TFD-CGAN

Shaded relief maps are a crucial means of terrain visualization. Existing neural network-based generation methods are predominantly data-driven and typically rely solely on digital elevation models (DEMs) as input. Such approaches often neglect structural semantics of landform and key terrain feature constraints, leading to discontinuous ridge representation, detail distortion, and limited generalization across different landform types. To address these limitations, this study constructs a terrain-feature-driven conditional generative adversarial network (TFD-CGAN). Using pix2pixHD as the baseline, topographic feature lines or other terrain factors are explicitly incorporated into the network. By optimizing the generator architecture, feature fusion strategies, and loss function design, the model improves its ability to capture landform structural characteristics while enhancing the controllability and stability of the generated results. Experiments conducted across multiple regions with diverse landform types are evaluated using both quantitative metrics and qualitative visual analysis. The results indicate that introducing topographic feature lines significantly improves ridge-line continuity and overall structural clarity. Among different terrain factors, aspect exhibits the most stable performance across landform types, curvature favors loess landforms, and slope demonstrates limited generalization in cross-landform applications. All in all, the proposed method outperforms DEM-only baselines and provides an effective approach for intelligent relief shading generation.