A Deep Learning Approach to Detecting Changes in Buildings from Aerial Images

摘要

Detecting building changes via aerial images acquired at different times is important in the urban planning and geographic information updating. Deep learning solutions have high potential in improving detection performance as compared with traditional methods. However, existing methods usually carry out detection for whole images. Non-building interferences involved may result in an increase of false alarm rate, a decrease in accuracy rate, and a heavy computational load. In addition, they mostly utilize supervised deep learning networks dependent highly on massive labeled samples. In this study, we present an unsupervised deep learning solution with detection only on segmented building areas. We first employ a masking technique based on building segmentation to remove non-building interferences. We then use a classification model combing an unsupervised deep learning network PCANet and linear SVM to realize building change detection. Experimental results show that our method achieves 34.96% higher accuracy rate, 45.18% lower missed detection rate, 37.92% lower false alarm rate, and 50.12% lesser computational time than the whole-image detection method without building segmentation.