Discriminative Marginalized Least Squares Regression for Hyperspectral Image Classification

Least squares regression (LSR)-based classifiers are rarely used for hyperspectral image classification. The reason is that their limited projections result in the loss of much discriminant information, and they focus only on exactly fitting samples to the label matrix while ignoring the problem of label overfitting. To solve this issue, discriminative marginalized least squares regression (DMLSR) is proposed to learn a more discriminative projection matrix with consideration of class separability and data reconstruction ability simultaneously. In the proposed method, Fisher criterion is employed to avoid the overfitting problem and enhance class separability; furthermore, a data-reconstruction constraint is imposed to preserve more discriminant information on limited projections, thereby enhancing classification performance. Experimental results on two hyperspectral datasets demonstrate that the proposed method significantly outperforms some state-of-the-art classifiers.