LIE-DSM: Leveraging Single Remote Sensing Imagery to Enhance Digital Surface Model Resolution

Digital surface models (DSMs) have numerous valuable applications in infrastructure and industrial development. However, the spatial resolutions of DSMs are often limited due to data acquisition constraints, resulting in potential inaccuracies in these applications. Recently, deep learning-based super-resolution algorithms have been utilized to enhance the accuracy of DSMs. Despite their success, these algorithms still possess limited accuracy and robustness. As such, a new DSM super-resolution algorithm named LIE-DSM is proposed, which leverages single remote sensing imagery (RSI) to enhance the accuracy of low-resolution (LR) DSM. Specifically, we introduce a dual-input neural network tailored to the characteristics of RSIs and DSMs to generate a high-resolution (HR) output. Experiments show that LIE-DSM attains outstanding performance in all metrics and the improvements exceeds 15% in high-ratio upsampling. Moreover, the visualized results showcase more accurate shapes, crisper edges, and a distribution closely resembling the ground truth.