Artefact and risk area detection of the tanDEM-X 30m edited DEM using an integrated machine learning and dem-of-difference framework: a case study in the Alps

The TanDEM-X 30m Edited Digital Elevation Model (TDX30) is widely used but impaired by complex artefacts from InSAR processing. We propose a multi-strategy detection framework that classifies TDX30 pixels into four categories (Normal Terrain, Minor Artefacts, Severe Artefacts, and Risk Areas) by fusing CatBoost terrain classification, directional edge detection for seamline artefacts, and dual-reference DEM of Difference (DoD) analysis against AW3D30 and SRTM V3. Applied to the entire European Alps, the CatBoost classifier achieves a test accuracy of 97.43% (±0.49%) and a Kappa coefficient of 93.70% (±1.10%). Independent ICESat-2 validation confirms strong inter-class separation: normal terrain yields a mean absolute error (MAE) of 5.34 m, while minor artefacts, severe artefacts, and risk areas reach 54.29 m, 294.96 m, and 415.75 m, respectively. Risk areas, which appear visually normal but exhibit catastrophic absolute elevation errors (RMSE=707.86 m, LE90=1,079.79 m), are detectable only through DoD, demonstrating the necessity of multi-strategy fusion. Crucially, the product provides data producers with a spatially explicit diagnostic layer for locating distortion centres that require review, flagging, or replacement, while allowing end-users to mask or down-weight high-risk pixels before downstream analyses, thereby reducing severe error propagation in DEM-based applications.