Extraction Of Marine Aquaculture Rafts From Multi-source Remote Sensing Images Based On MDOAU-net Mode

China has the world’s largest marine aquaculture area.And floating raft aquaculture is a vital production mode in marine aquaculture.However,floating raft aquaculture areas are growing continuously because of the high yield and significant return.The continual expansion of the floating raft aquaculture will cause severe marine environmental problems and reduce production in the marine aquaculture area.Therefore,extracting and monitoring the floating rafts is urgent for the healthy development mariculture industry and marine environment protection.Remote sensing is widely used in floating raft aquaculture information extraction because it has the advantages of a comprehensive monitoring range,fast speed,and easy long-term dynamic monitoring.Traditional remote sensing floating rafts extraction methods usually need to manually design feature extraction algorithms,while deep learning does not.And the deep learning method has strong feature construction ability and robustness,so it has become an essential method for the extraction of floating rafts from remote-sensing images.Remote sensing combines deep learning methods to extract floating rafts quickly,accurately,and intelligently which is of great significance to building a green and sustainable mariculture industry.Single technology of remote sensing has lots of limitations.However,multi-source remote sensing can break these limits.But each type of remote sensing image has different characteristics.To solve the problem,this paper selected multi-source(optical,SAR,PolSAR)remote sensing data as research data.The MDOAU-net model is an excellent deep-learning model to extract floating rafts from SAR images.To achieve the aquaculture floating raft extraction from multi-source remote sensing images.This paper improved the MDOAU-net model according to the characteristics of different remote-sensing images.The main research contents and results of this paper are as follows:(1)This paper proposes the MDOAU2-Net model,which adopts the structure of encoding and decoding to extract floating rafts of optical remote sensing images.The model can maintain the discriminative features and filter the fake objects in the floating rafts extraction.In experiments,compared with other models which have similar structures,MDOAU2-Net has the highest average Io U,reaching 71.33%.(2)The fast PMVOAU-Net model based on the encoding and decoding structure is designed to extract floating rafts from PolSAR remote sensing images.PolSAR multi-scattering feature extraction block is adopted to extract information from all scattering components.Moreover,VGG blocks are used for feature extraction to reduce the extraction time of the model while maintaining high accuracy.The experimental results show that the average accuracy of the fast PMVOAU-Net model is the highest of all contrast models,extracting the floating rafts from PolSAR images,reaching 94.15%.The model costs less time than MDOAU-net to extract rafts,and it can overcome the noises from other information like speckle noise,land,and ship.(3)To extract floating rafts from SAR remote sensing images under high sea states.Firstly,combined time-space information from SAR data and the climate inversion information provided by ERA5 to get the data of wave height and wind speed.Next,describe the sea state level of SAR images through the Beaufort scale.Then,analyzing the SAR images under different sea states,proposing a W-Net model to extract the floating rafts in SAR images under high sea states.According to experiment results,W-Net has a higher extraction accuracy under low sea states,and it can keep high accuracy under high sea states when the MDOAU-net model is useless.