Research On Hyperspectral Image Feature Extraction And Classification Algorithm

Hyperspectral remote sensing technology realizes the imaging of ground scenes through hundreds of spectral narrow bands,with high spectral resolution,continuous wave bands and strong feature recognition capabilities.It is widely used in geological mapping and exploration,atmospheric or vegetation ecological monitoring,product quality inspection,precision agriculture,urban remote sensing and military battlefield reconnaissance.Hyperspectral remote sensing images have the characteristics of high redundancy,strong correlation and large amount of data,which bring many challenges to image classification and recognition.In view of the characteristics of hyperspectral images,this paper studies the feature extraction algorithms and image classification methods based on spectrum and space based on the existing image feature extraction methods and classification recognition methods.The main contributions as follows:(1)Research and analyze the current hyperspectral image feature extraction method,hyperspectral image pattern recognition method,and introduce commonly used hyperspectral image evaluation classification indicators and data sets.(2)A hyperspectral image feature extraction algorithm with hierarchical fusion of spectral space features is proposed and classified on an unsupervised extreme learning machine.For hyperspectral images,the unsupervised extreme learning machine will produce a long calculation time and low classification accuracy during classification.To solve this problem,a layered fusion algorithm of spectral space features is added to the unsupervised extreme learning machine.The algorithm first uses the spectral dimension-based feature extraction algorithm LDA to reduce the original hyperspectral image data to a certain dimension and extract spectral information.Then the multi-scale adaptive weighted filter AWF is used to filter the dimensionality-reduced data to extract spatial information.The two algorithms are used alternately to design a layered fusion framework,which effectively extracts important spectral space information in hyperspectral images.In order to verify the effectiveness of the proposed algorithm,three hyperspectral images were used for experiments.The experimental results show the effectiveness of the proposed algorithm.(3)The algorithm of deep network of hyperspectral image classification fused with propagation filtering is proposed.This method includes three steps: preprocessing,deep feature extraction and classification.The pre-processing uses propagation filtering,which is used to smooth the image,remove noise,shadows and other information that is not related to the spectral characteristics,and obtain a hyperspectral image whose spectral characteristics are closer to the actual.The deep feature extraction part is composed of LPP algorithm and AWF algorithm,and performs hierarchical cross-learning of spectral features and spatial features to obtain deep spectral empty features.Classification uses a weighted kernel extreme learning machine as the classifier.In order to verify the feasibility of the proposed algorithm,two hyperspectral images were used for experiments.The experimental results show that the proposed algorithm has a better effect.