Construction of a small-scale relief shading neural network model based on the attention mechanism
Wenping Jiang, Yue Wang, Haijun Ding, Han Jiang, Daping Xi, Yuan Wang, Peiyang Ma · Cartography and Geographic Information Science · 2025
Relief shading is a primary technique for representing the three-dimensional effects of terrain on a two-dimensional plane. This study applies deep learning to generate small-scale Swiss-style relief shading maps. An attention module is defined to focus on key information in feature maps. Based on the characteristics of relief shading and digital elevation model (DEM) data, U-Net is adjusted and optimized, resulting in the design and construction of an end-to-end relief shading neural network model (Attention Hillshading U-Net, A-UNet) built on a limited training dataset. By learning the terrain-shaping patterns from Swiss-style shading maps, the model overcomes the challenges posed by high terrain complexity and insufficient representation of landform morphology in small-scale relief shading maps. The study further investigates the impact of hyperparameters on the performance of the model in generating small-scale relief shading maps. Based on the quantitative performance of the model under different hyperparameter settings and adaptability to lower-resolution DEMs, the optimal hyperparameters for the model are determined. Additionally, experimental comparisons of small-scale relief shading map generation using A-UNet and other network models show that, compared to U-Net and its variants, A-UNet demonstrates superior adaptability to different pixel sizes, better terrain simplification, and enhanced generalization to various landform types.