Research On Improved Hyperspectral Image Feature Learning Algorithm Based On ELM Autoencoder

Hyperspectral image is a combination of spectral and image information,which is also one of the most prominent developments in remote sensing technology since the 20 th century.It has made significant contributions to the development of medical,mineral exploration and agriculture.The significant difficulties in processing hyperspectral data are a large amount of data,complex structure as well as redundant information and noise.This paper aims to study the hyperspectral feature learning algorithm that can eliminate noise,fully exploit data information and process at a fast speed so that reliable and efficient processing of hyperspectral image data can be achieved.In this thesis the related principles of extreme learning machine and auto-encoder are studied.Extreme learning machine has the features of fast speed and excellent generalization ability,and the auto-encoder based on the extreme learning machine can quickly learn the characteristics of the data by taking advantage of the extreme learning machine.For hyperspectral data,this thesis analyzes the spatial and spectral characteristics of hyperspectral images and introduces the MH prediction algorithm.To improve the feature extraction efficiency and classification accuracy of hyperspectral image,two algorithms are designed based on ELM-AE:(1)Improved MH-ELM-DAE hyperspectral feature learning algorithm based on Gaussian white noise and spatial spectrum;(2)Improved MH-HELM-AE hyperspectral feature learning algorithm based on deep learning and spatial spectrum.In this thesis,two algorithms are used to make experiments on four data sets to verify the effect of the improved algorithms,and then we compare the experimental results of the Indiana Pines data set with the results of other algorithms.The experimental results show that both the improved MH-ELM-DAE and MH-HELM-AE improve the feature extraction effect of the original data.In comparison with other algorithms,MH-ELM-DAE and MH-HELM-AE excel in accuracy,and the speed of two algorithms are far ahead of other algorithms under the premise of reliable accuracy.Therefore,the improved algorithm can provide reliable accuracy for hyperspectral data processing.At the same time,it accelerates the feature extraction speed,reduces the consumption of hardware resources and enriches the research of hyperspectral data processing.