Training strategy over architecture: A systematic ablation study for Spartina detection from aerial imagery with small geospatial datasets : Ablation Results

Comprehensive Ablation Study Results for Spartina alterniflora Detection from Aerial Imagery: SegFormer-B2, U-Net (ResNet-34), and DeepLabV3+ (ResNet-34) Description This dataset provides the complete and reproducible results of a systematic ablation study accompanying the research article: Le Guillou, A., Stéphan, P., Dauvergne, X., Ammann, J., Brehier, M. (2026). Training Strategy Over Architecture: A Systematic Ablation Study for Spartina Detection from Aerial Imagery with Small Geospatial Datasets. Submitted to Remote Sensing of Environment. Context & Methodology The study evaluates the influence of four critical training strategy axes—Data Augmentation, Loss Function, Regularization, and Optimization—across three state-of-the-art segmentation architectures (U-Net, DeepLabV3+, and SegFormer-B2). The experiments were conducted for the automated detection of Spartina alterniflora using BD ORTHO CIR (20 cm GSD) and LiDAR HD nDSM imagery in the Bay of Brest (Brittany, France). To ensure statistical robustness, a total of 180 training runs (12 configurations × 3 architectures × 5-fold cross-validation) were performed and analyzed. Dataset Content & Structure The repository is organized by architecture: /SegFormer, /UNet, and /DeepLabV3. Within each folder, the data is structured as follows: 1. Raw Logs and Global Summaries ablation.txt: Comprehensive log summarizing all training runs, including per-fold metrics and hyperparameter details. progress.log: Chronological training log with precise timestamps for reproducibility. 2. Statistical Analysis Subfolders Each directory (analysis_augmentation/, analysis_loss/, etc.) contains: Statistical Reports (.json): Detailed analysis including Friedman tests, pairwise Wilcoxon tests, Cohen's $d$ effect sizes, and per-fold metric distributions. Summary Tables (.csv): Condensed cross-validation results (Mean $\pm$ Std Dev for IoU, F1-score, Precision, Recall, and Overall Accuracy). Manuscript Ready Tables (.tex): LaTeX-formatted tables optimized for direct inclusion in scientific publications. 3. Visualizations (.png) boxplots_comparison.png: Distribution of IoU and F1 metrics across different strategies. effect_size_heatmap.png: Cohen's $d$ heatmap for pairwise strategy comparison. fold_by_fold_iou.png: Evolution of IoU across the 5 folds. critical_difference_*.png: Nemenyi post-hoc test diagrams visualizing statistical significance. Technical Specifications Total File Count: ~108 files. Data Format: JSON (Stats), CSV (Tables), TXT (Logs), PNG (Visuals), TEX (LaTeX). Model Weights: Due to the significant volume (over 150 GB), the 180 pre-trained .pth model weights are not included in this deposit. They are available upon reasonable request to the corresponding author ([email protected]) via institutional file transfer. Licensing and Reuse This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0). Users are free to share and adapt the material, provided that appropriate credit is given to the original authors and the associated research article is cited. Related Identifiers Imagery Source (BD ORTHO): https://geoservices.ign.fr/bdortho Elevation Data (LiDAR HD): https://geoservices.ign.fr/lidarhd Software Framework: https://github.com/aleguillou1/SemanticSeg4EO Keywords: Semantic segmentation, Ablation study, Spartina alterniflora, Deep learning, Coastal vegetation, Remote sensing, U-Net, DeepLabV3+, SegFormer, Transfer learning, Small dataset, Bay of Brest. Data Availability Statement (for Manuscript) The complete ablation results (per-fold metrics, statistical reports, visualizations, and summary tables) for all 180 training configurations across the three architectures are publicly available on Zenodo at 10.5281/zenodo.18836432. Trained model weights (.pth files, cumulative volume > 150 GB) are available upon reasonable request to the corresponding author. The BD ORTHO CIR orthoimagery and LiDAR HD point cloud data used in this study are freely accessible from the IGN Geoservices platform (https://geoservices.ign.fr).