A Classification Method of Spectral Tensor Dimensionality Reduction Based on Multilinear ICA

Abstract

The invention discloses a classification method of spectral tensor dimensionality reduction based on multi-linear ICA. In the method, factors affecting the spectral characteristics of ground objects are regarded as intra-class factors, and intra-class factors, class and pixel spectra are respectively regarded as a mode Construct a third-order tensor, and perform dimension reduction based on low-rank tensor decomposition; perform multi-linear ICA (independent component analysis) decomposition on the third-order tensor D to obtain a space-like matrix C class , Intraclass factor space matrix C within‑class ; classify the class-free test hyperspectral image d using a supervised classifier. The invention can classify hyperspectral images after the model is established without adjustment, while other tensor modeling methods need to repeatedly set and adjust parameters to achieve the best classification effect; the invention maps all pixel spectra of a class to the same coefficient vector, so that the influence of various factors is minimized, which not only improves the classification accuracy, but also stabilizes the result; when classifying the spectrum of an unknown pixel, it can be deduced which factor it is affected by.

Description

technical field [0001] The invention belongs to the technical field of remote sensing image processing, and in particular relates to a classification method of spectral tensor dimensionality reduction based on multi-linear ICA. Background technique [0002] Hyperspectral images provide a detailed and rich description of the spectral characteristics of ground objects, greatly improving the classification ability of ground objects, and have been widely used in geological exploration and earth resource investigation, urban remote sensing and planning management, environmental and disaster monitoring, precision agriculture, surveying and mapping and archaeology. [0003] However, hyperspectral images are composed of a large number of band data, and these bands constitute a high-dimensional feature space, and its processing requires a huge amount of calculation, causing a "data disaster". To solve this problem, the most effective way is dimensionality reduction. Principal Compo...

AI Technical Summary

Problems solved by technology

However, this method only simply models the hyperspectral cube as a third-order tensor, without considering the real reason that affects the accuracy of hyperspectral classification: the spectral characteristics of ground objects are affected by various factors such as illumination, mixing, atmospheric scattering, and atmospheric radiation.

Images