Unsupervised Change Detection Technology In High Resolution Multispectral Remote Sensing Images Based On Superpixel And Siamese Convention Neural Network

With the rapid development of satellite remote sensing technology,high resolution multispectral remote sensing images have become an important data source for change detection.High resolution remote sensing images contain richer information,but also introduce more interference factors,presenting new challenges for how to use these information in target change detection reasonably and effectively.In recent years,due to its strong expressive ability and high generalization ability,the application of deep learning technology in the field of remote sensing is becoming more and more in-depth,and put forward a new idea for the optimization of the change detection algorithms.Through the investigation and analysis of the research status of change detection in remote sensing images at home and abroad,this paper summarizes the technology of change detection as six phases:data source selection,data preprocessing,change feature extraction,change analysis,post-processing and accuracy evaluation.On this basis,aiming at the characteristics of high resolution remote sensing images,which are large data volume,high precision,complex features and hard to label,this paper proposed an unsupervised change detection scheme based on superpixel and Siamese convolutional neural networks.Firstly,after necessary preprocessing,such as orthorectification,image registration and histogram matching,the multi-temporal remote sensing images are segmented using superpixel segmentation and integration algorithm.The calculation of local features and the selection of samples are performed in units of superpixels to realize the automatic labeling of significantly changed and unchanged areas in the image.Afterwards,a Siamese convolution neural network is trained with these labeled results to classify image changes,and then simple post-processing methods such as noise reduction and morphological filtering is performed to obtain the final change detection result.Experiments show that the indicators of this scheme are much better than the traditional change detection algorithms on the GF-2 satellite remote sensing image dataset,the kappa coefficient increases 0.3 on average,the average overall error is less then 3.5%,and the results have better practical value.