Bangladesh floods every monsoon. Existing systems rely on manual satellite interpretation, cloud-obscured optical imagery, and coarse delayed outputs. FloodGuard-BD closes the gap — Sentinel-1 SAR to pixel-level flood map, with calibrated confidence at every pixel.
The 2024 monsoon inundated more than 40% of the country. Millions displaced annually. Agriculture and infrastructure losses in the billions. These are not projections — they are last year's record.
FFWC and UNOSAT rely on optical imagery that clouds make unreliable precisely when floods are forming. Manual tracing misses the window for pre-emptive resource deployment.
The data exists. Sentinel-1 SAR passes over Bangladesh regardless of cloud cover. The gap is the intelligence layer that converts those passes into actionable maps within minutes — with a confidence score per pixel so disaster managers know where to trust the output.
"Every dollar of infrastructure translates directly into flood maps that reach disaster managers before the water does."
NASA and IBM's geospatial ViT masked autoencoder, pre-trained on global HLS Sentinel-2 data. We extend patch projection layers to ingest Sentinel-1 SAR alongside Sentinel-2 bands and HAND terrain derivatives — a multimodal stack that sees through monsoon cloud cover.
Standard softmax gives overconfident predictions in ambiguous regions — urban water versus shadow, submerged vegetation. A Dirichlet distribution head jointly predicts flood probability and calibrated uncertainty: both aleatoric (data noise) and epistemic (model ignorance).
Final calibration with isotonic regression on held-out UNOSAT maps achieves ECE <0.03. Uncertainty maps correctly flag regions where annotator agreement is also lower — confirming the uncertainty head is meaningful, not decorative.
Disaster managers get a trust score per pixel. That is the difference between deploying resources and waiting.
| Model | Dataset | Dice |
|---|---|---|
| Baseline U-Net (threshold) | Gaibandha 2020 | 0.781 |
| Prithvi-100M fine-tuned (pilot) | Gaibandha 2020 + synthetic weak labels | 0.830 |
| Target — full pipeline | 2019–2024 monsoon seasons | >0.850 |
Haor region inundation — validated against SPARRSO official records. Worst-in-decades event.
Brahmaputra floodplain — primary training and validation set. Char territory inundation.
Most recent monsoon event — current benchmark against DDM field records.
Sentinel-1 GRD + Sentinel-2 L2A acquisition, Lee speckle filtering, radiometric terrain flattening, dynamic 512×512px tiling with 256px overlap, COG-format export. Non-commercial tier, $0.
Raw COG tile archive with Standard→Coldline→Archive lifecycle policies, public flood map output bucket, model artifact registry. Cost-optimized for 3 TB scale.
Single NVIDIA T4 spot instance, ~2 hours per training run, TFRecord streaming from Cloud Storage. Approximately $12 per run at spot pricing.
T4 prediction endpoint, ONNX Runtime inference, autoscaling. Real-time inference triggered per new Sentinel-1 scene via Pub/Sub.
New Sentinel-1 scene notifications trigger automated pipeline. Cloud Function calls GEE Python API, exports tiles, fires Vertex inference, writes output GeoTIFF and Firestore metadata.
Next.js dashboard with Leaflet map, temporal slider, uncertainty toggle, district statistics. BigQuery stores active learning corrections for weekly retraining cycle.
Weekly Cloud Build trigger for active learning retraining cycle. Corrections from human validation feed back into the model — adapting to evolving flood patterns without manual relabeling.
All code is released under Apache 2.0. A Colab demo notebook ships with the repository. The system is designed for handover to FFWC, the Department of Disaster Management, SPARRSO, or UN OCHA at zero licensing cost.
Partnership discussions with SPARRSO — Bangladesh's national space research agency — are already underway. The validation methodology is explicitly calibrated against their official flood event records.