A generalized deep learning framework for automated marine oil spill detection using multi-source SAR cross-sea datasets

Marine oil spill detection using Synthetic Aperture Radar (SAR) imagery remains challenging due to complex sea-surface scattering mechanisms and interference from look-alike phenomena, which limit the generalization capability of deep learning models across different marine environments. To address this challenge, this study develops a multi-source remote sensing deep learning framework for automated oil spill detection based on Sentinel-1 SAR imagery. The research focuses on the integrated application of two publicly available datasets—DARTIS and CSIRO—to enhance the model’s adaptability across various marine environments. The results indicate that: (1) the proposed multi-source dataset fusion strategy improves the cross-regional generalization capability of the model compared with single-dataset training strategies; (2) scene-level validation using 27 Sentinel-1 SAR scenes demonstrates that the proposed framework achieves reliable oil spill mapping with Precision, Recall and F1-score of 0.811, 0.956 and 0.878, respectively; (3) the automated monitoring framework developed in this study enables end-to-end processing from SAR image acquisition to oil spill distribution mapping, providing timely monitoring and alerting results. It supports a transition from traditional passive post-event interpretation to an active automated monitoring paradigm, with improved efficiency and timeliness in marine oil spill surveillance.