Spaceborne Flood Damage Intelligence

Multimodal satellite AI for flood mapping, infrastructure assessment, and identifying isolated communities.

Comment

FloodLens AI: Turning Satellite Data into Disaster Intelligence

💡 Inspiration

During a major flood, the biggest challenge is not always finding out where the water is. It is understanding what happens next. Which roads might be blocked? Which bridges could be affected? Which villages may lose access to hospitals and emergency services?

We wanted to explore how artificial intelligence and satellite technology could help answer these questions, especially when ground communication and infrastructure are disrupted. This led us to create FloodLens AI, a disaster-intelligence platform that connects satellite-based flood mapping with infrastructure analysis and route connectivity assessment.

Our vision is to make complex geospatial information easier to understand and more useful for emergency planning.

🌍 What We Built

FloodLens AI brings together satellite imagery, machine learning, geographic information systems, and graph algorithms in a single application.

Its main capabilities include:

  • Satellite Flood Mapping: Processes Sentinel-1 radar imagery to identify areas that may be inundated. Optical satellite data can provide additional information when suitable imagery is available.
  • Infrastructure Exposure: Combines detected hazard regions with geographic information about roads, bridges, buildings, and settlements to highlight infrastructure that may be affected.
  • Settlement Connectivity Analysis: Represents the road network as a graph and evaluates whether settlements remain connected to selected emergency facilities.
  • Alternative Route Exploration: Identifies possible routes around road segments flagged as potentially affected.
  • AI-Assisted Incident Information: Provides a question-answering interface for exploring available disaster information in English and Nepali.
  • Situation Report Generation: Organizes analytical results into a structured summary that can be reviewed and shared.

Rather than presenting satellite images alone, FloodLens AI aims to connect image-based observations with geographic context and infrastructure connectivity.

⚙️ How We Developed It

We designed the application as a collection of connected processing modules.

The first stage analyzes Earth-observation imagery. Sentinel-1 synthetic aperture radar is particularly useful for flood assessment because it can collect observations in cloudy conditions. Where compatible optical imagery is available, Sentinel-2 data can contribute additional information about the affected landscape.

The next stage transforms the image analysis into geographic information. Predicted flood regions are aligned with mapped infrastructure so that potentially exposed roads, bridges, and settlements can be examined.

We then use graph-based analysis to study road connectivity. In this representation, locations become graph nodes and road segments become edges. When a road segment is flagged as potentially disrupted, the analysis can evaluate how its removal or restriction changes connectivity to an emergency-service hub.

Finally, a web dashboard brings these outputs together, allowing users to explore map layers, inspect infrastructure exposure, review connectivity results, and access generated incident summaries.

Our implementation uses Python, PyTorch, FastAPI, geospatial processing libraries, NetworkX, and an interactive web-mapping interface.

🧩 Challenges We Encountered

One of the main engineering challenges was bringing together information from different sources. Satellite images, elevation models, and infrastructure maps can use different coordinate systems, spatial resolutions, and observation dates. These differences must be handled carefully to avoid misleading overlays or area calculations.

Another challenge was interpreting the results responsibly. Detecting water near a road does not automatically mean that the road is impassable, and a predicted hazard does not independently confirm structural damage. Our analysis therefore treats these findings as potential impacts that require further verification.

We also had to account for limitations such as radar distortions in mountainous regions, missing map features, uncertain segmentation boundaries, and delays between satellite observations. These issues reinforced the importance of testing the pipeline and documenting its assumptions.

📚 What We Learned

Building FloodLens AI helped us see how different technical disciplines can work together to address one real-world problem.

We explored satellite-image segmentation, geospatial data processing, spatial overlays, graph connectivity, API integration, and interactive visualization. We also learned that the value of an AI system depends on more than model accuracy: its outputs must be understandable, geographically meaningful, and connected to a useful decision-making process.

Perhaps our most important lesson was that responsible disaster technology must communicate uncertainty clearly. It should help people investigate a situation without presenting estimates as confirmed facts.

🚀 Future Improvements

We want to strengthen FloodLens AI through more extensive testing on independent flood events, improved model evaluation, better infrastructure data, and additional geographic case studies. Future development could also improve the assessment of road disruptions, refine route comparisons, and make the interface more accessible to disaster-management teams.

Our long-term goal is to develop a system that transforms Earth-observation data into understandable, evidence-based insights about flood extent, infrastructure exposure, and community connectivity.

FloodLens AI brings satellite intelligence and geospatial reasoning together to help answer a critical question during disasters: when the landscape changes, which places and connections need our attention first?

Built With

Share this project:

Updates

Submission history