Multiscale Feature Fusion for Hyperspectral Marine Oil Spill Image Segmentation

Oil spills have always been a threat to the marine ecological environment; thus, it is important identify and divide oil spill areas on ocean surface into segments after an accident occurs protect environment. However, area segmentation using ordinary optical images greatly interfered with by absorption of light deep sea distribution algal organisms surface, difficult improve accuracy. To address above problems, hyperspectral image model multiscale feature fusion (MFFHOSS-Net) proposed. Specifically, dataset was created data from NASA for Gulf Mexico spill, small-size waveband filtering were generated annotated. The makes full use having different layers characteristics fusing maps scales. In addition, attention mechanism used effectively fuse these features region A case study, ablation experiments evaluation also carried out in this work. Compared other models, our proposed method achieved good results according various metrics.