Research On Semantic Segmentation Based On Deep Learning Of High Resolution Remote Sensing Image

With the rapid development of remote sensing technology,we can obtain more high-resolution remote sensing images with clear texture details and rich spectral features.How to effectively use these remote sensing images with rich information has become a hot research topic The traditional machine learning method has certain complexity in the process of feature analysis and feature extraction of high-resolution remote sensing images,and has significant limitations in dealing with the spatial structure and edge feature information of the object effectively.However,deep learning breaks the limitations of traditional machine learning methods,and can extract features of high-dimensional information in texture,spatial structure and spectrum.What is more,it has been applied to semantic segmentation in natural disaster monitoring,urban planning,land cover monitoring and other areas.This paper mainly explores semantic segmentation of high-resolution remote sensing images based on deep learning models,and the related researches are as follows(1)An improved deconvolution network model is proposed,which connects the corresponding feature layers of encoding structure and decoding structure.It can extract the feature information of sample space structure and edges in a deeper level.Experimental results using the model on a building remote sensing images data set in Massachusetts show that the improved model has 3%?6%higher pixel accuracy than other classic semantic segmentation network models.(2)Deep learning convolutional neural network models often require large amounts of labeled data for training.In contrast,the number of labeled images compared to numbers of remote sensing images is small,which makes it difficult to train deep learning network models.The data pre-training model is designed in this paper for the dynamic loading of the input data without over-fitting of the network model.On the basis of not changing the original data,the operations of clipping,affine transformation and elastic transformation on the remote sensing images are combined in this model to expands the data.(3)In this paper,we use the focal loss function to optimize the deeplabv3+network model,and extend the data with the pre-training model for improving the accuracy of the network model.DeeplabV3+uses multi-scale method to perform hole convolution on the feature map of each scale,then connects these feature maps.The separable convolution is used in the Xception structure to optimize the model.The experimental results show that the optimized DeeplabV3+ model can extract the edges and spatial features of samples effectively,and improve the segmentation effect of remote sensing images.