Least squares regression (LSR)-based classifiers are effective in multi-classification tasks. For hyperspectral image (HSI) classification, the spatial structure information usually helps to improve the performance, however, most existing LSR-based methods use the spectral-vector as input which ignore the important correlations in the spatial domain. To solve the drawback, a tensor-patch-based discriminative marginalized least squares regression (TPDMLSR) is proposed to modify discriminative marginalized least squares regression (DMLSR) with consideration of inter-class separability by employing the region covariance matrix (RCM). RCM is adopted to exploit a region of interest around each hyperspectral pixel to characterize the intrinsic spatial geometric structure of HSI. Specifically, TPDMLSR not only maintains the ascendancy of DMLSR, but also preserves the spatial-spectral structure and enhances the ability of class discrimination for regression by learning the tensor-patch manifold term with a new region covariance descriptor and measuring the inter-class similarity more accurately. The experimental results on membranous nephropathy (MN) dataset validate that TPDMLSR significantly outperforms LSR-based methods reflected in sensitivity, overall accuracy (OA), average accuracy (AA) and Kappa coefficient (Kappa).