Abstract
<title>Abstract</title> <p>Digital terrain models (DTMs) from unmanned aerial vehicle (UAV) photogrammetry are limited by the effectiveness of ground filtering. Conventional slope-based filtering (SBF) that utilizes a single global slope threshold is insufficient in areas with dense vegetation and closely spaced buildings. This study introduces a dual-scheme adaptive SBF that segments a photogrammetric digital surface model (DSM) into two distinct domains prior to thresholding. Scheme A, designed for variable terrain and vegetation, employs a roughness raster based on the local standard deviation of elevation. Scheme B, intended for flat, building-dominated terrain, utilizes an edge-density raster generated via Canny edge detection. Voids are addressed using a hybrid approach that combines inverse-distance-weighted and nearest-neighbor interpolation with Gaussian smoothing. The method was evaluated using an oblique UAV survey (4.15 cm ground sampling distance, 145 points per m2) covering 121 hectares of the ITERA campus in Lampung, Indonesia, an area characterized by dense secondary vegetation, buildings, and 70 meters of relief. Comparison with twelve GNSS-surveyed independent check points yielded a vertical root mean square error (RMSE) of 0.148 m and a linear error at 90% confidence (LE90) of 0.244 m. The observed 0.004 m increase over the DSM remains below the validation noise, indicating no measurable loss of terrain fidelity. However, visual inspection revealed over-smoothed surfaces in extensive vegetated regions, suggesting that void extent, rather than classification error, constitutes the primary limitation.</p>