Deep Learning and Superpixel Feature Extraction Based on Contractive Autoencoder for Change Detection in SAR Images

• Published in: IEEE transactions on industrial informatics Vol. 14; no. 12; pp. 5530 - 5538
• Main Authors: Lv, Ning, Chen, Chen, Qiu, Tie, Sangaiah, Arun Kumar
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01-12-2018; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Change detection; Change detection algorithms; Clustering; Clustering algorithms; Coding; Deep learning; Feature extraction; Image detection; Image segmentation; Iterative methods; Land cover; Markov chains; Noise; Principal components analysis; Radar detection; Radar imaging; simple linear iterative clustering (SLIC); stacked contractive autoencoder (sCAE); Synthetic aperture radar; synthetic aperture radar (SAR) image
• ISSN: 1551-3203; 1941-0050
• DOI: 10.1109/TII.2018.2873492

Image segmentation based on superpixel is used in urban and land cover change detection for fast locating region of interest. However, the segmentation algorithms often degrade due to speckle noise in synthetic aperture radar images. In this paper, a feature learning method using a stacked contractive autoencoder (sCAE) is presented to extract the temporal change feature from superpixel with noise suppression. First, an affiliated temporal change image, which obtains temporal difference in the pixel level, are built by three different metrics. Second, the simple linear iterative clustering algorithm is used to generate superpixels, which tightly adhere to the change image boundaries for the purpose of acquiring homogeneous change samples. Third, a sCAE network is trained with the superpixel samples as input to learn the change features in semantic. Then, the encoded features by this sCAE model are binary classified to create the change result map. Finally, the proposed method is compared with methods based on principal components analysis and Markov random fields. Experiment results show that our deep learning model can separate nonlinear noise efficiently from change features and obtain better performance in change detection for synthetic aperture radar images than conventional change detection algorithms.