Researches On Spectral-spatial Hyperspectral Image Classification Based On KNN Non-local Filtering

Hyperspectral remote sensing can capture images of the target region by tens to hundreds of consecutive and subdivided spectral bands. Each pixel of the hyperspectral image contains abundant spectral information, which provides the possibility for object recognition and classification with high accuracy. Therefore, hyperspectral image classification has become one of the core technologies of hyperspectral remote sensing technology. However, hyperspectral image classification is also facing many difficulties:How to deal with the Hughes phenomenon of high dimensional data, how to use the spatial information of hyperspectral images and how to select training samples. In this paper, a KNN non local filtering method is proposed for classification of hyperspectral images, namely using of support vector machine (SVM) and K-Nearest Neighbor (KNN) non local filtering to extract the spectral information and spatial information of hyperspectral images for classification. Furthermore, Active learning is applied to reduce the human efforts spent on labeling.1. Classification of hyperspectral images based on KNN filtering:First, the support vector machine (SVM) is adopted to obtain the initial classification probability maps which reflect the probability that each hyperspectral pixel belongs to different classes. Then, the obtained pixel-wise probability maps are refined with the proposed KNN filtering algorithm that is based on matching and averaging nonlocal neighborhoods. Finally, the classification result is obtained by choosing the maximum of probabilities. Experimental results show that the SVM in the classification process can be very good to deal with hyperspectral images of high dimensional data of the Hughes phenomenon; Non local filtering based on KNN can effectively improve the classification accuracy by using the spatial information of hyperspectral image.2. Hyperspectral image classification based on active learning and KNN non local filtering:Through using the optimization probability obtaining by the KNN non local filtering method, active learning iteratively guided classifier to select the most informative samples to label for the establishment of a set of compact and efficient training sample set.3. Hyperspectral image classification software system:The software system was programmed based on MFC application framework, computer vision open source library OpenCV and an external link library provided by MATLAB R2014b.The software system can achieve the reading and displaying of hyperspectral images, the tag of training samples, the classification of hyperspectral images and other functions.Experimental results show that the proposed method can dramatically improve the classification accuracy since it can make full use of spatial information of hyperspectral image effectively. Furthermore, the proposed method can be easily combined with active leaning, which effectively guides the classifier to select training samples, and thus greatly reducing the number of training samples.