The water benders

Project Description

⚠️ The Problem - Background information

In the last few years, both Romania and the wider European region have faced a "whiplash" effect, swinging from record-breaking droughts to devastating flash floods.

Droughts have been catastrophic for summer crops in the 2024-2025 season, with maize production plummeting by 31.5% due to extreme heat stress and soil moisture deficits. Agricultural production saw double-digit percentage drops across all categories (cereals, oilseeds, vegetables). Bulgaria was affected similarly, making it a transboundary environmental disaster.

Flash floods severely affected Dambovita and Arad counties in 2025, southern Romania in late 2024, and most famously France and Germany in early 2026, leading to saturated soils and waterlogging.

✅ Our Solution - Disaster Risk Monitoring & Early Warning System

AgroSentinel is an advanced geospatial intelligence platform that provides AI-powered risk assessments for multiple natural disasters: floods, droughts, wildfires, storms, and heatwaves. Users draw their fields directly on an interactive map, and the platform then automatically refreshes a per-field risk profile every hour by fusing real-time weather data, multi-mission Copernicus satellite imagery, and historical climate patterns — delivering actionable insights for disaster preparedness and resilience planning.

Core Features

🌍 Multi-Sensor Data Fusion

• Real-time weather forecasts from Open-Meteo, with a configurable forecast window (24 h default)
• Multi-day predictions (5-day default) for proactive planning
• Satellite evidence from Copernicus Data Space (Sentinel-1, Sentinel-2, Sentinel-3, Sentinel-5P, MODIS)
• Copernicus Land Monitoring Service products: SWI, surface soil moisture, NDVI, burned area, water bodies
• Historical climate baselines (10-year) and anomaly detection (90-day lookback)
• Terrain elevation and hydrological context, plus OpenStreetMap-derived land-use layers
• Regional threshold calibration: Mediterranean, continental Europe, northern Europe, Middle East dry, tropical

📍 Spatial Risk Mapping

• Polygon-first design: users trace any field shape on the interactive map and area in hectares is computed automatically
• Dynamic zone-based exposure analysis with customizable grid resolution
• Interactive map visualizations with real-time, status-coloured heatmaps (healthy / flood / drought / alert)
• Relative exposure indices computed per monitored parcel and refreshed hourly
• Per-field caching so dashboard reloads are instant even at scale

🛰️ Advanced Satellite Integration

• Sentinel-1 SAR: VV-band water detection for flood early warning
• Sentinel-2 L2A: Vegetation health and optical data quality, with six spectral indices computed in a single Process API call (NDVI, NDWI, NDMI, NBR, BSI, NDBI) and automatic cloud filtering
• Sentinel-3 LST: Land surface temperature anomalies for heatwaves
• Sentinel-3 FRP: Active fire detection for wildfire alerts
• Sentinel-5P: Atmospheric aerosol and carbon-monoxide indicators
• MODIS: Burned area tracking and NDVI vegetation indices
• CLMS: Soil moisture and water stress indicators, plus optional LCFM LCM-10 land-cover at 10 m
• 14-day STAC catalogue lookback per analysis run

📊 Decision-Support Scoring

• Five calibrated risk indices: Flood, Drought, Storm, Heatwave, Wildfire
• Risk scales: None (0–5%), Extremely Low (5–15%), Very Low (15–30%), Low (30–45%), Moderate (45–60%), High (60–80%), Extreme (80–100%)
• Multi-component weighted scoring — e.g. flood = 65% forecast precipitation + 25% recent wetness + 10% terrain runoff, with resource modifiers for water proximity and urban density
• Satellite confidence gating capped at 25% of total weight, so weather and context remain dominant when satellite evidence is stale or missing
• Random-Forest disaster-type classifier (normal, flood, fire, deforestation, landslide) from before/after Sentinel-2 image pairs
• Explainable: each score includes methodology and input factors

⚙️ Enterprise Ready

• JWT-secured REST API for programmatic access
• Redis-backed per-field cache (~65-minute TTL) with PostgreSQL/Supabase historical archive
• Automated hourly synchronisation via cron worker, with throttled per-field analysis to respect Sentinel Hub quotas
• JSONL/CSV/GeoJSON export formats
• STAC-compliant geospatial asset delivery, plus CoverageJSON-style weather and GeoJSON event/prediction layers
• Manifest-based site bundle: a single manifest.json entrypoint links the dashboard payload, maps, events, and predictions
• Horizontal scalability for multi-region deployments

Use Cases

✓ Emergency Management Agencies — Real-time disaster risk situational awareness

✓ Agricultural Sector — Drought monitoring & irrigation planning

✓ Food Markets — Producer price trends & forecasts

✓ Urban Planners — Long-term climate resilience assessments

✓ NGOs & Development — Community early warning systems

Methodology

The project does not predict event probabilities. Instead, it computes operational risk indices that combine:

• Weather signals: precipitation, temperature, wind, atmospheric instability (CAPE), evapotranspiration, vapour pressure deficit
• Environmental context: seasonal baselines, multi-year historical anomalies, and elevation-derived terrain runoff modifiers
• Satellite observations: real-time water, vegetation, fire, and thermal data, gated by acquisition recency and cloud cover
• Spatial exposure: zone-level vulnerability to specific hazards, with per-region threshold calibration based on the field centroid

Each of the seven warning bands — None, Extremely Low, Very Low, Low, Moderate, High, Extreme — is calibrated against local thresholds and should be validated against official warnings and historical event data before operational deployment. The bands Moderate, High, and Extreme are flagged as active risks and surface in the dashboard's alert feed.

👨🏻‍💻 Meet the Team

Vesel Matei-Denis

Meet the lead backend developer of our website. Student at the Faculty of Automatic Control and Computer Engineering in Iasi, with him 5 years of experience in Node programming, he linked the "face" and the "brains" of our programme with unrivaled efficiency to compute the many requests our site will receive.

Matei Marcel-Rares

Student at the Faculty of Automatic Control and Computer Engineering in Iasi and a natural-born mathematician, having participated at multiple international contests including SEEMOUS, he created our AI model which analyses and interprets the satelite data to offer precise weather forecasts and recommendations for disaster prone areas.

Manolica Matei-Theodor

Student at the Faculty of Automatic Control and Computer Engineering in Iasi and an adaptable member in any team, I am a jack of all trades who worked on the website of our platform and also on our final presentation to make everything as clean as possible.

Hanganu Daniel-Nicolae

Three-time gold medal winner at International and European Geography Olympiad during High School, now student at the Faculty of Automatic Control and Computer Engineering in Iasi. Provided the team with real-world areas of interest and case studies which were used to train and verify our language model.

Corduneanu Alin-George

A student at the Faculty of Automatic Control and Computer Engineering in Iasi with natural inclination towards STEM. Participated at SEEMOUS and other local hackathons such as Prove It. Helped the team with creating the server for the AI and modifying the frontend.

Coșa Andrei-Valeriu

Master student at the Faculty of Automatic Control and Computer Engineering in Iasi with a crazy passion towards robotics and emerging technologies. Brings experience of previous hackatons presentations to a young new generation of recruits.

Presentation Materials

💻 GitHub Repository

https://github.com/ManolicaMatei/Cassini-AgroSentinel

📀 Slides Presentation

https://docs.google.com/presentation/d/1eB2fGdJ9KVHN9CveuCCHdZxA956WAKN1/edit?usp=sharing&ouid=114752394227165589023&rtpof=true&sd=true

📹 YouTube Presentation

https://youtu.be/7HC2YIuoSrk?si=nEcuqKsOcj2AmBxh