Advancing Landslide Understanding in Forested Areas: High-Resolution Digital Twins for Novel DTM Extraction and Digital Dendrogeomorphology

Landslides are among the most frequent natural hazards, posing threats to human lives, infrastructure, and the environment. Effective mitigation of the associated risks requires improved landslide understanding based on high-resolution data.               
Digital twins, created using optical or LiDAR-based remote sensing techniques alongside the derived digital terrain models (DTMs), offer significant potential to enhance landslide modeling and therefore deepen our understanding of their dynamics. However, the generation of such 3D models remains challenging in forested areas as most remote sensing techniques, particularly those utilizing aerial platforms, lack the ground point density needed to capture the surface precisely. Additionally, ground point classification (GPC) becomes more difficult at increasing resolution and forest cover.

To address these challenges, we explore novel under-canopy ground data acquisition methods for digital twin generation and present innovative approaches for DTM generation and refinement, along with digital dendrogeomorphology.         
Conducted at a representative study site, close-range panoramic terrestrial photogrammetry (_crpTP) and terrestrial laser scanning (_crpTLS) were employed and evaluated for their effectiveness in generating high-resolution 3D models. We utilized a custom R-tool applied to the derived point cloud for ground surface extraction and refinement. The presented script overcomes challenges in DTM generation from high-resolution point clouds by correcting misclassifications typical in the GPC process at fine scales, using a combination of tree detection and generalized additive modeling. Additionally, we analyse the morphology of detected trees by automatically fitting least square ellipses to stem segments and deriving the shape, eccentricity, inclination, and tilting direction of tree stems. This data-driven, digital approach to dendrogeomorphology offers potential solutions to current challenges in conventional dendrogeomorphological surveys, such as time and labour intensiveness or bias in the selection of sample and reference trees.

Our study demonstrates that both _crpTP and _crpTLS are capable of producing highly accurate digital twins of forested landslides. The models generated through photogrammetric data acquisition for two small-scale test plots are characterized by low check point RMSE values ranging from 0.16 to 0.19 cm, indicating high model accuracy. In comparison, the _crpTLS digital twin of the entire study area showed an RMSE value of 1.87 cm. The accuracy of the derived DTMs, validated using independent sets of ground reference points, ranged from 0.92 to 1.90 cm. The high filtering accuracy demonstrated the capability of the presented approach to reduce misclassification and propagation errors in the generated DTMs.      
By employing digital dendrogeomorphology, we were able to fit least square ellipses to stem segments at RMSE accuracies close to the voxel size of the point cloud. Based on these accurately fitted ellipses, their respective centroids and shift across stem segments, we demonstrate the feasibility of automatically extracting tree morphology indicators.

Our study shows that the combined extraction of DTMs and tree morphology indicators from digital twins can provide valuable data on both surface and subsurface processes, which, when applied over multiple time periods, can enhance landslide understanding. However, as the study site is characterized by little understory, the applicability at other sites with different forest structures has yet to be explored.