Research Of Change Detection Method In Hyperspectral Images Based On Deep Siamese Network

Hyperspectral remote sensing image change detection technology is now a popular research direction for dynamic monitoring of land cover changes,and has been widely used in land resources survey,land using condition,disaster monitoring and other fields.Hyperspectral remote sensing images have rich spatial-spectral information,but traditional change detection methods have limited ability to express hyperspectral remote sensing image features,and it is difficult to identify complex detail features,semantic features and spatiotemporal correlation features in bitemporal hyperspectral remote sensing images.Therefore,it is of great research significance to effectively use hyperspectral remote sensing images to accurately identify land surface changes.This paper takes deep learning and metric learning as the core technology,introduces the spatial-temporal attention mechanism,and proposes a hyperspectral image change detection method based on the deep twin network to realize the spatial-temporal-spectral feature extraction and change detection of dual-temporal hyperspectral images.The specific research contents and innovative ideas are as follows:First,for the problem of hyperspectral data feature extraction and selection,a dual-branch feature extractor combining semantic segmentation network and deep Siamese network is proposed.A dual-branch model framework is built with the deep Siamese network as the main body,and the semantic segmentation network is used to extract spatial-spectral joint features rich in semantic information in each branch structure.Compared with the traditional method of using convolutional neural network to extract hyperspectral data features,it can simultaneously extract space-spectral features of dual-temporal hyperspectral data and introduce spatial semantic information,which enriches the feature expression information.Secondly,for the extremely unbalanced distribution of data categories in the hyperspectral image change detection task,a batch balance measurement algorithm is introduced,which can make different change categories have the same contribution to the loss function calculation under the condition of different data volumes,and the balance model has no effect on the change pixels.and invariant pixel sensitivity.Experiments prove that the algorithm can better detect the change area when applied to the hyperspectral image change detection task.Finally,for the joint representation of spatial-temporal spectral information of bitemporal hyperspectral data,a spatial-temporal spectral attention mechanism is proposed to focus on time-invariant space-spectral features,combining semantic segmentation network and deep Siamese network from two dimensions of space and channel Capture the inherent connection of image data and enhance feature recognition capabilities.Experimental results show that the proposed algorithm outperforms other comparative methods in detection accuracy.