DTM extraction from DSM using a multi-scale DTM fusion strategy based on deep learning

Abstract

Extraction of digital terrain model (DTM) from Digital Surface Model (DSM) still faces many problems in a complex scene with geometric ambiguities such as steep slope forested environments and contiguous non-ground regions. In this paper, an approach based on deep learning is proposed to generate DTM directly from DSM without applying filtering methods for eliminating non-ground pixels. In this regard, first, in the preprocessing step, the data is prepared for entering into the proposed deep network. Then, a hybrid deep convolutional neural network (HDCNN) is proposed which is a combination of the U-net architecture and residual networks. In this network, effective features are generate4d in different scales during the downsampling process from the input DSM and the DTM is extracted during the upsampling process. To rectify the results, a multi-scale fusion strategy is proposed to produce the final DTM by fusing the generated DTMs at different scales and with different spatial shifts. The performance of the proposed approach is analyzed by implementing four different evaluation scenarios in five different datasets. The evaluation results demonstrated significant performance and high generalizability of the proposed approach. The proposed network also outperforms the deep learning-based filtering methods and two reference DTM extraction algorithms especially in challenging regions.