Guided Filter of Random Patches Network and Relaxed-Collaborative-Representation-Based Hyperspectral Image Classification

Feature extraction and accurate classification are crucial tasks in the land-cover classification of the hyperspectral image (HSI). We propose a guided filter (GF) of a random patches network (RPNet) and a relaxed collaborative representation (RCR)-based HSI classification (HSIC) method called GRR. The shallow and deep features are extracted using RPNet that requires no pretraining stage. In addition to the obtained feature set, the original HSI and extracted features are then filtered by GF to preserve the edge details. After that, all the distinct feature sets are separately concatenated with the original HSI to keep the original structure of the data. The high-dimensional feature sets are then processed by a linear discriminant analysis (LDA) to increase class separability and to select the most representative features. Since few train samples are available in the HSIC task, the efficiency of LDA is improved using superpixel segmentation to generate pseudosamples. In the final stage, the reduced-dimension feature sets are classified by the use of superpixel-guided RCR, which utilizes the resemblance and discrimination of the feature sets efficiently. The extensive experiments on the real HSIs are carried out to validate the efficacy of the proposed method.