Hyperspectral image unmixing method based on multi-layer stacked autoencoder

Abstract

The invention discloses a hyperspectral image unmixing method based on a multi-layer stacked autoencoder, comprising the following steps: training three autoencoders one by one, forming a stacked autoencoder, and extracting high-order spectral features of an image; constructing a multilayer autoencoder and use the parameters trained in step 1 to initialize the multi-layer autoencoder network; use the gradient descent algorithm to train the multi-layer autoencoder network until the weight between the input layer and the output layer of the multi-layer autoencoder network structure error is the smallest; the present invention realizes the learning of high-order spectral features of hyperspectral images by using stacked autoencoders, provides better initialization for multilayer autoencoders, and adds sparse terms and regularization to the loss function of multilayer autoencoders The transformation term constrains the endmembers and abundance to improve the unmixing accuracy. This method does not require any prior knowledge, performs unsupervised learning, has strong generalization ability, and has achieved good results in endmember extraction and abundance estimation. Effect.

Description

technical field [0001] The invention relates to the technical field of hyperspectral image unmixing, in particular to a hyperspectral image unmixing method based on a multi-layer stacked automatic encoder. Background technique [0002] The widespread existence of mixed pixels has become a prominent problem restricting the application of hyperspectral remote sensing. It not only affects the recognition accuracy of ground features based on hyperspectral images, but also has a great impact on image target detection and sub-pixel classification. Therefore, , how to effectively solve the problem of mixed pixels is one of the problems faced by hyperspectral image processing technology; [0003] At present, the most effective method to solve the problem of mixed pixels is mixed pixel decomposition. The unmixing technology has been developed for decades. The traditional unmixing methods at home and abroad are mainly divided into four categories: geometric analysis methods, statistic...

AI Technical Summary

Problems solved by technology

The autoencoder automatically learns the characteristics of the data by reducing the reconstruction error, and the decoding process matches the unmixing process, making it an ideal model to solve the problem of mixed pixels; however, many autoencoder algorithms unmix with a single automatic The encoder is a model, which is not conducive to the learning of high-order features of the image and thus affects the unmixing accuracy;

[0005] Most of the existing hyperspectral image unmixing methods use traditional algorithms, which are divided into linear unmixing models and nonlinear unmixing models. Although they can obtain better unmixing effects, they also have the following disadvantages: The unmixing model of the unmixing model requires prior knowledge or corresponding feature information, and the universality of the model is poor. At the same time, it also requires a large number of input parameters, which brings difficulties to practical applications; 2) Based on the linear unmixing model, the endmembers are extracted and abundance estimation are carried out step by step, it is easy to produce errors in the first step that affect the abundance estimation, and then cause low unmixing accuracy to have adverse effects on target detection and sub-pixel classification. Therefore, the present invention proposes a multi-layer stack method based Hyperspectral image unmixing method for autoencoders to address deficiencies in existing techniques

Images