Intelligent oil spill detection and segmentation using deep CNN and transformer architectures

Abstract

Introduction Oil spills pose significant threats to marine ecosystems and coastal environments, making rapid and reliable detection essential for effective environmental monitoring and response. Satellite imagery enables large-scale monitoring; however, differences among greyscale Synthetic Aperture Radar (SAR), pseudo-RGB SAR, and optical RGB imagery introduce challenges for developing a unified oil-spill detection framework. Conventional approaches based on a single image modality or fixed thresholds may also confuse oil spills with visually similar phenomena, including low-wind areas, shadows, and biogenic films. Methods This study proposes a three-stage Convolutional Neural Network (CNN)-based framework for oil-spill analysis across multiple image modalities. The first stage classifies the input image modality, the second performs binary oil-spill detection, and the third generates pixel-level segmentation maps. Image enhancement techniques, including speckle filtering, Sakaguchi contrast enhancement, and Local Contrast Ratio (LCR), are incorporated to improve image quality and enhance discriminative features for model learning. Results Experimental evaluation demonstrated successful performance of the implemented modality classification, binary oil-spill detection, and Optical RGB segmentation components. The Phase 3 Optical RGB U-Net achieved a Dice/F1 score of 0.8756, an IoU of 0.7787, a precision of 0.8101, and a recall of 0.9525 on a 560-image evaluation dataset. The high recall indicates that the model effectively identified most oil-spill pixels, while the comparatively lower precision indicates some misclassification of background pixels as oil-spill regions. SAR-based segmentation experiments were affected by challenges associated with ground-truth mask processing and computational limitations. Discussion The results demonstrate the potential of the proposed multi-stage CNN framework, particularly for Optical RGB oil-spill segmentation, where the model achieved strong pixel-level detection performance. However, the incomplete SAR segmentation experiments limit quantitative conclusions regarding SAR-based segmentation. Therefore, the present findings primarily establish the effectiveness of the implemented Optical RGB segmentation component, while comprehensive SAR segmentation evaluation remains an area for further investigation.