of every column of the index files and of the result tables CHANGELOG.md Changes since version 1 Where to start To map new imagery, use production/best_val_miou.pt with the inference scripts of the toolbox (executer_inference.py, configured in config_inference.py). To evaluate on a given zone, use the cross-validation model that held it out. The correspondence between folds and zones is given in folds.json and in MODEL_CARD.md. The production model has seen every zone and cannot provide an unbiased evaluation. To follow the whole chain, read SCRIPTS.md. Performance SegFormer with a MiT-B2 encoder initialised from ImageNet, binary target (dense Spartina against the rest), trained with a combined cross entropy and Tversky loss, radiometric gain augmentation and a negative tile pool. Mean over the five spatial folds, each evaluated on zones never seen during training: mIoU Dense Spartina IoU Recall Precision False detection, negative pool test block 0.6052 0.4087 0.5603 0.6046 1.38 % On the same negative pool test block, the production model reaches 0.80 % of falsely detected area with its best-mIoU checkpoint, against 4.74 % with its best-validation-loss checkpoint. The best-mIoU checkpoint is the one recommended. The fold pli_4 (Arcachon and Canche) gives the weakest result in every configuration tested. Known limitations The annotation confidence varies from site to site, and so does the measured performance. The confidence tier of each site is given in index_tuiles.csv. The areas of interest were framed around known Spartina stands. The negative pool is a targeted selection of Spartina-free sites, not a random sample of the coast. The performance depends on the absolute radiometric level of the imagery. The training applies a gain augmentation (0.45 to 2.0), and the response of the false detection to the gain is documented in benchmark_results/radiometric_gain/. Licence The record is distributed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). The code of the toolbox carries its own licence file (LICENSE_CODE_MIT.txt).