SipamMar

Un sistema autónomo brasileño de detección y modelado de derrames de petróleo

Abstract

The "SipamMar" project developed an automated operational system for alerting and simulating oil spill dispersion in Brazilian jurisdictional waters, aiming to mitigate the Blue Amazon's vulnerability to spills, a deficiency exposed by the 2019 disaster. This system integrates Synthetic Aperture Radar (SAR) remote sensing, enabling continuous detection regardless of weather conditions, with advanced machine learning and deep learning (ML/DL) techniques, specifically Convolutional Neural Networks (CNNs) like U-Net and ResNet-50, trained with Sentinel-1 SAR images for automated spill detection. Detection accuracy is enhanced by integrating auxiliary environmental data (e.g., chlorophyll-a, wind, currents), which helps reduce false positives (lookalikes) and speckle noise. Upon detection confirmation, the system activates a numerical oil dispersion modeling module using the MEDSLIK-II model, configured to simulate multiple oil types and physical weathering processes, utilizing data from Copernicus Marine Service, ERA5, and GFS/NOAA. The entire process is fully automated by scripts, from data acquisition to graphical output and hourly animation generation, with results interoperable with GIS environments, providing crucial spatial support for decision-making and rapid emergency response. Case studies demonstrate SipamMar's robustness and operational applicability in scenarios with and without false positives, representing a significant advance in Brazil's capacity for monitoring and protecting the Blue Amazon.