Dual Linear Regression in Tensor Space for Hyperspectral Image Classification

Recently, linear regression has been a popular image classification technique since it is efficient and easy to implement. However, the current traditional regression methods require to convert samples to vectors even if the hyperspectral image (HSI) has a three-dimensional structure. Furthermore, the traditional linear regression methods can not fully utilize the spatial information in HSI data, which plays an important role in HSI classification. To address these problems, this paper proposes a novel dual linear regression in tensor space (TSDLR) for HSI classification. Specifically, the subspace representation of the HSI data is generated via tensor decomposition-based regression and the ridge regression is adopted to reveal the potential relationship between the subspace representation and label information. More importantly, the intrinsic spectral-spatial information of HSI is well discovered for classification by the tensor decomposition-based regression. The experimental results on two real HSI datasets demonstrated the effectiveness of the proposed TSDLR method.