Research On The Method Of Raft Aquaculture Area Extraction Based On Fully Convolutional Network And Deep Feature Fusion

Remote sensing has become a primary technology for monitoring raft aquaculture area.In recent years,deep learning has made great progress in the field of remote sensing.However,due to the complexity of the marine environment in the raft aquaculture area,which will result in’adhesion phenomenon’ when using the method of depth learning to extract the aquaculture area from the remote sensing image.There are two main reasons for ’adhesion phenomenon’.On the one hand,the boundary of raft aquaculture area in remote sensing image is often blurred due to the complex marine environment;on the other hand,affected by the growth stage and harvest state of aquaculture crops,some rafts will be lower than the sea water,and the spectral value of raft aquaculture area is very similar to the surrounding sea water.Therefore,when extracting the raft aquaculture area from the remote sensing image,the seawater between the adjacent raft aquaculture areas is easy to be identified as the aquaculture area,which will make the extracted multiple raft aquaculture areas stick together.In order to improve the precision of raft aquaculture area extraction based on deep learning method,in this paper,an end-to-end raft aquaculture area extraction model(UP-Net)is proposed based on fully convolutional neural networks.On the basis of U-Net,this paper proposes pyramid upsampling module,and adds 5 branches to obtain multi-scale feature maps.The large-scale feature maps are used to extract detailed boundary information to prevent boundary-blurring,and the small-scale feature maps are used to extract context information to accurately locate the specific location of the raft aquaculture area.The experimental results show that UP-Net improves the extraction accuracy of the raft aquaculture area,but there is still a small amount of ’adhesion phenomenon’ in the extraction results.Convolutional neural networks can extract the multi-scale feature maps from the remote sensing images.The key to improve the convolutional neural networks is to fuse the detailed boundary information and context information of the multi-scale feature maps.On the basis of UP-Net,UPS-Net is proposed,which uses the deep feature fusion module to learn the weight of each feature maps and carry out weighted fusion.In addition,compared with other neural networks,by using the depth-wise separable convolutions and pyramid upsampling module,UPS-Net have less parameters.Taking the raft aquaculture area extraction of Lianyungang coastal waters as an example,the experimental results show that compared with several state-of-the-art models,the proposed UPS-Net model performs better at extracting raft aquaculture areas and can significantly reduce the ’adhesion phenomenon’.