Change Detection of Hyperspectral Remote Sensing Images Based on Deep Belief Network

To address the performance bottlenecks of existing methods for change detection of hyperspectral remote sensing (HSRS) images, a new scheme for change detection of HSRS based on deep belief network (CDHSRS-DBN) is proposed. First, the HSRS images collected at two different phases (dual-temporal HSRS images) are pre-processed and registered, and then the spectral-difference images of the dual-temporal images are computed. Next, the endmember spectrums and the abundance-difference images are extracted using the pixel unmixing method based on independent component analysis (ICA). The low-layer feature vector of deep learning in CDHSRS-DBN adopts the integrated feature vector that consists of the pixel spectral-difference vector, endmember abundance-difference vector and the pixel spectral feature angle vector. Finally, a deep belief network (DBN) model that contains multi-layer restricted Boltzmann machines (RBM) and a support vector machine (SVM) is devised, and the weights of connections between visible and hidden layers are adjusted through pre-training. The accuracy of change detection is further improved by fine-tuning all the weights via the SVM classifier. In order to evaluate the performance of the proposed CDHSRS-DBN method, four pairs of EO-1 Hyperion test images at different phases, which collected in four different experimental zones, are used as the test data and CDHSRS-DBN is compared with six other typical HSRS change detection algorithms (CDHSRS-SCD, CDHSRS-MPD, CDHSRS-ICA, IR-MAD, CD-PCA and PCCD). The experiments focus on the detection of land-use changes. The average recall and precision reach 90.39% and 87.10%, respectively. The average value of F-Score and time consumption is 0.8871 and 242.5, respectively. Experimental results demonstrate the better performance of CDHSRS-DBN to detect changes of multi-temporal HSRS images accurately and efficiently.