Deep learning performance for digital terrain model generation

Abstract

Traditional methods of photogrammetric image processing allow reconstructing a digital terrain model with high accuracy needed for producing different geographic information system (GIS) products such as maps, orthophoto etc. The accuracy of correspondence problem solution for good imaging conditions can reach the level of hundredth of a pixel. Recently deep learning techniques have been applied for dense stereo matching of images. They allow receiving depth information directly from stereo images and show impressive results for such tasks as scene segmentation, object detection and classification. The aim of the study was to evaluate the performance of deep learning techniques for accurate digital terrain models generation. Some of the state-of-the-art deep learning stereo matching models were evaluated along with original convolutional network. To get reliable accuracy estimation of the considered techniques the results of 3D reconstruction were compared with the reference surface obtained by 3D scanning. Also the imagery acquired from unmanned aerial vehicle during the mission on digitizing an archaeological site was used for performance evaluation by comparing with digital terrain model generated by photogrammetric technique.