Land Cover Classification Based On Deep Learning For Remote Sensing Image

Land cover classification of remote sensing data is one of the important application fields of remote sensing.With the rapid development of satellite imaging technology and aerospace technology,more and more high-resolution remote sensing satellites have been put into use,and the number of remote sensing images that can be obtained by humans is constantly increasing and the resolution is constantly improving.In high-resolution remote sensing images,the category features of different feature elements are significant,and the texture details are rich,which are an important prerequisite for realizing remote sensing image feature classification.However,compared with natural images,high-resolution remote sensing images have richer information and more complex scenes.Therefore,direct application of general computer vision processing algorithms to remote sensing images often results in poor results.Based on the imaging characteristics of high-resolution remote sensing images,this paper improves the existing algorithms for pixel-level classification tasks of natural images.In this article,the existing algorithms are optimized from the three directions of the dataset,the network structure of the deep convolutional network,and the loss function to improve the final performance of the algorithm.The main research contents of this paper are as follows:1.This thesis proposes a low-level feature extraction module and an iterative update strategy for labels.The low-level feature extraction module aims at the feature that the remote sensing image has more representation information at each pixel,and extracts more features from the shallow features for the final classification decision;the iterative update strategy addresses the problem of manual annotation errors in remote sensing images dataset.The output of the network corrects the original manual annotations,uses the corrected labels to continue training the algorithm model,repeats the ”training-correction” step until the network converges again,and uses the information learned by the network itself to correct the performance degradation of the network model caused by the data deviation to improve the final performance of the algorithm.2.This thesis proposes a deep feature fusion method based on U-shaped network.The Ushaped network not only loses the resolution in the process of encoding and decoding,but also insufficient fusion of shallow and deep features.In order to improve the feature fusion ability of the algorithm,the attention module is introduced,which is used to improve the extraction ability of finer edges in the fusion of shallow features and deep features;the spatial pyramid pooling module is introduced,which is used to supply the lost resolution information which is undermined during encode and decode phase.In order to solve the problem of the lack of interaction between different depth information of the U-shaped network,the network cross-layer feature fusion module is introduced to iteratively integrate the features of different layers,so as to better supplement the detailed information.3.This thesis proposes coordinate loss based on label residuals.The most commonly used cross-entropy loss in current pixel-level classification only pays attention to whether the classification of a single pixel is correct,but it cannot pay attention to the loss of image global context information.In order to encourage the algorithm model to pay attention to the loss of the image level during the training process,this paper proposes a coordinate loss based on the label residual.By modeling the residual information of the image label in two spatial directions,it uses two one-dimensional features to represent the image level difference to assist the cross-entropy loss function to optimize the network.Through the joint constraints of cross entropy loss and coordinate loss,the final performance of the network is improved.