A Hyperspectral Image Classification Method Extracted from Local to Global Context Information

Abstract

Aiming at hyperspectral images, the present invention discloses a deep learning hyperspectral classification method based on context information extraction based on a local-global attention mechanism, which mainly includes: a full-volume classification framework using encoding and decoding, which can take into account the global spatial spectrum at the same time Information, to achieve rapid classification; imitating the human visual perception mechanism, design a local to global context information perception network architecture in the encoder module of the network, to achieve advanced semantic feature extraction taking into account context information; in the decoder module of the network, design a A module of channel attention to achieve adaptive fusion of local-global information. The present invention can be applied to fine classification of hyperspectral images with massive high-dimensional nonlinear data structures, and greatly improves the phenomenon of misclassified isolated regions in hyperspectral image classification results. Requirements for real-time, rapid and fine-grained classification and mapping.

Description

Technical field [0001] The present invention belongs to the field of remote sensing image processing technology, particularly relates to a deep learning hyperspectral classification method based on the extraction of contextual information from local to global attention mechanisms. Background [0002] Classification has always been an important research area in hyperspectral remote sensing image processing and application, and its rich spectral information can accurately identify the attribute categories of features. At present, with the development of hyperspectral imaging technology, hyperspectral observation platforms such as spaceborne, airborne, and unmanned aerial vehicles provide rich data sources for fine identification of hyperspectral features. However, the high correlation between hyperspectral image bands, the high degree of nonlinearity of the data, and the "spectral variation" of homogeneous figures make model-driven classification methods seriously challenging in th...

AI Technical Summary

Problems solved by technology

However, this method still faces the following problems: (1) only local spatial information can be used, which will lead to misclassified isolated regions in the classification results; (2) the optimal spatial block size is difficult to determine, and its Determined by the spatial resolution and the homogeneity of the distribution of ground features

(3) The calculation cost is high. When predicting the classification map, this method needs to take spatial blocks pixel by pixel for prediction. For some relatively large images, it takes a very long time

Images