Exploiting Spectral–Spatial Information Using Deep Random Forest for Hyperspectral Imagery Classification

Abstract

In recent years, deep learning methods have been widely applied to hyperspectral image (HSI) classification. Besides convolutional neural network (CNN)-based deep learning, deep random forest (RF)-based method, such as densely connected deep RF (DCDRF), was also developed for HSI classification which utilized the spectral–spatial information to improve the classification accuracy. In DCDRF, evenly distributed image patches with a fixed patch size are utilized to extract the spatial information of ground objects. However, the spatial information in each patch is not always correct, especially when the patch center is close to the edge of ground objects. In this letter, we propose a new classification method called spectral–spatial deep RF (SSDRF) which can fully utilize the spatial information existing in HSIs to further improve the classification accuracy. The joint region that combines both the fixed-size patch and shape-adaptive superpixel is proposed to exploit more accurate spatial information. The RF used in the classification model is replaced by extremely random forest (EF) to avoid overfitting. Moreover, the majority voting is conducted within superpixels and among different scales of superpixels to optimize the classification. The experimental results on three HSIs demonstrate that the proposed SSDRF can achieve satisfactory classification results and outperforms patched-based DCDRF.