-
-
Poster
Inspiration
In the Himalayan region, floods, landslides, and extreme rainfall can isolate entire communities within hours. Damaged roads, inaccessible bridges, difficult terrain, and limited ground-level information make disaster response particularly challenging.
Satellite imagery offers valuable information after disasters, but converting raw observations into actionable geographic evidence requires specialized processing, infrastructure mapping, and careful interpretation.
We built HimaRoad + Flood Copilot to explore a fundamental question: Can satellite observations, geospatial analysis, and artificial intelligence work together to help responders understand where hazards occurred, which infrastructure may be exposed, and which settlements could become isolated?
Our motivation was to create an accessible, transparent, and evidence-grounded disaster-mapping prototype that communicates not only what satellite data reveals, but also what remains uncertain.
What it does
HimaRoad + Flood Copilot is an integrated satellite-based flood-mapping and disaster-impact screening platform designed for mountainous regions.
The system combines satellite imagery, geospatial analysis, road-network modeling, and a bilingual AI-assisted reporting interface.
Its primary capabilities include:
Satellite-based hazard mapping: Compares pre-event and post-event Sentinel-1 radar observations to identify potential new water and terrain disturbances, with Sentinel-2 optical imagery providing additional corroboration when available.
Infrastructure exposure assessment: Overlays detected hazards with historical OpenStreetMap data to identify potentially exposed roads, bridges, and buildings.
Cut-off settlement analysis: Constructs vehicle and pedestrian road networks to evaluate settlement connectivity under four hypothetical road-disruption scenarios.
Terrain-based flow tracing: Uses Copernicus DEM elevation data and D8 flow routing to trace downstream terrain paths and identify nearby settlements.
Interactive geospatial dashboard: Presents satellite-derived hazard layers, infrastructure exposure, settlement accessibility, and supporting evidence.
Bilingual Flood Copilot: Answers predefined, evidence-grounded questions in English and Nepali, with traceable claim references.
Automated situation reports: Generates one-page PDF reports in both languages, alongside geospatial outputs including GeoJSON and KML.
A central design principle is uncertainty-aware mapping. Areas hidden by radar shadow or terrain effects remain explicitly marked as unknown rather than incorrectly classified as unaffected.
The platform is an educational research prototype, not an operational early-warning or emergency-navigation system.
How we built it
We developed the platform as three interconnected components: HimaRoad, the geospatial processing engine; Flood Copilot, the evidence-grounded reporting service; and a unified web interface.
- Satellite data acquisition and preprocessing
We integrated Copernicus Data Space Ecosystem services to retrieve Sentinel-1 radar and Sentinel-2 optical imagery.
Sentinel-1 observations were paired using matching orbital tracks and processed using terrain-corrected gamma0 backscatter. All geospatial layers were aligned to a common 20-meter UTM grid.
We incorporated Copernicus GLO-30 elevation data to support terrain analysis, radar visibility masking, and downstream flow tracing.
- Hazard detection and deep learning
We implemented a rule-based radar change-detection pipeline and trained a UNet semantic segmentation model from scratch using the Kuro Siwo flood dataset.
Our selected model used four input channels representing pre-event and post-event VV/VH radar observations.
Training incorporated:
Binary Cross-Entropy + Dice loss
AdamW optimization
Four registered model variants
Development-based threshold selection and calibration
Held-out Himalayan evaluation
On the development dataset, the selected UNet achieved an IoU of 0.7656 and F1-score of 0.8672, outperforming the rule baseline.
However, independent testing on three Himalayan flood events revealed a significant generalization failure. We preserved and documented this result rather than presenting development accuracy as real-world performance.
- Infrastructure and network analysis
We integrated historical OpenStreetMap snapshots to construct directed vehicle and pedestrian networks.
Hazard intersections were used to flag potentially exposed infrastructure. Four hypothetical disruption scenarios evaluated how road-link removals could affect settlement connectivity to towns and healthcare facilities.
- Evidence-grounded Flood Copilot
We developed a FastAPI-based companion service that transforms saved geospatial outputs into structured, verifiable evidence claims.
Every claim includes provenance information such as its source file, method, and checksum.
The copilot uses deterministic question-answer templates to avoid generating unsupported statistics or unsafe navigation recommendations.
English and Nepali outputs share the same underlying evidence references.
- Dashboard and reporting
We integrated the mapping engine and copilot into a unified web application, with a Streamlit workbench for initiating new analyses.
The platform supports interactive visualization, bilingual situation-report generation, and geospatial data exports.
Challenges we ran into
Developing a satellite-based disaster-mapping system for Himalayan terrain introduced several substantial challenges.
- Complex mountainous terrain
Steep valley walls create radar layover and shadow, making some areas impossible to observe reliably.
In our Trishuli case study, 24.9% of the study area was unobservable, while 56.1% of road links had insufficient observational coverage.
