feat: SAR-based flood detection

config.yaml

@@ -75,6 +75,27 @@ analysis_types:
       - "flooded_area_km2"
       - "mndwi_stats"
 
+  # Flood Detection (SAR / Sentinel-1) -- all-weather, ensemble-based, no trained
+  # weights required. Separates permanent water from flood given a JRC GSW
+  # reference or a pre-event scene.
+  flooding_sar:
+    enabled: true
+    display_name: "Flood Detection (SAR)"
+    description: "All-weather flood detection from Sentinel-1 VV/VH using a physics-based ensemble"
+    model:
+      architecture: "ensemble"   # LIST + DLR + TUW majority vote; no neural weights
+      in_channels: 2             # VV, VH
+      num_classes: 3
+    bands: ["VV", "VH"]
+    classes: ["dry_land", "permanent_water", "flooded"]
+    thresholds:
+      alert_flood_area: 5.0
+      critical_flood_area: 20.0
+    metrics:
+      - "flooded_percentage"
+      - "flooded_area_km2"
+      - "permanent_water_km2"
+
   # Drought Monitoring
   drought:
     enabled: false  # Not yet implemented

config/train_flood.yaml

@@ -0,0 +1,59 @@
+# ============================================================
+# ClimateVision — Flood Detection Training Config
+# ============================================================
+
+# --- Data --------------------------------------------------
+data:
+  dir: data/processed/flood
+  image_size: 256
+  batch_size: 8
+  num_workers: 4
+  use_weighted_sampler: true
+  pin_memory: true
+
+# --- Model -------------------------------------------------
+model:
+  architecture: flood_unet_s2only
+  in_channels: 3
+  num_classes: 3
+  encoder: efficientnet-b7
+
+# --- Loss --------------------------------------------------
+loss:
+  type: combined
+  focal_weight: 0.5
+  focal_alpha: 0.25
+  focal_gamma: 2.0
+  use_class_weights: true
+
+# --- Optimiser --------------------------------------------
+optimizer:
+  learning_rate: 1.0e-4
+  weight_decay: 1.0e-4
+  min_lr: 1.0e-6
+
+# --- Schedule ---------------------------------------------
+schedule:
+  epochs: 20
+  warmup_epochs: 3
+  checkpoint_interval: 5
+
+# --- Regularisation / Tricks ------------------------------
+training:
+  mixed_precision: true
+  grad_clip: 1.0
+  use_ema: true
+  ema_decay: 0.99
+  early_stopping_patience: 10
+
+# --- Outputs ----------------------------------------------
+output:
+  save_dir: models
+  run_name: ""
+
+# --- Normalisation stats ----------------------------------
+normalizer_stats: ""
+
+# --- Analysis type ----------------------------------------
+analysis:
+  type: flooding