Feature Extraction of Hyperspectral Images Based on Deep Boltzmann Machine

High dimensionality and lack of labeled samples are the difficulties in feature extraction for hyperspectral image (HSI) processing. In this letter, a deep-learning-based feature extraction method is proposed. First, the guided filter is used to preprocess the original HSI data. The result data contain the joint spectral and spatial information of the objects. Second, the local receptive field and weight sharing are introduced into deep Boltzmann machine(DBM) to establish a novel feature extractor, called local-global DBM (LGDBM). The LGDBM has two advantages: 1) it can learn both the local and global features of the high-dimensional input data and 2) it has much fewer parameters than the DBM. Therefore, only a few labeled samples are needed for training, and the local and global spectral-spatial features are extracted intrinsically. A group of classification experiments are performed to evaluate the advantages of the feature extraction method.