Research And Application Of Remote Sensing Image Feature Extraction Based On Deep Learning

With the development of satellite remote sensing technology,the image resolution of remote sensing images has been continuously improved,and the amount of data in remote sensing images has also increased dramatically,making traditional feature extraction techniques can’t to meet the interpretation needs of high-resolution remote sensing images;At the same time,as an important part of the field of computer vision,deep learning has gained widespread attention and rapid development in recent years,deep learning-based semantic segmentation technology has become a very active research direction in the field of Computer image processing and deep learning.This thesis is combining the characteristics of high-resolution remote sensing images,and based on the basic theory of Neural network,analyze and compare the performance of existing typical semantic segmentation networks in the remote sensing image feature extraction task,and improve the model,applying it to the field of high-resolution remote sensing image processing,has greatly improved the processing speed of remote sensing image feature extraction tasks.On this basis,this thesis designs and completes an intelligent remote sensing interpretation platform based on deep learning,which can intelligently interpret high-resolution remote sensing images automatically.The main contents of this thesis are as follows:(1)Established a remote sensing image semantic segmentation dataset that can be used for deep learning training.For the existing high-resolution remote sensing images and manually labeled vector files,by developing corresponding format conversion algorithms,this thesis converts vector files into truth map labels that can be used for deep learning.At the same time,remote sensing images were cut and data augmented,and finally a large remote sensing image semantic segmentation data set that can be used for deep learning training was constructed.(2)Through classical semantic segmentation network FCN,U-Net,Deeplabv3+analysis and comparison,this thesis puts forward a modified model based on U-Net,the concept of dilated convolution is integrated into U-Net,and the transposed upsampling convolution in U-Net is replaced by parameterless bilinear upsampling.Finally,the accuracy of the model has reached the same level as DeeplabV3+on the basis of reducing the amount of parameters.Improved the speed of model training and prediction.(3)This thesis designs and implements an intelligent remote sensing interpretation platform based on semantic segmentation.A deep learning model was built using TensorFlow,and a remote sensing interpretation platform was implemented through web frameworks such as Django and Vue.For the four types of remote sensing images of cultivated land,natural resources,roads and houses,this platform implements functions such as data upload,data preview,automatic interpretation,result preview,result download,etc.This platform also supports viewing and downloading of historical tasks.Finally,this thesis carried out a system test and interface display of the platform.