Hyperspectral image change detection method based on spatio-temporal joint graph attention mechanism

Abstract

The invention discloses a hyper-spectral remote sensing image change detection method based on a spatio-temporal joint graph attention mechanism, which introduces the spatio-temporal joint graph attention mechanism, utilizes the relevance between different time phase earth surface coverage to spread earth surface coverage information, and effectively makes up for the deficiency when a traditional convolutional neural network is used for extracting features. Meanwhile, the attention mechanism of the space-time joint graph performs feature propagation in a semi-supervised manner, so that the dependency of the network on the number of marked samples can be reduced to a certain extent. Besides, a network framework of a super-pixel-level branch and a pixel-level branch is introduced, the super-pixel-level feature and the pixel-level feature of the hyperspectral image are extracted respectively, the two branches cooperatively work in a complementary mode, and the change detection precision is effectively improved. Experimental results show that the overall precision of the method on three data sets of driver, farm and USA reaches 96.91%, 98.40% and 96.87% respectively, and the Kappa coefficients are 79.57%, 96.14% and 90.99% respectively.

Description

technical field [0001] The invention relates to the field of remote sensing image processing, in particular to a hyperspectral image change detection method based on a spatiotemporal joint graph attention mechanism. Background technique [0002] Remote sensing image change detection (CD) can identify changes in land cover in the same area at different times, and has been applied to such fields as geographic national conditions detection, land survey, urban change analysis, ecosystem detection, and disaster monitoring and assessment. Hyperspectral imagery (HSI) is a detailed spectral sampling in a wide spectral wavelength range, which can reflect the spectral information of ground objects with approximately continuous electromagnetic spectral reflection characteristics while obtaining the spatial information of ground objects, so as to obtain more detailed information. Ground change information lays the groundwork. [0003] At present, the research on hyperspectral change de...

AI Technical Summary

Problems solved by technology

There are the following problems: use a fixed-size convolution kernel to extract features, so its feature extraction ability will be limited by the receptive field; feature extraction can only be performed in Euclidean space, and cannot adapt to the difference between different land covers in hyperspectral images. Therefore, the correlation between land cover cannot be effectively used; the label acquisition of hyperspectral imagery is difficult

However, so far there has been no report on the application of graph attention networks to hyperspectral image change detection.

Images