Research On Semantic Segmentation Of Remote Sensing Images Based On Multi-source Data And Multi-scale Feature Fusion

With the development of remote sensing technology,the number of remote sensing images acquired is increasing,and their applications have become more extensive.The extraction of remote sensing image information is the basis of remote sensing image application.Among them,the semantic segmentation of high-score remote sensing images is the extraction of remote sensing information An important branch.However,there are many differences between high-resolution remote sensing images and ordinary images: First of all,there are many types of ground features in high-resolution remote sensing images,the problem of "same object with different spectra,same-spectrum foreign objects" is common,and the band data of high-resolution remote sensing images is limited,and the spectrum Insufficient information richness,a single data source affects the effect of semantic segmentation of high-resolution remote sensing images;secondly,there are differences in the size of different features in high-resolution remote sensing images,and there are differences in the location of the same kind of features,using a single-size convolution kernel,It is easy to miss the context information of the high-resolution remote sensing image;finally,the boundary between the classes in the high-resolution remote sensing image is not obvious,and the ground objects overlap each other,which is easy to produce the wrong division phenomenon.In order to solve the above problems,this paper deeply studies the semantic segmentation method of high-score remote sensing images based on deep learning,and proposes a dual-stream deep fully convolutional neural network model based on multi-source data and multi-scale feature fusion(TMdeep Net),and The error point correction loss is proposed to optimize the loss function.The main research of this article is as follows:(1)Aiming at the problem of “same spectrum,same spectrum foreign material” and insufficient spectral information of ground features in high-resolution remote sensing images,a dual-stream multi-source data feature extraction and fusion method is proposed.Based on the original remote sensing image,this method uses the spatial information and position information provided by multi-source data to make up for the lack of spectral information.At the same time,in order to avoid the mutual influence between multi-source data,the dualstream input method is used to separately extract the features of the spectral data and the multi-source data,and perform feature fusion on the extracted multi-source features.The experimental results show that the rational use of multi-source data can effectively improve the classification of complex features.(2)Aiming at the difference in area of different features in high-resolution remote sensing images and the position difference of the same feature in different remote sensing images,a depth separable convolution structure DDVR based on different void rates and an image pyramid based on void convolution are used Structure ASPP to extract multi-scale features from high-resolution remote sensing images.Among them,DDVR obtains the multiscale features of the image during downsampling through receptive fields of different sizes.ASPP performs multi-scale feature extraction and fusion of semantic features at the same location on the network.The experimental results show that the use of DDVR and ASPP structure can improve the segmentation effect of network features,while using separable convolution to compress the model,effectively reducing the model parameters.(3)In view of the wrong points in the classification results,a loss function Error point correction(EPC)was proposed.By increasing the module length of the prediction probability of wrong points and the prediction probability of correct points in the loss function,the network punishment for right and wrong points was increased.The experimental results show that EPC loss can effectively correct the wrong points in the segmentation results and improve the classification accuracy of the network.