HimalayaSentinel-AI: Spaceborne Flood Damage & Evacuation Intelligence

Team: Anotics Hackathon: IIT Mandi Multimodal AI Hackathon 2026 — Track B: Mapping Flood Damage from Space

1. Inspiration

Floods in mountainous regions can be especially dangerous because steep terrain, river valleys, and limited road connectivity make emergency response difficult. Knowing where floodwater may be present is only one part of the problem. Emergency teams also need to understand which settlements have road access and how nearby hospitals or relief facilities can be reached.

This inspired us to build HimalayaSentinel-AI, a geospatial disaster-analysis system focused on the Trishuli–Bhote Koshi corridor in Nepal. Our goal is to bring satellite-based flood analysis, road-network mapping, settlement accessibility, and emergency situation reporting together in one interactive dashboard.

2. What We Built

HimalayaSentinel-AI combines three main components:

• Satellite-based AI: A deep-learning model that classifies land pixels as dry land, permanent water, or flood inundation.
• Infrastructure and accessibility analysis: A road-network graph used to analyze connectivity between settlements and emergency facilities.
• Emergency dashboard and reporting: An interactive map and a one-page Emergency Situation Report that summarize the available analysis and its limitations.

The study area covers approximately 68,317 hectares, within the coordinates 85.10°–85.35° E and 27.90°–28.15° N. Our case study concerns the reported flood event of August 26, 2026.

3. How the AI Model Works

We use Sentinel-1 Synthetic Aperture Radar (SAR) imagery from before and after the event, together with Copernicus WorldDEM elevation data.

The model receives five input channels:

1. Pre-event radar VV
2. Pre-event radar VH
3. Post-event radar VV
4. Post-event radar VH
5. Elevation from the digital elevation model

We use a ResUNet-18 segmentation architecture. In simple terms, the model examines each pixel and estimates whether it belongs to dry land, permanent water, or flood inundation.

The model was trained and evaluated using the Kuro Siwo flood benchmark. It achieved a validation Flood IoU of 73.66% and a held-out test Flood IoU of 68.82%. These are benchmark results, not verified accuracy measurements for Nepal.

4. Road Networks and Emergency Accessibility

We use OpenStreetMap data to construct a road-network graph for the study area. The current network contains 161 mapped highway ways and 5,481 graph segments.

Using Dijkstra's shortest-path algorithm, our system calculates routes through the mapped network and evaluates road connectivity between six settlements and emergency facilities, including hospitals, relief facilities, and a helipad.

Five settlements are connected in the current road graph. Kaptali is disconnected because no connecting vehicular road exists in the baseline network. We do not attribute this disconnection to flooding.

The current overlay identifies zero flood-induced road closures. Because the Nepal flood mask is unverified, this must not be interpreted as proof that all roads are safe or that no flood-related disruption occurred.

5. Interactive Dashboard and Situation Report

Our dashboard organizes the analysis into five views:

• Overview: Study-area context, key statistics, mapped roads, facilities, and evacuation routes.
• Flood AI: Benchmark metrics, model predictions, input details, and the Nepal inference limitations.
• Infrastructure: Road-network statistics, passability overlays, bridges, and infrastructure assessment status.
• Settlements: Settlement connectivity and routes to nearby emergency facilities.
• Data & Ethics: Data sources, model-estimate disclosures, ground-truth availability, and attribution requirements.

We also generate a one-page Emergency Situation Report summarizing benchmark performance, Nepal inference results, road-network analysis, settlement accessibility, and operational cautions.

6. Challenges and Limitations

Applying a model trained on one geographic environment to another is difficult. Our Nepal inference illustrates this challenge.

The standard raw-argmax prediction produced zero flood pixels. Our data audit also found that approximately 80% of post-event VV pixels reached the preprocessing clipping ceiling, indicating substantial radar backscatter saturation. Steep Himalayan terrain and differences in elevation and spatial resolution may contribute to the model's difficulty.

We display an exploratory candidate mask of approximately 2.6 hectares using a probability threshold of 0.20. This mask has not been verified against independent ground truth and is not presented as a confirmed flood extent.

Independent local ground-truth flood polygons were not available in our analysis environment, so we cannot report verified Nepal-specific precision, recall, or IoU. Structural building damage detection is also not implemented.

We consider these limitations important findings rather than hiding them. In a disaster-response setting, clearly distinguishing model estimates from confirmed observations is essential.

7. What We Learned and What's Next

This project helped us bring together deep-learning segmentation, satellite data, geospatial processing, graph algorithms, and interactive visualization into a single workflow.

Our next steps are to improve terrain-aware radar preprocessing, investigate domain adaptation for mountainous regions, and validate predictions against reliable local reference data. Further work could also incorporate verified infrastructure-damage observations and more comprehensive emergency-response information.

Our vision is to build a transparent decision-support tool that helps responders investigate potential flood impacts, understand mapped accessibility, and identify where additional verification is needed.

HimalayaSentinel-AI — Team Anatics: Combining satellite intelligence, AI, and geospatial analysis for more informed disaster-response planning.

Built With

• computervision
• data
• deeplearning
• geospatial
• openstreetmap
• python
• pytorch