Nelvinebi/Niger-Delta-Oil-Spill-Monitoring-System

This project applies deep learning to synthetic SAR data for oil spill detection and GIS-based impact mapping in the Niger Delta, producing georeferenced raster and vector outputs suitable for environmental monitoring, research, and spatial analysis workflows. # 🛢️ Niger Delta Oil Spill Monitoring System (NDOSMS) > A physics-informed **deep learning pipeline** that detects oil spills from **Synthetic Aperture Radar (SAR)** imagery using a **54-layer U-Net** with attention gates, **Monte Carlo Dropout uncertainty quantification**, and a **FastAPI deployment layer** achieving **94.38% pixel-wise validation accuracy** on synthetic Sentinel-1 data simulated with Bragg scattering physics and oil-damping models. --- ## 📌 Problem The Niger Delta home to over 30 million people and one of the world's most biodiverse wetland ecosystems has experienced contamination equivalent to more than **13 million barrels of crude oil** since the 1950s. Mangrove forests are declining at an estimated **5,644 hectares per year**, with cascading impacts on fisheries, livelihoods, and water security. Conventional oil spill monitoring depends on expensive field surveys and passive optical satellite observation both of which are blocked by the near-constant cloud cover over the Niger Delta, and both of which detect spills **days or weeks after occurrence**. There is an urgent need for an automated, all-weather, near-real-time detection system that can penetrate cloud cover, produce confidence-scored alerts, and integrate directly with regulatory reporting workflows. NDOSMS addresses this by applying **Synthetic Aperture Radar (SAR)** which operates through clouds, rain, and darkness combined with a deep learning segmentation model that produces pixel-level spill masks and uncertainty maps from each satellite pass. --- ## 🎯 Objectives - Simulate **physics-based synthetic SAR data** using Bragg scattering ocean backscatter models and oil-damping effects across multiple weather and thickness scenarios - Train a **U-Net semantic segmentation model** to detect oil spills at pixel resolution from SAR GRD imagery - Implement **Monte Carlo Dropout** at inference time to produce per-pixel confidence and epistemic uncertainty maps al …