Abstract:

Based on minimum noise fraction transformation, we introduce a novel wavelet kernel method, which improves the minimum noise fraction transformation by replacing the traditional kernel function with the wavelet kernel function, for its feature of multi-resolution analysis can improve the nonlinear mapping capability of the kernel minimum noise fraction transformation method.The relevance vector machine classification of hyperspectral images is a new classification method which combines the novel kernel minimum noise fraction transformation with the relevance vector machine. Simulation results show that, the wavelet kernel minimum noise fraction transformation method reflects the nonlinear characteristics of the hyperspectral images. The proposed method is applied to the HYDICE data (shoot over in Washington DC Mall), and compared with the compare algorithm, its classification accuracy can be increased by 3%~8% and the classification precision of areas with small sample data can be improved effectively.

Key words: RVM;hyperspectral classification;kernel method;minimum noise fraction