Supervised hyperspectral multi-scale image convolution classification method

Abstract

The invention relates to the field of hyperspectral remote sensing intelligent information processing, in particular to a hyperspectral multi-scale image convolution neural network for realizing fine classification of ground surface coverage in a hyperspectral image scene in a supervised mode, and more particularly relates to a supervised hyperspectral multi-scale image convolution classification method. According to the invention, for a high-dimensional nonlinear hyperspectral data structure, a small amount of known sample information is supervised to be used for training a multi-scale graph convolutional neural network by constructing global, local and spectral index adjacency matrixes, so that the graph convolutional neural network can effectively adapt to hyperspectral data for feature learning and label prediction; therefore, the graph expression capability of the nonlinear characteristics of the hyperspectral data can be enhanced, and the precision of land cover classification and recognition can be improved.

Description

technical field [0001] The present invention relates to the field of hyperspectral remote sensing intelligent information processing, that is, to integrate multi-modal feature data and adopt a standard supervised learning method to train and design a multi-scale graph convolutional neural network to realize hyperspectral image classification. Specifically, it relates to a supervised Convolutional Classification Methods for Hyperspectral Multiscale Graphs. Background technique [0002] Hyperspectral data has the characteristics of "integration of map and spectrum", including high-resolution spatial and spectral information and rich electromagnetic spectrum and radiation features. The classification of land cover has strong advantages, and there are many challenges in the field of hyperspectral intelligent information processing and analysis. The acquisition and information processing of hyperspectral images is the process of high-dimensional signal acquisition and representa...

AI Technical Summary

Problems solved by technology

In particular, the existing graph convolution classification methods mostly use the hyperspectral image as the entire graph as input, and only use the spectral features, which not only has a high computational cost, but also does not consider the local spatial structure information embedded in the hyperspectral data, and its graph The representation learning process does not conform to the standard semi-supervised or supervised learning assumption of independent and identical distribution of samples

Images