Hyperspectral image classification using K-plane clustering and kernel principal component analysis

In this paper, we present a new approach for hyperspectral image classification. The pixels’ spectra are grouped into clusters in an unsupervised manner using an improved version of plane based clustering. Since the pixels containing the same substances are linearly correlated, the proposed plane-based clustering can effectively group the data points. Plane-based clustering is a more appropriate choice than point based clustering schemes for grouping the datasets which are distributed around hyperplanes instead of hyperspheres. Then, Kernel Principal Component Analysis (KPCA) is applied to each cluster individually to obtain multiple kernel vectors for each data point. Applying non-linear kernels, can greatly increase the discrimination power of the acquired features. The feature vectors are extracted by a weighted linear combination of the kernel components obtained from each cluster. We compute optimal weights using the cluster hyperplane parameters. Since the whole procedure is performed in an unsupervised manner, the proposed approach can enhance the generalization power of the extracted features. Then, morphological attribute filters are applied to the feature maps to effectively utilize spatial relations. Hence, the acquired compact feature vectors include both spectral and spatial information. SVM is used for classification. The experiments performed on three well-known hyperspectral datasets reveal the effectiveness of the proposed feature extraction approach.