Enhancing oil spill detection with controlled random sampling: A multimodal fusion approach using SAR and HSI imagery

Oil spills from offshore drilling and coastal refineries pose significant threats to coastal environments. Despite the proven efficacy of multimodal image fusion in various domains, the combined use of multimodal data for oil spill detection (OSD) remains underexplored due to dataset constraints. To explore the efficiency of multimodal image fusion for OSD, this paper first introduces a novel data coregistration strategy to generate paired SAR-Hyperspectral image (HSI) datasets, enabling comprehensive algorithm testing to elucidate the characteristics of unimodal and multimodal data. We then develop a SAR and HSI fusion network (SHNet), setting a new baseline for multimodal OSD. Our findings indicate that while SAR images effectively differentiate oil spills from water surfaces, they show significant variance in distinguishing thin from thick oil, with accuracy discrepancies of approximately 6.41-56.98%. In contrast, HSIs excel in identifying various types of oil, although they exhibit limited generalization capabilities compared to SAR images, as evidenced by a Kappa reduction of around 7%–45%. The SHNet effectively harnesses the strengths of both SAR and HSI, achieving superior performance in oil type discrimination and overall OSD through hierarchical feature extraction from both modalities. Our results suggest that while SAR imagery is optimal for rapid, large-scale OSD, the fusion of HSI and SAR data provides more precise oil type estimation within identified spill areas.