Deep-water oil-spill monitoring and recurrence analysis in the Brazilian territory using Sentinel-1 time series and deep learning

•This is the first oil-spill deep learning study in the Brazilian territory.•We compared three deep learning architectures and four backbones.•We detected hotspots using recurrence analysis from deep learning predictions.•A recurrence threshold greater than two eliminates nearly all look-alike features.•Most oil-spills events in the Campos Basin are related to oil platforms. Oil spills are a worldwide concern since they cause environmental problems and financial losses. Automatic detection plays a crucial role in rapid decision-making to reduce damage. In this context, deep learning and remote sensing are powerful tools with successful applications in many regions. However, there are still no studies on deep-water zones and in the Brazilian territory. The present research has three contributions: (1) create an oil spill dataset on the Brazilian region, (2) compare state-of-the-art deep learning models for this task, and (3) propose a novel application on the Campos Basin (Brazil's largest producer) using Sentinel-1 time series to identify more susceptible regions, encompassing 138 images from 2016 to 2021. The experts manually labeled the dataset with confirmation from the oil company and federal inspection, resulting in 800 images with 512 × 512 spatial dimensions and their respective annotations (600 for training, 125 for validation, and 75 for testing). The study compared three semantic segmentation architectures (U-net, DeepLabv3+, and LinkNet) with four backbones (ResNet-101, ResNet-50, Efficient-net-B0, and Efficient-net-B3) resulting in 12 models. The U-net with the Efficient-net-B3 backbone presented slightly better results (98% accuracy, 75% precision, 78% recall, 76% F-score, and 62% IoU). The time-series analysis considered images with 8,192 × 11,264 spatial dimensions. The entire image classification used a sliding window approach with a 128-pixel stride (averaging the overlapping pixels). Moreover, recurrence analysis detected the more susceptible areas and eliminated the look-alike. False-positive features have a high spatial presence, but with low recurrence (rarely reaching 3 events), which facilitates their elimination using a threshold recurrence. This novel analysis for oil spills is suitable for understanding the patterns along the time and providing regions for the authorities in real-time.