Metric learning and local enhancement based collaborative representation for hyperspectral image classification

Collaborative Representation (CR) models have been successfully employed for Hyperspectral Images (HSIs) classification because of the effectiveness and simplicity. However, when dealing with high-dimensional and noisy HSIs data, CR models which try to seek an approximation of each testing sample by a linear combination of training/dictionary data with different regularizators in Euclidean space could be ineffective. Although there are some variants of CR adopt various distance measurements like Mahalanobis distance or subdictionary data to address these issues, the performance of these models could be compromised due to the inaccurate distance calculation between the testing and dictionary/subdictionary data. In order to further improve the performance of CR models, we propose three novel CR models for HSIs classification. Firstly, Metric-learning based CR (MCR) adopts metric learning method to adaptively learn the correlation between the spectral bands from HSIs data which leads to more accurate distance measurement in CR. Then, Local-enhancement based CR (LCR) tries to enhance the contribution of the neighbor training samples and impose penalty over the non-neighbor training samples of each testing sample in the regularization term of CR models, leading to effective local structures representation. Finally, Metric-learning and Local-enhancement based CR (MLCR) is proposed to combine MCR with LCR, which could provide better classification accuracy. To further consider the spatial information in CR, the Metric-learning and Local-enhancement based Spatial-aware CR (MLSaCR) is also developed. The experimental results on three HSIs datasets demonstrate the effectiveness of the proposed models where the novel models outperform the classical and state-of-the-art CR models for HSIs classification.