Multiphysical Interpretable Deep Learning Network for Oil Spill Identification Based on SAR Images.

The application of deep learning algorithms to oil spill identification in synthetic aperture radar (SAR) remote sensing images has enabled substantial progress. However, the end-to-end learning approach of deep learning has inefficient interpretability, making it difficult to ensure the reliability of oil spill identification tasks. In addition, how to effectively apply the physical information of SAR images to make intelligent decoding methods more interpretable is the key to understanding and evaluating its oil spill identification results. To address these issues, this paper proposes a multi-physical interpretable deep learning network (MIDLN) for oil spill identification based on SAR images. MIDLN defines a dual neural network consisting of a multi-physical deep convolutional neural network and an adaptive grad-weight selection class activation explainer. The explainer adaptively selects the most important interpretation weights for the category, mitigating visual background noise in the explainer results. The high-quality results of the explainer are utilized to provide feedback signals to the network, and the learning strategy of the network is adjusted to structure a deep learning framework with interpretable feedback. Meanwhile, to enhance the physical interpretability of the network, this study employs a scribble-style interactive sampling to streamline the oil spill physical information extraction process. The effectiveness of two novel oil spill physical features is validated through the combination of pixel-level oil-water feature analysis and deep learning methods. In addition, a multi-physical feature extraction head is designed for different features to enhance the uniqueness of physical information and reduce the redundant interference of multiple features. The superiority of the proposed method over existing algorithms is demonstrated by experiments on a large number of Sentinel-1 oil spill datasets. The code for this work will be made available at https://github.com/fjc1575/Marine-Oil-Spill/tree/main/MIDLN for the sake of reproducibility.