Change Detection In High-Resolution Remote Sensing Images Based On Deep Learning

In recent years,remote sensing technology and deep learning have developed rapidly,and remote sensing image change detection has made a major breakthrough in low-resolution image data sets.This technology plays an extremely important role in the national economy and national defense construction.Intelligent and accurate change detection algorithms are of great significance for exploring the potential of multi-temporal remote sensing image data.For high-resolution remote sensing images,problems such as highly complex spatial structure,difficult preprocessing,and difficulty in manual labeling exist,resulting in degradation of performance of most algorithms.In view of the great achievements in the field of computer vision,deep learning has been applied to remote sensing image change detection tasks,and has obtained high detection accuracy.Based on the research on high-resolution remote sensing image characteristics and change detection algorithms,two variation detection models based on deep convolutional neural networks are proposed.This article mainly includes the following two tasks:(1)This paper proposes a change detection model based on Gaussian pyramid and deep convolutional neural network.This model introduces a multi-scale representation of the image in the Gaussian pyramid,improving the difference between the sample in the data set and the unchanged sample.Firstly,Gaussian blur and downsampling are performed on multi-temporal pixel blocks.Then,the basic network model is used to extract difference information from multi-temporal pixel blocks of different scales,and extract feature information such as contour and background from pixel blocks with lower resolution.The spatial detail information such as texture is extracted from the higher resolution pixel block,and finally the high-level semantic feature is merged by the full-connection layer,and the central pixel variation probability is output through the logical regression layer.Experimental results show the effectiveness of the algorithm.(2)This paper proposes a change detection model based on Hebbian sparse theory and deep convolutional neural network.The sparse convolution module designed in this paper adopts multi-pass design.Each channel uses different numbers of 3*3 convolutional layers and 1*1 convolutional layers to extract features of different grades and different scales,including 1*1 convolutional layer.To achieve information integration across channels,the highly correlated features in the same spatial location but different channels are aggregated to make the model sparse in the channel direction.The sparse model can greatly reduce the interdependence of parameters,and to some extent alleviate the occurrence of over-fitting problems,so that the model has efficient learning ability and high-capacity expression ability.Experiments on high resolution remote sensing images of different scenes show the effectiveness of the algorithm.