Research On Ground Segmentation Method For Remote Sensing Image

Object segmentation of remote sensing image plays an important role in the research of satellite image field.It is an important basis for extracting objects from remote sensing image,and provides a reliable theoretical support for marine pollution monitoring and crop yield estimation.Remote sensing image is different from ordinary three-channel image(RGB image).It contains abundant channel types and a large number of spectral features.It is difficult for traditional unsupervised clustering and machine learning methods to make full use of the spatial features of image and to achieve high-precision object segmentation.Ordinary pixel classification methods,such as Support Vector Machine(SVM)and Random Forest(RF),can only extract pixel statistical features and vegetation index features,and can not obtain rich spatial information in remote sensing images.Besides,their capability of feature expression is limited,the segmentation results are rough and the recognition accuracy is not high.Moreover,the results of pixel classification obtained by statistical machine learning method need to be stitched to generate the image of segmentation results,which has a relatively high cost of time.In order to solve the above problems,we studied the segmentation model of ground objects,and designed several algorithms.The effectiveness of our algorithms are verified by experiments..In order to solve the problem of the low accuracy of traditional methods in remote sensing image segmentation task,we proposed a fusion model that combines deep learning based feature extraction with image semantic segmentation algorithm.Fusion model can extract the complex spectral features of remote sensing images,and employ attention mechanism to balance the attention of different pixels by CNN.In addition,we introduce and improved U-net model to enhance the information expression capability of remote sensing images,and to some extent,improve the segmentation accuracy of remote sensing images..In addition,in order to improve the accuracy of remote sensing image segmentation and solve the problem of unprecise edge recognition,we combined the U-net model with the conditional random field model(CRF)to form a joint learning segmentation and recognition method,using our improved U-net model as the front model of the joint learning framework.The conditional random field model(CRF)is transformed into a gradient-optimized RNN algorithm and serves as the back model.The front and back parts of the model to interact with each other so as to accurately obtain the information of the target location and the category of can have impact on each other so as to detect the location of objects and types.Finally,Experimental results on the open data set show that our method can achieve 86.1% accuracy of ground objects segmentation,which is a significant improvement..