Huge spectral-spatial information available in 3d-hyperspectral images is exploited to learn optimal dictionaries for classification. Unlike deep neural network-based methods, sparse representation for hyperspectral image processing requires low computational resources. In this work, specific constraints on the coefficients and the dictionary result in two convex objective functions that guarantee unique optimal solutions for the dictionary and the coefficient matrix. When coded with a single shared discriminative dictionary, pixels from the same region share the same sparsity profile. The proposed method uses the sparse representation of HSI data with adaptive elastic-net regularization over the coefficient matrix for robust spatial sparse feature extraction. These spatial features are inputs to a multilayer perceptron of a few hidden layers for classification through supervised learning. The experimental evaluation of the model on five HSI datasets shows that the model exhibits competitive performance with lower complexity compared to the state-of-the-art methods.