Classification Research Of Hyperspectral Images Based On Broad Learning

In recent years,the accumulative development and increasing maturity of remote sensing and imaging technologies have led to the widespread study and application of hyperspectral image classification.Unlike traditional images or multispectral images,hyperspectral images contain a greater number of spectral bands,enabling them to describe more information about an object.In general,the reasonable utilization of image information can identify objects accurately,which will contribute to various fields such as ocean exploration,defense and military,and atmospheric environment.Currently,most of the existing hyperspectral image classification algorithms are based on deep learning and are expected to achieve high classification accuracy.However,existing deep learning algorithms are usually subject to a series of limitations such as numerous parameters,complex models and time-consuming operation.Meanwhile,the problem of sample labeling cost has not been considered in the existing researches on the broad learning classification algorithms.To address the above problems,this thesis combines different scales of superpixel segmentation,active learning methods and broad learning systems for the classification of hyperspectral images,respectively,where the works are list as follows.（1）In Chapter 3,we firstly combine the superpixel segmentation,which can effectively utilize the spatial information of hyperspectral images,with broad learning system,and propose a multiscale kernel superpixel based broad learning system classification method.This method obtains the corresponding first principal component data from the spectral and spatial information,respectively.For enhancing the non-linear data processing capability of the system,each superpixel data is reduced in dimension using the kernel principal component analysis method.Then,after combining the reduced dimensional data into the broad learning system for training,we fuse all the sample results of the predictions obtained into two-dimensional data.Finally,the majority voting algorithm is employed to get the final classification results of the test set samples.From the classification results on the three datasets,it is evident that this method has superior generalization ability as well as the higher classification accuracy compared with other advanced algorithms.（2）In Chapter 4,to alleviate the problems of difficult sample annotation and timeconsuming training,we propose an active broad learning system based hyperspectral image classification method.Different from the existing classification methods,the proposed algorithm adopts a broad learning system with incremental learning as the classifier of the system,and instead of using all training samples,the high-quality samples selected and labeled by active learning are fed into the classifier for training,which can greatly reduce the repetitive training of samples and thus reduce the training time.In addition,we select the best one in this method for the proposed algorithms based on the classification accuracy value among three different active sampling strategies.To further verify the significant difference between the proposed method and the other advanced algorithms,the classification results obtained on the three datasets are evaluated using the Friedman test,from which the efficiency of the classification performance is proved.