3D CHANGE DETECTION IN URBAN AREAS BASED ON DCNN USING A SINGLE IMAGE

Abstract. In this paper, a novel approach is proposed for 3D change detection in urban areas using only a single satellite images. To this purpose, a dense convolutional neural network (DCNN) is utilized in order to estimate a digital surface model (DSM) from a single image. In this regard, a densely connected convolutional network is employed for feature extraction and an upsampling method based on dilated convolution is employed for estimating the height values. The proposed DCNN is trained using satellite and Light Detection and Ranging (LiDAR) data which are provided in 2012 from Isfahan, Iran. Subsequently, the trained network is utilized in order to estimate DSM of a single satellite image that is provided in 2006. Finally, the changed areas are detected by subtracting the estimated DSMs. Evaluating the accuracy of the detected changed areas indicates 66.59, 72.90 and 67.90 for correctness, completeness, and kappa, respectively.