Method of offshore aquaculture area extraction based on strip convolution and context awareness
Aiming at the problems in medium resolution remote sensing image such as fuzzy boundary,inter-class interference in rafts and nets farming,the MSUResUnet model with strip convolution module and context-aware unit is proposed in this study based on ResUnet to improve the feature extraction ability,which can improve the problems of missing extraction and adhesion in offshore aquaculture extraction tasks.The strip pooling module in the model is used to enhance the interaction between the encoding and decoding layer information.The multi-directional strip convolution module can better capture the linear characteristics of aquaculture,and the context-aware unit can obtain the rich multi-scale context information of aquaculture area.Experiments results on Sentinel-2 MSI data show that among the six models participating in the comparison,the MSUResUnet model has the best accuracy,and its kappa,MIoU,OA and F1-score reach 89.17%,84.33%,96.38%and 91.19%,respectively.The MSUResUnet model achieves high accuracy in farming extraction in the more intensively farmed waters around Xinghua Bay,Sansha Bay and Luoyuan Bay.MSUResUnet model has stronger feature extraction and anti-interference ability,which can meet the needs of high-precision large-scale medium-resolution image offshore aquaculture information extraction.
raft and cage aquaculturedeep learningResUnet modelmulti-directional strip convolu-tion modulecontext-aware unit
万方数据
维普
