UNet-NNI: A Collaborative Inpainting Method Based on Deep-Sea Multibeam Backscatter Intensity Images

The long acoustic path and the complexity of deep-sea environments make it difficult to obtain complete and reliable multibeam backscatter intensity data, often resulting in central beam anomalies and/or missing and erroneous measurements. To address this challenge, this study developed a joint UNet-NNI (nearest neighbor interpolation) inpainting method for deep-sea multibeam backscatter intensity images. First, a UNet-MCSA (multiscale channel spatial attention) model combining shortcut connections and the MCSA mechanism was developed and trained on backscatter data obtained from UK waters, with data from a Pacific deep-sea region used for validation. Then, a collaborative UNet-NNI framework was constructed by integrating histogram matching with NNI to improve clarity in large-gap regions. Experimental results demonstrated that the proposed approach achieves higher structural fidelity and better intensity consistency than realized using classical interpolation and conventional deep learning methods, across both simulated and real data. The method effectively inpaints large-scale missing regions, and produces visually and physically consistent intensity images, thereby providing a practical deep–shallow hybrid paradigm for deep-sea multibeam backscatter inpainting.