Abstract: Hyperspectral images have abundant spectral information, and thus the adjacent single-band images usually contain similar and redundant information. To simultaneously reducing spectral redundancy and retaining more effective discriminant information, a novel classification algorithm based on multitask learning and sparse representation is proposed in this paper. In this method, the spectral redundancy is transformed into effective information and a band cross-grouping strategy is used to construct sub-tasks. Sparse representation model is introduced to jointly represent all classification tasks. Finally, the test samples are classified according to the cumulative residual of all tasks. An experiment is performed between multitask learning and the single task learning. The experimental results show that the classification performance of the former is better than the later.