Research On Target Change Detection Method Based On High Resolution Remote Sensing Image

Change detection based on remote sensing image refers to the process of extracting change regions from two or more remote sensing images of the same scene and different phases,which can judge whether each pixel in the scene has changed.At present,change detection based on remote sensing images has played a huge role in land use/coverage,disaster assessment,medical diagnosis,video monitoring and other fields,so it has attracted extensive attention.However,the traditional change detection has strict requirements on the image preprocessing,and many links need manual intervention,which will cause great interference in the processing of multi-source image data and the accuracy of the result is lower.With the continuous improvement of image resolution and computer performance,deep learning is gradually enter people’s vision.High resolution remote sensing image contains large amount of information and complicated,and deep learning can extract the abstract feature in the image and solve the complex mapping relationship.Therefore,this paper studies the change detection based on the deep learning,and resolves the problem of small sample.This paper mainly carried out the following aspects of work:(1)The traditional change detection method was studied and improved.First,the difference method based on OTSU was used to generate the result of change detection,and the algorithm was improved to calculate the threshold based on the traditional OTSU problems,and the threshold value was calculated by combining the distance between the two classes with no change and the distance within the class.At the same time,the evaluation parameter commonly used in change detection was used to evaluate and analyze the effect of the algorithm.The results of change detection were compared with those of manual calibration,and the overall classification accuracy was improved to 87.7%.(2)Aiming at the problem of small samples in deep learning,a method of sample amplification based on Deep Convolutional Neural Network(DCNN)was used.Based on DCNN network,the data set needed for change detection is generated.Since this paper using the original data set only 13,non-generative and generative data amplification methods are adopted to generate sample data respectively The data generated in the learning process of the network is selected as the sample data set,and the change detection area is manually marked to solve the problem of small samples.(3)In this paper,based on DeepLabv3 network whth higher accuracy in current classification algorithm,a remote sensing image change detection model based on deep convolutional network is constructed.The model realizs end-to-end training and prediction of remote sensing image change detection,and sub-pixel convolution is used to improve the algorithm.Spatial pyramid pooling module and encod-decoding module in DeepLabv3 have a very significant effect on pixel-level classification.Meanwhile,sub-pixel convolution is used to replace the deconvolution layer in DeepLabv3 to improve the up-sampling effect.The experiment proved that Kappa coefficient of the improved network increased from 0.64 to 0.75,and the improved network accuracy rate increased to 95.1%.Finally,the traditional method,the U-Net based change detection method and the improved DeepLabv3 method are compared to prove the superiority of the method combining deep learning and change detection.