Remote Sensing Image Change Detection Based On Convolutional Neural Network

Change detection of remote sensing image is a task of analyzing and detecting multiple remote sensing images acquired at the same geographical location at different times,and obtain the region where the ground object type changes.The traditional change detection algorithm extracts change features from the difference map of multi temporal images,and obtains the change map by threshold segmentation.The quality of difference map and the selection of threshold value have a great impact on the accuracy of change detection.The quality of the difference map is often affected by the imaging factors,and the threshold value is usually difficult to determine,which leads to the low accuracy and low automation level of change detection.The convolution neural network can automatically extract the change class feature information and generate the change graph after a large amount of data training.Therefore,this thesis uses convolutional neural network to detect the change of remote sensing image,and the main work is as follows:(1)Aiming at the problem of uneven color distribution caused by band splicing in data set images,a chromatic aberration equalization and stretching algorithm is proposed to eliminate the trace of band splicing;In view of the problem that remote sensing image is too large to participate in training directly,this thesis proposes to preprocess the remote sensing image by data segmentation and augmentation to generate experimental data set.(2)Aiming at the phenomenon of over fitting in U-Net network training,an asymmetric convolution block is proposed to replace the standard convolution operation of U-Net network feature extraction module to enhance the robustness of convolution kernel and prevent over fitting;Aiming at the problem that the background of remote sensing image is complex and the change of small target is easy to be missed,this thesis proposes to introduce attention mechanism into U-Net network,adjust the weight of feature map,strengthen the learning of changing features,and extract more appropriate features for change detection.The experimental results show that,compared with the traditional method and the improved U-Net model,the accuracy of the improved U-Net model in change detection is significantly improved.(3)In order to solve the problem that the model can not make full use of the features of each image due to the existence of merging channel operation in U-Net network forchange detection,a change detection model of remote sensing image based on dual branch convolutional neural network is designed and implemented.The features of each image are extracted by two identical convolutional neural networks;Aiming at the problem that the information loss caused by the pooling operation of feature extraction affects the accuracy of change detection,a convolution layer with step size of 2 is proposed to replace the pooling layer to reduce the information loss of feature map.The experimental results show that the accuracy of change detection of this model is further improved compared with the change detection model based on improved u-net and the change detection model based on twin network,and the model has certain generalization ability.