A multiple-feature-based improved sparse representation (MFISR) method is proposed herein for the classification of hyperspectral images. The spectral feature, Gabor feature, and local binary pattern (LBP) feature are extracted from the hyperspectral image; subsequently, the sparse coefficients are solved and a 2-paradigm constraint is added. These obtained coefficients are used to determine the final class label of each test pixel. The experimental results demonstrate that the proposed MSIFR method exhibits excellent results for the detection of small samples, and its classification performance is stable and good.