Build-up Land Extraction From Medium And High Resolution Remote Sensing Images

With the development of sensor technology,remote sensing technology has gradually become an important method to observe the earth`s surface.Deep learning has achieved considerable development and good effect in the field of computer vision and natural language processing.Therefore,deep learning algorithm is gradually introduced into the field of remote sensing image processing such as land cover classification,surface change detection,target detection in remote sensing image,etc.Different from the traditional remote sensing image processing methods,such as remote sensing index threshold method,support vector machine,decision tree classifier,etc.constructed features artificially.Deep learning can automatically extract image features by using learnable weights in the neural network,and can also better fit the essential features of data.In this paper,the application of deep learning semantic segmentation methods in the field of land cover classification and artificial surface coverage extraction is studied and some improved methods are proposed.This paper mainly studies the how to transfer the semantic segmentation method in field of computer vision to the task of artificial surface coverage extraction and improve the performance of remote sensing images.Secondly,this paper improves the performance of the semantic encoder by using sparse Non-local spatial attention.Optimize the decoding performance by integrating the channel attention mechanism into the decoder for fusing the high-level feature and low-level feature.Third,considering the problem of uneven distribution of sample,an online hard example mining algorithm is introduced to the cross-entropy loss function to help the convergence of model.Finally,according to the characteristics of remote sensing images,combing with the method of multi data fusion,this paper improves the application of deep learning in remote sensing images,and proposes a set of methods for artificial surface classification extraction of medium-resolution remote sensing images.In this paper,the feasibility of algorithm improvement has been analyzed on the DeepGlobe public data set firstly.Compared with Deeplab-v3 +,the algorithm proposed in this paper has a performance improvement of 4% on the mean intersection over union.In addition,based on Gaofen satellite images,this paper also built a medium-resolution artificial surface semantic segmentation data set,which was applied to the task of 16-m artificial surface coverage extraction in China.Through the method proposed in this paper,270 Gaofen-1 satellite remote sensing images were used to extract 16-m artificial surface coverage in the whole China region,and the recognition accuracy was verified in the experimental sample area.Compared with the traditional support vector machine classification method,the algorithm proposed in this paper has higher accuracy and faster extraction speed,and eliminates the complex process of manually selecting sample points to build features.