Abstract: Linear features extraction methods for hyperspectral imaging reduce feature class separability. Aiming to solve these problems, this paper introduces a novel kernel method based on minimum noise fraction transformation. This method focuses on the kernel function in the minimum noise fraction transformation and replacing the traditional kernel function with a wavelet kernel function. This new method improves the nonlinear mapping capability of kernel minimum noise fraction transformation by exploiting the features of multi-resolution analysis. Classification experiments on hyperspectral image data combined the novel kernel minimum noise fraction transformation and support vector machine; simulation results show that the wavelet kernel minimum noise fraction transformation method is suitable for the nonlinear characteristics of hyperspectral images. The proposed method was applied to HYDICE and AVIRIS data, and compared with other algorithms. Classification accuracy increased 3%-9%.