Adaptive digital elevation models construction method based on nonparametric regression

Interpolation is one of the most critical factors affecting the quality of digital elevation models (DEM) generated from sampling point data. However, existing methods have not taken into full account that the interpolation function should be consistent with the density and direction characteristics of sampling points. Therefore, an adaptive DEM construction method based on nonparametric regression (AdpNPR-DEM) is proposed, which incorporates an adaptive kernel that changes according to local landform types. The method first estimates local terrain features, and second the dominant orientation of the local gradients. The orientation information adaptively "steers" the local kernel to produce elongated, elliptical contours that spread along the directions of local terrain features. The effectiveness of the AdpNPR-DEM is verified based on directional features, the number of sampling points, and typical landform types. The results show that, compared with triangulated irregular network (TIN), natural neighbor interpolation (NNI), and the Australian National University digital elevation model (ANUDEM) interpolation method, AdpNPR-DEM has the highest accuracy for all landform types and a better level of robustness. Our method improves the quality of DEM in areas with complex landforms. It could significantly promote the high-quality production of DEMs, and also promisingly broaden and deepen its applications.