A Feature-Fusion Convolutional Neural Network for Terrain Generalization

Terrain generalization—the scale-driven abstraction of digital elevation models (DEMs) that retains dominant landforms while filtering trivial microtopography—has long been dominated by rule-based algorithms, whereas deep learning has mainly targeted the inverse task of DEM super-resolution. This study proposes TG-CNN, a feature-fusion convolutional neural network for DEM down-scaling generalization. The network compresses the high-resolution DEM through a max- or average-pooling front end, combines normalized elevation with slope and sine–cosine circularly encoded aspect as complementary input channels, and is trained with a composite elevation–slope–aspect loss. Training, validation, and testing are conducted on geographically independent mountainous areas under a leave-one-area-out scheme, followed by adjustable IDW fusion post-processing. On the test region, classical bilinear interpolation retains the best elevation accuracy, while the standalone TG-CNN does not reach a comparable level; fused accuracy improves monotonically as the IDW share increases, indicating that the fusion benefit originates from the interpolation component. On the ridge–valley structure, however, TG-CNN behaves as an aggressive yet reliable filter: it achieves the highest precision and the smallest positional offset among all compared methods but retains only part of the reference skeleton, leaving its F1-score below the interpolation baseline. TG-CNN thus realizes a genuine, controllable terrain-abstraction effect, yet deep learning does not yet surpass mature interpolation in pixel-level elevation accuracy—a negative result that clarifies current limitations and points toward residual learning and information-preserving architectures.