1 · Landscape
Four paper study areas (configs/study_areas.yaml)
2 · Scene
Preprocessed composites written by scripts/acquire_study_area.py (temporal median, 10 m, UTM)
3 · Segmentation model
Trained checkpoints under outputs/segmentation — metrics are agreement with the GMW reference map on held-out tiles
Result label: DEVELOPMENT — not final (runs from the UI are never labelled final)MMU 2 ha · k = 3 · τ = 5 km (configs/graph.yaml)
Run
Runs as a background job on the server — this page follows its progress
Landscape · Kerala Coast
Scene · —
Model · —
Threshold · 0.5
Sign in to run analyses (GIS officer, analyst or admin).
What is real here
• Scenes: Copernicus Sentinel-1 RTC (Planetary Computer) and Sentinel-2 L2A (Earth Search), temporal medians for the year.
• Labels used for training: Global Mangrove Watch v3 2020 (weak labels) — the model learns a reference map, not field truth.
• Development checkpoints (B0):
multi_* models were trained on all four study areas (pooled 2020 tiles); kerala_* models on Kerala only. The 4-area model maps Sundarbans and Bhitarkanika well and Kerala / Gulf of Mannar poorly (thin mangrove fringes).• Graph, indices and criticality follow the paper exactly (Eqs. 3–11); costs are never invented, so restoration is ranked by connectivity gain unless you upload costs.