End-to-end pipeline diagram: Sentinel-2 RGB (visible bands) + Sentinel-2 NIR (water-sensitive) → Dual encoders (self-attention per branch) → Cross-attention (fused at every skip scale) → Flood extent masks (zero-shot sensor transfer) — Pipeline overview

Situation

Disaster teams need flood maps to plan routes and supplies. However, manual GIS work, which maps and analyses locations, takes days or weeks. That delay leaves little time during an emergency’s first 72 hours. The 270 km Citarum River basin in West Java houses more than 25 million people. Therefore, it needed accurate automated flood maps across cities, farms, and forests despite partial cloud.

The Challenge

U-Net models using only RGB colour images often confuse flood water with dark soil and shadows. Near-infrared (NIR) bands from Sentinel-2 help distinguish them. However, the model must combine information about light wavelengths and spatial patterns. Joining RGB and NIR only at the input leaves their features misaligned later. This happens at skip connections, which pass spatial detail to the decoder that builds the map.

The Approach

I designed ProCANet, a U-Net with separate RGB and NIR encoders that extract image features. At every skip connection, cross-attention combines features from both inputs. It works at 4 scales, 1/2, 1/4, 1/8, and 1/16. First, self-attention captures long-range patterns within each branch. This keeps coherent regions with different wavelength responses from interfering before the branches combine.

I trained ProCANet on Sen1Floods11, which contains Sentinel-2 images from 11 floods worldwide. Moreover, I tracked every experiment with Weights & Biases. These records included parameter sweeps, loss curves, and checkpoints showing each class’s intersection over union (IoU). IoU measures how much the predicted and reference regions overlap.

I tested the model on PlanetScope images without retraining. PlanetScope pixels represented 3 m on the ground, compared with 10 m in the training images. Across a 6,112 km² holdout area, the model reached 0.659 IoU. The reference labels came from NDWI, a water index based on light reflected at different wavelengths. This test reflects deployment across different training and inference sensors.

I compared ProCANet with UNet, PSPNet, LinkNet, MANet, PAN, and ConvNeXt V2. Each model used the same Sen1Floods11 test protocol. ProCANet reached 0.898 F1, which balances precision and recall, and 0.815 IoU. By comparison, UNet reached 0.884 F1 and 0.791 IoU. These gains matter on a benchmark already close to its current performance ceiling.

Impact

The system can support disaster response for more than 25 million people in the Citarum basin. It reduces days of manual GIS work to hours after a satellite pass. Therefore, Indonesia’s disaster agency, BNPB, can use earlier flood maps to plan evacuation routes and stage supplies.

0.898 f1 score 0.815 iou 0.659 zero shot iou 0.884 baseline f1 0.791 baseline iou 6112 coverage area km2 95.0 accuracy pct

Tech Stack

PyTorchPyTorch LightningGoogle Earth EngineSentinel-2PlanetScopeWeights & Biases (W&B)GeoTIFFmultispectral rasterNDWI

Collaborators

Monash University, Primary research host (A/Prof Risqi U. Saputra)
University of Indonesia, Domain validation

← All Use Cases View Publication ↗