Research On Semantic Interpretation Technology Of Remote Sensing Image Based On Deep Learning

With the continuous development of remote sensing surveying and mapping in our country,the remote sensing images taken by high resolution satellite have been widely used in production planning,environmental protection and other fields.For example,in recent years,the standardization action of projects under construction has made full use of remote sensing images to interpret ground feature information in a large range,in which the areas violating construction regulations are called disturbance spots.Due to the vast territory of China,it is difficult to detect spot targets quickly with manual labeling,and the accuracy cannot be maintained with manual labeling under heavy tasks.Using computer vision technology to interpret high resolution remote sensing images can improve the annotation accuracy and save a lot of manpower.As an important way of image interpretation,image segmentation can be used for pixel-by-pixel attribute annotation of remote sensing images.Traditional image segmentation algorithms are mainly based on the continuity and similarity of pixel values,so it is difficult to segment remote sensing images with complex textures.With the emergence of convolutional neural network,semantic information in advanced features of images can be obtained.The semantic segmentation algorithm based on convolutional neural network can be used for pixel level labeling of targets under certain semantic rules.However,for the complex geomorphic base map and the variable scale target in the remote sensing image,there are still some problems,such as incorrect interpretion,high omission rate of small target,unclear edge division and so on.Therefore,this paper mainly focuses on the semantic segmentation algorithm based on convolutional neural network and studies the application of complex semantic segmentation in remote sensing images.The main research contents are as follows:(1)In view of the strong correlation between objects in remote sensing images,a spatial attention structure based on multi-scale region search is proposed.The criterion of remote sensing spots is not simply to find a certain type of target,but to reason together with the information of surrounding categories.Spatial attention is introduced to connect the context information of the target more,and to judge the semantic attribute of the target by combining the surrounding and the category of the target.At the same time,global attention adds a lot of computation and distracting information to small targets.In order to deal with the multiscale target,the spatial attention of region search is used to reduce the redundant calculation,and the feature pyramid is used to integrate the region search at multi-scale,so as to retain the feature extraction results at each fine granularity as much as possible.(2)In order to select more effective feature channels from the multi-scale feature pyramid,a hybrid semantic segmentation model of attention is proposed,which integrates the relationship between spatial attention structure and network channels.By using the channel attention module in the splicing structure of multiple branches,the feature channels related to task objectives are given higher weight.In the public data set,the ablation experiments of each module of the mixed attention model were carried out,and its detection accuracy for complex semantic targets increased significantly.In addition,it was compared with the mainstream algorithm in multiple indicators and actual effects.The method also achieves good results in the actual task of spot labeling.(3)For the low fine-grained segmentation results caused by too much interpolation in the upsampling stage of the current semantic segmentation network,this paper proposes a nonuniform sampling upsampling recovery algorithm,which uses more intensive sampling points in the high-frequency region of the classification results to add more effective details to the recovery stage of the semantic segmentation network.In the mixed attention model,the upsampling recovery method under non-uniform sampling is used,and compared with the deconvolution upsampling method and the post-processing method of segmentation results in the traditional semantic segmentation algorithm,the proposed method can effectively improve the fineness of segmentation results.