Enhanced Minority Class for Oil Spill Detection by Integrating Weighted Data Class Into Mask R-CNN Architecture (WMask-RCNN)

Efficient detection and precise classification of oil spills in marine environments remain challenging, mainly due to the oil spill class imbalance within Synthetic Aperture Radar (SAR) datasets, which reduces deep learning model performance and often leads to false negatives. Moreover, traditional deep learning models have limited generalization efficiency, making their performance unreliable when tested on new oil spill patterns. Existing models, such as two-stage Region-Based Convolutional Neural Networks (R-CNN), are inefficient in detecting small oil spill patterns because of class imbalance. In response, this study introduces WMask-RCNN, an improved oil spill detection framework based on satellite imagery that focuses on improving minority class representation during training. The proposed architecture integrates weighted class data to reinforce learning of minority features and is enhanced with Squeeze-and-Excitation (SE) blocks, Convolutional Long Short-Term Memory (ConvLSTM) for temporal feature modeling, and the Convolutional Block Attention Module (CBAM) to improve spatial attention. A dynamically adaptive weighted cross-entropy loss function is employed to emphasize small oil spill patterns during training, thereby reducing false negatives and enhancing detection performance. The experiment was carried out on the MKLab/ITI benchmark SAR dataset and achieved higher performance with the proposed WMask-RCNN, which outperformed the latest architectures, including DL-ViT, FC-ViT, U-ViT, DeepLabV3+, and FC-DenseNet, in terms of accuracy, precision, recall, F1-score, and IoU, with values of 97%, 96%, 96%, 96%, and 81% respectively. The results presented here demonstrate the potential of the proposed model and its usefulness for reliable monitoring and detection of oil spills in complex offshore environments.