A Hyperspectral Image Classification Method And System Implementation Based On Sparse Convolution Filter Learning

Hyperspectral images(HSIs)typically contain hundreds of continuous spectral bands,which are rapidly developed in classification and recognition of land cover.Among them,HSI supervised classification methods aim to give information of the object category from the pixels in HSI,which can be widely used in many fields including mineral exploration,precision agriculture and environmental monitoring.However,due to the high dimensionality,multi-channel,unbalanced category and complex space information of the hyperspectral data,it is a research hotspot to combine spatial and spectral information to design a machine learning classification algorithm under limited supervision samples.This paper analyzes HSI classification method which based on sparse convolutional filter learning to make full use of the spatial-spectral features of HSI,so as to improve the accuracy and robustness of the classification results.The research work of this paper mainly includes the following aspects:(1)This paper study some basic HSI classification algorithms,including classification method based on composite kernel(SVM-CK),classification method based on multiple hypotheses(MH-SVM),classification method based on three dimensional discrete wavelet transform(3d DWT-SVM),and three classification algorithms based on convolutional neural network(CNN)of deep learning,including2D-CNN,3D-CNN and multi-channel convolutional neural network method(MC-CNN).We analyze the principles,advantages and disadvantages of each algorithm.At the same time,we analyze the experimental results and evaluate algorithm performance.(2)This paper propose a HSI classification algorithm based on sparse convolutional feature learning.Based on convolutional sparse coding,a group of convolutional filters are trained and learned from HSI,which are convoluted with the original HSIs to extract spatial features,and all the 2-D spatial features are stacked into a 3-D spatial-spectral feature map.The method of sparse convolutional feature learning realizes robust feature extraction.Experimental comparison shows that the proposed algorithm achieves excellent classification performance.(3)A multi-layer sparse separable convolutional model for HSI classification is proposed.The HSI classification task is modeled as a problem of deep feature extraction with sparse constraint.This model introduces sparse coding and utilizes the structural similarity(SSIM)algorithm to reduce the band dimension of HSI,so as to accelerate the process of effective feature learning.Rank-1 tensor decomposition realizes learning separable filters and accelerates the process of convolutional feature extraction;multi-layer cascade model achieves deep feature extraction of HSI.Experiments show that this algorithm is better than some classical deep learning classification algorithms.(4)This paper design and implement a system of HSI classification algorithm.The system integrates two classification algorithms and six contrast algorithms proposed in this paper,and implement image visualization as well as image classification of HSI.Specifically,it mainly realizes three function modules: image visualization,supervised classification,and performance evaluation.