Ressub-Net: Residual Subtraction Network For DTM Extraction From DSM

Digital terrain model (DTM) is of paramount importance in various applications, such as infrastructure planning. However, this model is not directly generated from remote sensing sensors. DTMs are generally generated by filtering digital surface models (DSMs) as a product of these sensors. In this regard, deep learning (DL) techniques have been successfully used compared to traditional ones in recent years. DL-based DTM extraction methods, the performance of which is dependent on training losses, are usually followed by ad-hoc post-processors, which makes the efficiency of the DL part distorted. Accordingly, this article presents an independent novel network, a residual subtraction network (ResSub-NET), where different loss functions and their combinations are also assessed. The proposed method is a simple version of residual-based networks, the skip connections of which play a subtraction role. This is attributed to the nature of the problem, which is filtering. The experimental results in three different datasets, related to urban, hilly, and mountainous show that the proposed method outperforms common traditional methods as well as the current state-of-the-art one, named DeepTerRa.