A Two Stage Learning Algorithm for Hyperspectral Image Classification

Since the excellent performance of Support Vector Machine (SVM) in handling with high-dimensional data, it is often used in the field of hyperspectral image (HSI) classification. However, traditional SVM methods only uses a single Mercer kernel function as base kernel, which does not represent the similarity of samples well. Meanwhile, it cannot utilize the spatial background information to enhance the HSI classification results. To address these issues, the paper proposes a two-stage learning (TSL) algorithm for HSI classification. In the first stage, a new Kernel Singular Value Decomposition-Multiple Kernel learning (KSVD-MKL) method is proposed for SVM Multiple Kernel Learning (MKL), which allows the best combination of kernels to be composed by using Gaussian kernels with different bandwidth scales. In the second stage, the KSVD-MKL classification is used as the initial spectral term classification results. Then, spatial information is modeled by using Conditional Random Field (CRF) observation fields and labels, and the KSVD-MKL classification results are optimized. Experiment results on public Indian pines and Botswana datasets demonstrate that the classification accuracy of the proposed method is effectively improved against existing algorithms.