-
-
Dashboard
-
-
-
-
-
FloodLens: Project Journey
What Inspired Us Disaster response in mountainous and remote river corridors often fails not because of a lack of satellite data, but because of latency and fragility in spatial pipelines. During emergency windows, analysts face rigid API timeouts, missing pre-event road infrastructure data, and fragmented water bodies that obscure true flood impact.
We were inspired to build FloodLens to bridge the gap between raw Sentinel-1 synthetic aperture radar (SAR) imagery and actionable local intelligence. Our goal was to create an automated, end-to-end resilient system that marries AI-driven water extraction with robust infrastructure safety caching, ensuring that disaster teams get reliable impact assessments even when network conditions fail.
How We Built It FloodLens combines geospatial signal processing, machine learning, and resilient local engineering into a cohesive pipeline:
Core Processing & AI Engine: We integrated a hybrid approach combining custom rule-based change detection with fine-tuned U-Net models for water body identification, paired with HAND (Height Above Nearest Drainage) filters using pysheds to filter out false-positive terrain anomalies.
Resilient Infrastructure Data (Overpass Caching): To conquer API rate-limits and server blocks (like 504 Gateway Timeouts), we engineered an intelligent Overpass cache-warming script. It splits large geographic areas of interest (AOIs) into custom 11km tiles, uses exponential backoff, and saves immutable historical snapshots locally.
User Interface & Dashboard: We built a fully interactive frontend using Streamlit (app/frontend/Home.py) to give responders a clean dashboard for viewing flood spreads, infrastructure impact, and validation metrics.
Challenges We Faced API Rate Limits & Timeouts: Overpass and geospatial servers frequently threw connection errors or blocked requests under heavy hackathon workloads. We solved this by implementing smart grid tile-splitting and strict local caching to make executions fully idempotent.
Narrow Mountain Channel Detection: Early tests showed the U-Net model excelled at capturing broad water pools but struggled along narrow 30m mountain debris corridors. We honestly documented this limitation and balanced our pipeline fallback rules to handle regional variations transparently.
Environment & Dependency Clashes: Managing specific scientific libraries (numpy, pysheds) under tight time constraints required strict version pinning and isolating environment configurations.
What We Learned Resilience Over Feature Creep: In emergency or hackathon scenarios, making existing pipelines bulletproof (like handling server timeouts gracefully) matters infinitely more than adding speculative features.
Honest Validation: Real-world geospatial models have blind spots (such as narrow channel misses). Embracing these findings through transparent metrics comparisons (against EMSR datasets) yields much stronger system credibility.
Smart Caching Saves Deadlines: Implementing robust caching layers early prevents last-minute panic when external APIs drop offline.
Here is the final part of your submission documentation, outlining your core accomplishments, key takeaways, and the future roadmap for FloodLens:
Accomplishments That We're Proud Of Robust Server Resilience: Successfully engineered an intelligent Overpass cache-warming script with grid tile-splitting and exponential backoff, conquering stubborn API gateway timeouts and making historical analysis completely reliable.
Hybrid Water Extraction: Harmonized custom rule-based change detection with fine-tuned U-Net models and pysheds-backed HAND (Height Above Nearest Drainage) filters to effectively target broad water bodies while filtering terrain noise.
Transparent Benchmarking: Validated the system against real-world EMSR927 activation data, documenting both broad water detection successes and narrow mountain channel constraints with complete honesty.
What We Learned Resilience Trumps Feature Creep: In high-stakes disaster and hackathon scenarios, ensuring pipelines gracefully handle server drops and API throttling matters vastly more than piling on speculative features.
Embrace Real-World Limitations: Acknowledging model blind spots—such as narrow mountain stream detection limits—and addressing them through transparent fallback rules creates a much stronger and more trustworthy technical report.
Immutable Caching Wins: Implementing smart local caching early completely eliminates last-minute panic when external mapping servers or data APIs experience heavy traffic.
What's Next for FloodLens Multi-Scale Model Generalization: Fine-tuning the segmentation architecture to capture narrow 30m mountain debris channels and micro-corridors alongside wide water bodies.
Automated Multi-AOI Expansion: Extending the pipeline to handle dynamic, multi-zone disaster activations concurrently.
Enhanced Decision Support Integration: Expanding the Streamlit dashboard layout to provide real-time evacuation prioritization metrics for first responders along with Cloud removal feature
Log in or sign up for Devpost to join the conversation.