Abstract

In the conservation of earthen sites in humid environments, damage induced by vegetation root systems constitutes one of the primary threats. However, quantitative assessment of root system architecture remains severely constrained by the challenges of belowground sampling. This study proposed a root assessment framework that integrated multi-resolution segmentation with partial least squares regression (PLS), specifically designed for earthen sites where belowground root samples are scarce. Taking the earthen remains of the ancient city wall in Shugang, Yangzhou as a case study, we developed predictive models for root density across shallow-, medium-, and overall layers, as well as maximum rooting depth. These models integrated GPR-derived measurements from 28 non-standard sampling plots and regional-scale ALS point cloud data. Results indicated that multi-resolution segmentation units derived from canopy structure significantly outperformed traditional fishnet-based methods, effectively enhancing model performance and stability under small-sample conditions. The optimal validation R² values reached 0.861 for maximum rooting depth and 0.804 for overall root density. Component loadings analysis further revealed that spectral features dominated shallow-layer root prediction, whereas elevation-derived structural metrics exhibited strong explanatory power for maximum rooting depth. This framework not only overcomes the constraints of sample scarcity and non-standard sampling plots commonly encountered in archaeological contexts, but also provides an applicable technical pathway for regional-scale mapping and preventive management of vegetation root-related risks at earthen sites.