Mangrove classification dataset and reproducible workflow, Mandeh coast, West Sumatra

This deposit contains the data, code, and outputs for a binary mangrove / non-mangrove classification of a coastal area on the Mandeh coast, West Sumatra, Indonesia. Contents: Data/feature_stack.tif: 12-band raster stack, 455 x 689 pixels, EPSG:4326, pixel size about 10 m. Bands: B2, B3, B4, B8, B11, B12 (Sentinel-2 surface reflectance), DEM (DEMNAS elevation), and five indices derived from the bands (MVI, NDVI, EVI, NDMI, NDWI). All layers were rescaled to [0, 1] over the whole scene before the stack was written. Data/reference_points.shp (with .dbf, .shx, .prj, .cpg): 2,500 reference points. KELAS is the binary label (1 = mangrove, 0 = non-mangrove). GRID_CODE is the sampling stratum (0-3 non-mangrove, 4 mangrove), with 500 points per stratum. POINTID is the unique identifier. Mangrove_mandeh_classification_user.ipynb: the analysis notebook with stored outputs. Set the three paths in the configuration cell before running. results/: classification maps for Random Forest, SVM (RBF), and XGBoost; the calibrated SVM probability map; model performance, hyperparameters, SHAP contributions and fold stability, ablation results; and the reference table with extracted feature values. Method summary: points with missing feature values were removed, leaving 1,508 of 2,500 (the missing values all come from DEM NoData over water). Reference points were grouped into 40 spatial blocks with k-means, and all cross-validation used these blocks (StratifiedGroupKFold). Hyperparameters were tuned with nested cross-validation (5 outer, 3 inner folds); XGBoost parameters were searched with Optuna. The final models were evaluated on a held-out spatial partition. Random seed: 42. Requirements are listed in requirements.txt. The full notebook takes roughly 40 minutesto run on the authors' machine.