This paper addresses the problem of automatic oil spill detection in Synthetic Aperture Radar (SAR) imagery, a critical capability for maritime monitoring and disaster risk reduction. Despite extensive research, reliable discrimination between oil spills and look-alike phenomena remains a major challenge, particularly under heterogeneous sea-state conditions and limited availability of labelled data. To overcome these limitations, this work proposes an end-to-end processing framework that integrates multi-resolution candidate detection, adaptive pixel-level segmentation, feature-based classification and contextual data fusion within a unified pipeline. The proposed approach introduces a multi-scale statistical modelling strategy for sea clutter characterization, enabling robust detection of heterogeneous regions while preserving sensitivity to potential oil slicks. A compact and discriminative feature space combining morphological and radiometric descriptors is optimized for Support Vector Machine (SVM) classification under limited training data conditions. A key contribution of this work is the incorporation of external semantic information, such as wind fields and potential pollution sources, as a reliability enhancement layer, bridging the gap between image-based detection and operational decision-making. The methodology is validated using SAR imagery acquired by PAZ and TerraSAR-X sensors over multiple maritime scenarios, including real pollution events. Results demonstrate that the proposed framework achieves a favourable trade-off between detection sensitivity and false alarm reduction, outperforming representative segmentation approaches while maintaining accurate delineation of oil slicks. The classification stage reaches an accuracy of up to 96.8% using a reduced feature set. Overall, the results confirm the potential of the proposed framework for robust, automatic oil spill identification in operational scenarios.
Benito-Ortiz et al. (Thu,) studied this question.