Remote Sensing Change Detection Based on Deep Belief Networks Optimized by Domain Knowledge

  Objectives  A method of deep learning change detection with domain knowledge as an optimization strategy was proposed to improve the change detection precision of high-resolution remote sensing im‍ages.

  Methods  The improved change vector analysis algorithm and grey-level co-occurrence matrix algorithm were used to obtain the spectral and texture changes of images, and reasonable thresholds were set to divide the changed samples from the unchanged samples based on the spectral and texture change intensity maps. The pattern shape index and spectral knowledge in domain knowledge were introduced as an optimization strategy to filter the changed samples for obtaining high-quality training samples. The deep belief network model was constructed and trained, and the results of deep learning change detection were optimized by the optimization strategy to reduce the influence of "salt and pepper noise" and false change zones on the detection accuracy.

  Results  The Results of change detection experiments show that the accuracies of Gaofen-2 and IKONOS imageswere increased by 7.58% and 14.69% and the recall by 17.08% and 23.87%, respectively, while the false alarms and were decreased by 30.22% and 23.30% and the missing alarms by 17.08% and 23.87%, respectively.

  Conclusions  Compared with the method before the optimization strategy was adopted, the proposed method in this paper can effectively improve the precision of change detection, and it provides a new way of using remote sensing images to improve the precision of deep learning change detection.