Hyperspectral Image Classification Based On Spatial Information And Network Learning

Because hyperspectral images contain abundant spectral curves of ground objects,it has become an effective tool for monitoring dynamic changes of the earth's environment and quantitative inversion of remote sensing.The analytical processing method of hyperspectral images has attracted more and more attention.Although the rich spectral information of hyperspectral images is beneficial to the classification of ground objects,too many spectral dimensions and redundant information bring some challenges for the classification algorithm,and the labeled samples in hyperspectral images are scarce.How to use effective feature extraction methods based on the characteristics of hyperspectral image data to provide discriminative information has always been the key problem.The multi-layer network learning method can describe the features of unlabeled data in multiple stages and extract various hidden features to form a more abstract semantic representation,which has better classification performance than the traditional and shallow method.Therefore,based on the multi-layer network learning model,this paper extracts the potential multi-layer features of the image,and combines the advantages of the spatial feature extraction method to study the more efficient classification algorithm.The main contents are as follows:(1)A hyperspectral image classification method based on Squeeze-and-Excitation aggregated residual network is proposed.The network model designed by this method combines the deep extensible structure of the aggregated residual network and the recalibration strategy of the Squeeze-and-Excitation module to the feature channel.It can effectively learn the spatial context information in hyperspectral images and selectively strengthen the feature that has important information,and extract deep level abstract semantic features.This method has excellent classification performance.(2)A hyperspectral image classification method based on multilayer extreme learning machine and active learning is proposed.Based on effectively representing of the spatial feature,this method utilizes the unique advantages of multilayer extreme learning machines in feature learning and data classification.Due to the fact that multi-layer networks dependent on the number of training samples,an ensemble network is constructed by using the idea of committee voting in active learning.In this process,unlabeled samples with large differences in prediction labels are selected to expand the training sample set and reduce the need for labeled samples in a more efficient way.(3)A hyperspectral image classification method based on edge-preserving filtering and deep network is proposed.From the angle of different texture information and multi-level learning,a large number of hyperspectral image data are studied unsupervised with the idea of reconstruction.First,the spatial information is extracted by the extended morphology profile after the principal component analysis,and the edge-preserving filtering method is used to effectively protect the edge.Then the feature set cascaded with the original spectral information,and the high-level feature of spectral-spatial information is extracted by the stacked autoencoder network.Finally,the softmax classifier is used to classify.