| Abstract: | Abundant spectral information of hyperspectral images (HSI) provides rich information for HSI classification, which often brings high dimensional data resulting in the dilemma between the demand for fine data and the limited resources such as computation, storage as well as transmission band-width. To address this issue, we propose a deep hierarchical feature representation model based on three-dimensional adaptive sampling and improved iterative shrinkage-threshold algorithm (ISTA) for HSI classification. Due to the adaptive sampling, we improve ISTA with deep learning network for spectral–spatial feature representation since the ISTA is no longer applicable for the sampled data reconstruction. Through end-to-end joint learning, the proposed method can not only effectively reduce the required data, but also learn discriminative features for HSI classification, which will be meaningful for the HSI’s transmission from the space satellites and fast classification. Experimental results demonstrate the effectiveness and superiority of the proposed method on three public HSI datasets. |