We addressed this by explicitly tracking visibility and preserving unknown classifications.
- Generalization of deep learning models
Although our UNet performed well on development data, it failed to detect labelled new-water pixels in the independent Himalayan test.
This exposed the difficulty of transferring flood segmentation models across different terrain characteristics, event types, and satellite-observation conditions.
- Distinguishing exposure from confirmed damage
A road intersecting a satellite-derived hazard does not necessarily mean that it is destroyed or impassable.
We designed the system to distinguish potential exposure, observed change, and unknown operational status.
- Satellite revisit and data availability
Satellite acquisitions do not always align with disaster timing. Missing same-track observations can prevent reliable comparisons.
Rather than producing misleading outputs, our pipeline explicitly terminates incomplete analyses with documented reasons.
- Computational performance
Large urban areas, road-network construction, and satellite acquisition introduce substantial processing overhead.
Our worldwide robustness tests revealed performance bottlenecks, particularly in dense urban environments.
- Reliable bilingual communication
Communicating complex geospatial evidence in English and Nepali required consistent terminology, traceable statistics, and reliable PDF rendering.
We addressed this through predefined templates, claim-ID parity testing, and native-speaker review.
Accomplishments that we're proud of
Our most significant accomplishment is building an end-to-end, evidence-grounded geospatial analysis and reporting platform, connecting satellite acquisition with infrastructure screening and accessible disaster-information summaries.
Key achievements include:
Developed an integrated system combining satellite processing, deep learning, geospatial network analysis, and bilingual reporting.
Trained and evaluated a UNet flood segmentation model, achieving 0.7656 IoU and 0.8672 F1 on development data.
Conducted independent Himalayan testing and transparently documented the model's failure to generalize.
Implemented four hypothetical road-disruption scenarios to assess potential settlement isolation.
Built a provenance-aware copilot that links generated answers and reports to verifiable evidence claims.
Successfully validated 32 predefined bilingual question-answer cases, maintaining exact evidence-reference parity between English and Nepali.
Completed native-speaker review of the Nepali wording inventory, incorporating corrections.
Generated verified one-page bilingual situation reports for both Trishuli and Melamchi.
Achieved warm median UNet detector execution times of 4.929 seconds on CPU and 1.702 seconds on GPU for a 600 km² input, excluding data acquisition and other pipeline stages.
Conducted six worldwide robustness drills, documenting successful runs and explicit failure conditions.
Implemented reproducibility safeguards, including frozen case-study outputs, model configuration records, input checksums, and validation-data separation.
We are especially proud of prioritizing scientific transparency and responsible AI over unsupported performance claims.
What we learned
This project strengthened our understanding of the intersection between remote sensing, deep learning, geospatial intelligence, software engineering, and responsible AI.
Our most important learning was that strong machine-learning performance on development data does not guarantee successful deployment in geographically different environments.
The Himalayan evaluation demonstrated how terrain complexity, sensor limitations, domain shift, and spatial resolution can significantly affect flood-detection performance.
We also learned that reliable disaster-information systems require more than accurate predictions. They must preserve data provenance, communicate uncertainty, distinguish observations from assumptions, and avoid presenting potentially dangerous recommendations as verified facts.
Building the bilingual copilot highlighted the value of deterministic, evidence-grounded generation in high-stakes applications where unsupported numerical claims could be misleading.
Finally, integrating satellite services, raster processing, graph algorithms, backend APIs, and interactive visualization gave us practical experience designing a complex, multidisciplinary AI system.
What's next for flood copilot
Our next objective is to improve the system's reliability, geographic generalization, and practical usability while preserving its evidence-first design.
Planned directions include:
Improving Himalayan flood detection: Investigating terrain-aware learning, domain adaptation, improved training datasets, and better handling of narrow river valleys.
Strengthening multimodal fusion: Exploring more effective integration of Sentinel-1 radar, Sentinel-2 optical imagery, and elevation-derived terrain features.
Expanding independent validation: Evaluating additional flood and landslide events across diverse mountainous regions using independently sourced reference data.
Improving infrastructure assessment: Incorporating independently verified closure information and strengthening road-network quality checks.
Enhancing computational efficiency: Optimizing geospatial intersections, graph processing, and satellite-data retrieval for larger areas.
Advancing the copilot: Investigating controlled natural-language question selection while maintaining strict evidence grounding and bilingual consistency.
Improving user experience: Expanding map visualization, uncertainty communication, and reporting capabilities.
Our long-term vision is to develop Flood Copilot into a more rigorously validated, transparent, and accessible geospatial decision-support system.
Our guiding principle remains simple: disaster intelligence should communicate what the evidence supports, clearly identify what remains unknown, and never confuse a satellite observation with confirmed ground truth.
Built With
- ai
- dl
- ml
- python
