Hyperspectral Image Classification Based On Active Learning And Spacial-spectral Information

Hyperspectral image technology is an important way to observe the ground objects.It overcomes the shortcomings of single-band and multi-band remote sensing images with low feature and less information about the features of the ground,which has made great contributions to modern military,agricultural,marine,ecological environments and so on.The hyperspectral images are characterized by combination of spacial-spectral Information and high spectral resolution,which provide a strong basis for the object recognition.However,scholars only use the spectral information without the spatial information in the early stages of classification,and the classification results are not very satisfactory.At the same time,it takes a great cost to obtain the labeled samples,and how to get the ideal classification result in the case of small sample becomes the research direction of the scholar.Support Vector Machine(SVM)-based active learning can solve this problem well,which chooses a small amount of information-rich marked samples with constant learning,and thus improve the performance of the classifier greatly.However,the traditional active learning only uses spectral information but ignores spatial information in the procedure of the selection of samples.Therefore,we propose the following methods of classification of active learning and spacial-spectral Information,which aims to obtain higher classification results with fewer labeled samples.The specific content is as follows:(1)Hyperspectral image classification technique of combination of active learning(AL)and semi-supervised learning(SSL)based on regional division strategy is proposed.In the algorithm,firstly,all the regions in the segmented graph are divided into two parts of trusted region and untrusted region by the region division strategy.Then,a sample is selected in each region,and the sample selected for different regions is marked by different ways,and then to achieve the purpose of combination active learning and semi-supervised learning.Among them,both of the spatial information and spectral information are taken into account when selecting the samples,which makes the selected samples more representative and achieves the effect of obtaining high classification accuracy in the case of few samples.(2)A method of combination improved active learning and spatial information for hyperspectral image classification is proposed.First,the representative samples of the boundaries of the segmentation map were chosen,which are rich in information and these can quickly improve the performance of the classifier.And then the improved active learning is executed,which is consistent with traditional active learning,the only difference is that the step of pre-processing of the test sample is added to the improved active learning in order to increase the number of the representative sample selected,so that the performance of the classifier is maximized.Finally,the neighborhood information of the sample is added to correcting the classification result of the improved active learning,so that the final classification accuracy can achieve the expected effect.(3)A hyperspectral image classification method based on segmentation result correction is proposed.The method is proposed for the problem of unreliability of the segmentation results of the area that including fewer pixels or more mixed pixels in the traditional method of classification results and segmentation map,and the final segmentation label of the region is obtained by fusing the statistical information of the region and its neighboring region,wherein the neighborhood region of the region is calculated by Fuzzy c-means(FCM).This fusion process not only considers the statistical information of the region itself but also refers to the statistical information of its adjacent region,which makes the segmentation result of regions with fewer pixels and more mixed pixels more reliable,and thus improve the final classification accuracy.