Research On The Hyperspectral Image Classification Algorithms Combining With The Selection Of Spatial Neighborhood

The rich spectral information of hyperspectral images makes it widely applied to target detection terrain classification、pollution prevention military reconnaissance, etc. The hyperspectral image classification being an important content in the processing and analysis of hyperspectral image attracts much attention. However because of reasons like "metameric substance of same spectrum"、"metameric spectrum of same substance" and insufficient labeled samples with a high dimension, the accuracy of Hyperspectral image classification in the early time cannot be improved effectively with only the spectral information. The spatial information describes the relationships between pixels to be classified and its neighbors, so how can we jointly exploit both the spectral information and spatial information to improve the accuracy has become the focus of attention. So far researchers have done much work to combine the spatial information with the original classification algorithms and made satisfactory progress. Though, it is still the key point to mine and utilize spatial information effectively and rationally for the Hyperspectral image classification. With the mentioned, we do research on the spatial information for extracting more effective features and designing a neighborhood collaborative classification algorithm, and hope to improve the classification results lastly. The main contributions are as follows:1. we propose two algorithms which combine the selection of spatial neighborhood with composite-kernels SVM for Hyperspectral image classification. Both of them belong to the spatial preprocessing way and the main idea is to extract better spatial features for improving the classification results.1)Propose an algorithm combining the watershed segmentation with composite-kernels SVM(WSCSVM). Firstly we perform the watershed segmentation on the image to mine the boundary information, and then we extract the spatial features from a more compact spatial neighborhood based on the selected segmentation regions. Through the composite-kernels SVM, we combine spatial features with spectral features and design our new algorithm for hyperspectral image classification.2)Propose an algorithm combining the MST neighborhood with composite-kernels SVM(MSTCSVM). We consider not only the boundary information but also the similarity between pixel and its adjacent pixels to select a more compact and discriminant neighborhood, then we extract the spatial features from it and design new algorithm combining spatial features with spectral features through composite-kernels SVM.The different advantages of two algorithms:the algorithm of WSCSVM has high efficiency, with just once segmentation, we can select appropriate spatial neighborhood for all pixels to be classified; while the algorithm of MSTCSVM needs to consider both the spectral and the gradient information for more effective spatial features, and we repeat the selection until the requirements meet which means more time and low efficiency. We do experiments on two datasets:Indian Pines and the University of Pavia (PU), and results prove the effectiveness of our algorithms.2. Propose an algorithm combining MST neighborhood with kNN for hyperspectral image classification. Firstly we denoise the image with bilateral filtering and select the MST neighborhood for pixels to be classified, further we classify them by our proposed neighborhood collaboration kNN algorithm exploiting the kNN results of adjacent pixels. Experimental results on Indian Pines and the University of Pavia (PU) datasets show our algorithm improve the accuracy effectively.