Satellite-Assisted Rescue & Assessment

Inspiration

During disasters, emergency teams need information quickly, especially in areas that are difficult or dangerous to access. We wanted to explore how satellite imagery and AI could help provide an early view of potentially affected areas.

This inspired us to build SARA (Satellite-Assisted Rescue & Assessment), a system that combines satellite SAR imagery, AI-based segmentation, change detection, and terrain analysis to support preliminary disaster assessment.

What it does

SARA analyzes Sentinel-1 SAR imagery to identify preliminary water candidates and changes between satellite observations.

It:

• Uses ESA SNAP for SAR preprocessing
• Uses a Small U-Net for water segmentation
• Performs multi-temporal SAR change detection
• Uses Copernicus WorldDEM-30 for terrain and slope analysis
• Visualizes results through an interactive web dashboard
• Provides geospatial information to support preliminary disaster assessment

The system is designed as a decision-support prototype. Its AI predictions are preliminary and require independent validation and field verification.

How we built it

We built SARA using Python, PyTorch, ESA SNAP, Rasterio, NumPy, FastAPI, Leaflet, and geospatial processing tools.

The workflow is:

Sentinel-1 SAR ↓ ESA SNAP Preprocessing ↓ Calibrated SAR GeoTIFF ↓ VV + VH + (VV - VH) ↓ Small U-Net ↓ Preliminary Water Candidates ↓ SAR Change Detection ↓ Terrain Analysis ↓ Interactive Dashboard

Challenges we ran into

One of our biggest challenges was processing very large satellite images with limited computing resources. Large SAR scenes can contain hundreds of millions of pixels, making memory management important. We addressed this using windowed raster processing and tiled AI inference. We also faced challenges with:

• SAR preprocessing
• Large GeoTIFF files
• Nodata regions
• Raster alignment
• Coordinate reference systems
• AI inference over large scenes
• Combining AI predictions with terrain information
• Visualizing large geospatial datasets Another important challenge was making sure that model predictions were presented responsibly. A predicted water region is not automatically confirmed flooding, so the system clearly treats its results as preliminary candidates.

Accomplishments that we're proud of

We built an end-to-end prototype connecting satellite data, SAR preprocessing, AI segmentation, change detection, terrain analysis, and an interactive dashboard. We are particularly proud of creating a workflow capable of processing large SAR scenes using memory-conscious techniques while bringing the results together in a single interface. We also integrated multiple types of geospatial information instead of relying only on a single AI prediction.

What we learned

Through this project, we learned that building a satellite-AI system requires much more than training a machine-learning model. We gained practical experience with:

• Sentinel-1 SAR data
• ESA SNAP
• Remote sensing
• GeoTIFF processing
• Raster alignment
• PyTorch and U-Net segmentation
• Windowed inference
• Terrain analysis
• Geospatial visualization
• FastAPI-based AI applications Most importantly, we learned the importance of validation and uncertainty. AI-generated results should be treated as evidence to investigate, not automatically as ground truth.

What's next for Satellite-Assisted Rescue & Assessment

We want to improve SARA by:

• Improving model generalization across different regions
• Adding stronger independent validation datasets
• Improving uncertainty estimation
• Incorporating additional satellite data
• Improving terrain and change-detection analysis
• Making the dashboard more useful for emergency-response workflows

Built With

• ai-for-good
• artificial-intelligence
• change-detection
• climate-tech
• computer-vision
• deep-learning
• disaster-management
• earth-observation
• emergency-response
• esa-snap
• flood-detection
• flood-mapping
• geospatial-ai
• gis
• machine-learning
• natural-disaster
• python
• pytorch
• remote-sensing
• sar
• satellite-imagery
• sentinel-1
• terrain-analysis
• u-net