Information Extraction Of Remote Sensing Image For Coastal Wetland Based On Semantic Segmentation

Coastal wetland is located in the transition zone between land system and ocean system,and it is one of the wetland environments that are key protected by the state.With the progress of wetland protection,field surveys by expert groups are no longer applicable to today’s large-scale wetland environment monitoring,and the development of remote sensing imaging technology has made it possible to carry out such large-scale monitoring.However,many vegetation in the coastal wetland have low coverage density or alternate growth with other vegetation around the distribution area,which makes it difficult to distinguish the boundary of ground objects clearly,which makes it difficult to extract information.When extracting information from remote sensing images for coastal wetland,the traditional method uses a single pixel or the area of pixel as the training sample,without considering the boundary area between different ground objects,which is the key to improve the accuracy of information extraction.To this end,this thesis introduces an idea of semantic segmentation of deep learning and proposes a method for extracting information for coastal wetland based on UNet.This method uses the junction area of different objects as training samples,and uses UNet model for training and testing.This thesis uses the GF-2 satellite to take a multispectral remote sensing image of the Yellow River Estuary Wetland in Kenli District,Dongying City,Shandong Province,China as an example.Experimental results show that compared with three traditional machine learning methods:Support Vector Machines,Mahalanobis Distance and Maximum Likelihood Estimation,the method proposed in this thesis improves the overall accuracy by more than 3%,kappa by more than 0.04,and the F1 scores of six classes are higher than those of the comparison method.When extracting information from remote sensing images for coastal wetland,UNet model has limited ability to maintain the integrity of the extracted object’s boundary contour.Therefore,this thesis proposes a semantic segmentation model WetlandNet for coastal wetland based on deconvolution.The model uses an encoder-decoder as the basic structure.It uses depthwise separable convolution instead of regular convolution to reduce the model size and parameter amount.It uses deconvolution to extract the boundary contour features of the target,and the extracted boundary contour features through the jump connection is spliced to the upsampled feature map,which enriches the feature information.The experimental results show that compared with representative semantic segmentation models UNet,PSPNet and DeepLabV3 Plus in deep learning,the WetlandNet proposed in this thesis obtains more accurate segmentation results,improving the overall accuracy by more than 5%,and Kappa improves above 0.07,and the F1 score of 5 of the 6 classes is higher than the comparison model,and the parameter amount is only 1/36 of UNet,1/42 of PSPNet,and 1/51 of DeepLabV3 Plus.