Research On Change Detection Method Based On Hyperspectral Remote Sensing Data

Remote sensing image change detection is a detection technique that identifies surface changes by repeatedly observing the same area.It is widely applied in urban layout,river monitoring,and disaster assessment.In recent years,with the development of hyperspectral remote sensing technology,hyperspectral remote sensing images have received widespread attention.Compared with other images,hyperspectral images contain rich spectral information and have a high spectral resolution,which has great advantages in distinguishing subtle land changes.Therefore,change detection based on hyperspectral images has become one of the research hotspots in recent years.However,there are still several issues at present: 1)the problem of nonlinear relationships between the spectral curves of the two time periods caused by changes in atmospheric and solar illumination;2)the problem of complex terrain targets making it difficult to accurately extract the features of each target;3)the problem of sample unevenness in hyperspectral images caused by the number of changed pixels being much smaller than the number of unchanged pixels.In this dissertation,a series of research has been carried out around hyperspectral image change detection in response to the above issues.Meanwhile,this dissertation combines hyperspectral change detection methods with hyperspectral feature generation to solve the problem of poor results in multispectral images change detection.The main research content of this dissertation includes the following four aspects:(1)This dissertation proposes a hyperspectral change detection method based on spatial weighted kernel spectral angle constraint to address the nonlinear problem between two temporal hyperspectral images.This method is based on the idea of image transformation,with a simple model,fast operation speed,and strong applicability.In this dissertation,the minimum spectral angle constraint and nonlinear function are used to transform the original image to reduce the impact of nonlinear points on the detection results,so that the changed pixels can be more easily distinguished in the new variable space.In addition,the spatial weight map is generated by the similarity of local space to assist in change detection,thereby reducing the interference of noise on the results.The experimental results show that compared with classical machine learning methods,the AUC value of this method has improved by 0.0067～0.0324.(2)To better utilize the feature of spatial-spectral integration in hyperspectral images and enhance the ability of image feature extraction,this dissertation proposes an unsupervised hyperspectral change detection algorithm based on Transformer boundary autoencoder,which uses the autoencoder to extract image features in an unsupervised manner.The encoder in this method consists of a Transformer and a spectral attention module,which extracts global features while selecting feature channels through a spectral attention mechanism,increasing the weight of important feature channels,and reducing feature redundancy.To address the difficulty of edge extraction for changed targets,a boundary reconstruction module is designed to assist the encoder in learning boundary features.The overall accuracy of this algorithm on three public datasets is better than 94%.Compared with classic unsupervised methods,the overall accuracy of this method has improved by 0.48%～1.24%.(3)To further enhance the ability of feature extraction and improve the accuracy of change detection in specific scenarios,this dissertation proposes a supervised hyperspectral change detection algorithm based on adaptive convolutional kernel.To solve the problem of uneven samples in supervised methods,this dissertation designs a weight loss function.In the adaptive convolutional kernel,two kinds of convolutional kernels are used to extract features from images,and feature fusion is performed after generating trainable weights for each pixel.This structure can effectively extract features of multi-scale targets by automatically selecting convolutional kernels.The new loss function allows the pixels that may change to take a larger weight in the training process to prevent the network from overfitting.The overall accuracy of this method is optimal on several datasets,which improves by 2.62%～5.19% compared to classic deep learning methods.(4)To increase the performance of change detection in multispectral images,this dissertation proposes a multispectral change detection method based on hyperspectral feature generation.This dissertation designs a structure that combines hyperspectral feature generation and change detection.This method utilizes hyperspectral data to assist in training the network.Once the network is trained,the multispectral data is first processed through a feature generation network to generate hyperspectral data with a large amount of spectral information,and then undergo change detection.The experimental results display that this algorithm can make the enhanced spectral information more accurately reflect the features of ground objects,and the overall accuracy of multispectral images change detection has been improved by 2.46%～2.76%.