Hyperspectral classification with limited training samples is challenging. The current work lies in two aspects: first, we change the statistical distribution of samples by iterative filtering based on the guide images. The filter is called a Simplified Bilateral Filter (SBF), which is a modified bilateral filter for clustering samples. Secondly, new structural convolution kernels are used to generate new hyperspectral data. Finally, the class label of the test sample after dimension reduction is determined by OMP classification or SVM classification. Experimental results on two hyperspectral datasets demonstrate the effectiveness of the proposed feature extraction method in improving classification accuracy with limited training samples.