A Hyperspectral Unsupervised Classification Method Based on Confidence Constructing Generic Dictionary

Abstract

The invention discloses an unsupervised classification method for hyperspectral images based on constructing a generic dictionary based on confidence. The method first constructs a two-dimensional spectrum-pixel matrix; normalizes rows and columns; extracts and selects features, and obtains each pixel Dimensionality reduction features; coarse classification and confidence evaluation, that is, using dimensionality reduction features to classify each pixel and calculate the Euclidean distance between each spectral pixel and its rough classification category center, as the confidence level, to obtain high confidence classification samples and low Confidence classification samples; finally, secondary classification is performed based on kernelized sparse representation, that is, a category dictionary is formed from high-confidence classification samples, and kernelized sparse representation is performed on low-confidence classification samples, and the low-confidence classification is determined by the minimum category reconstruction error Classification labels for spectral cells. The invention overcomes the problems of insufficient classification subspace description accuracy and high computational complexity caused by directly using all spectral data to construct a dictionary, improves the discrimination of dictionary subspaces, and reduces the error rate of misclassification.

Description

technical field [0001] The invention belongs to the unsupervised technical field of hyperspectral images, in particular to a hyperspectral unsupervised classification method for constructing a category dictionary based on confidence. Background technique [0002] Hyperspectral image data is decomposed according to dozens or even hundreds of continuous narrow-band wavelengths in the spectral interval, which has rich spectral feature information of ground features and can be widely used in fine classification of ground features, mineral surveys and other fields. How to use the massive data and high-dimensional characteristics of hyperspectral images to combine various features of hyperspectral images to study fast and efficient target recognition and classification algorithms has always been a hot spot in the research of hyperspectral image processing. [0003] For hyperspectral image unsupervised classification methods, classical methods often have low classification accuracy...

AI Technical Summary

Problems solved by technology

By using the original data samples of the image, some or all of the spectral pixels are used to construct a sparsely represented dictionary. Although this method is simple in construction and low in computational complexity, the discrimination of the sparse subspace is insufficient, which leads to this method constructing a dictionary for sparse representation. The misclassification rate is relatively high

The method of dictionary learning, the representation dictionary obtained through sample learning can better match the structure of the image itself, and has a sparser representation, but the computational complexity is high

Images