MS3OSD: A Novel Deep Learning Approach for Oil Spills Detection Using Optical Satellite Multisensor Spatial-Spectral Fusion Images
Kai Du, Yi Ma, Zhongwei Li, Rongjie Liu, Zongchen Jiang, Junfang Yang
Abstract
Marine oil spills pose a significant threat to ecosystems, highlighting the critical need for effective monitoring technology. Optical remote sensing technology plays a crucial role in monitoring marine oil spills. However, its performance is constrained by inherent tradeoffs among temporal, spatial, and spectral resolutions, making it difficult for a single sensor to fully meet the demands of oil spill monitoring. Furthermore, existing oil spill detection algorithms often prioritize surrounding spatial features while neglecting the contribution of central spectral features, resulting in reduced detection accuracy. To address these issues, this article proposes a joint framework for multisensor data spatial-spectral fusion and oil spill detection. This framework fuse images from the coastal zone imager (50 m, 4 bands) with images from the ultraviolet imager and the Chinese Ocean Color and Temperature Scanner (1000 m, 10 bands), all of which are onboard Haiyang-1C/D satellites, generating high temporal and spatial resolution ultraviolet-visible-near-infrared range images with 10 bands. The framework uses parallel branches, including a convolutional neural network and a vision transformer, to extract surrounding spatial features and central spectral features from the fused data. This design enables the effective combination of fine-grained spatial information with multiband spectral information, facilitating precise detection of oil spills in various emulsification states under different sun glint conditions. The proposed framework demonstrates strong performance, achieving F1-scores of 95.24% and 93.04% for detecting oil slicks and oil emulsions under weak sun glint conditions, and 90.06% for positive contrast oil spills under strong sun glint conditions. This study provides new insights for advancing oil spill monitoring and highlights the potential of multisensor data fusion in marine target detection.