In order to alleviate the drawback that the spatial information of any pixel in a superpixel for generating the spatial-spectral kernel is totally determined by the same biased superpixel feature

especially for spatial information of the pixels located at the boundary

we propose an edge-modified superpixel based spatial-spectral kernel method for hyperspectral classification.On one hand

we combine the fixed window and superpixel to determine the homogeneous regions in a weighting strategy

in which the weights for pixels outside the fixed window are set to zero.Then we obtain the modified spectral-spatial kernel based on the weighted homogeneous regions.On the other hand

by considering the correlation among adjacent superpixels

we extract the spatial features among those superpixels to generate the inter-superpixel based spectral-spatial kernel.Finally

we combine the two spatial-spectral kernels in a convex way and employ support vector machine (SVM) for classification.Experimental results on two real hyperspectral data sets indicate that the proposed method could overcome the instability caused by superpixel-based spatial information extraction technique

and lead to better classification results than other state-of-the-art classifiers.