Geospatial Deep Learning for Pomelo Tree Detection in Mixed Orchard Systems: A UAV-Based Assessment of YOLO Models

Abstract

Identification of reactivation zones in thick volcanic soils is critical for understanding landslide mechanisms. Volcanic soils are clay-rich, making slopes highly susceptible to water saturation and renewed movement. However, scientific understanding of the relationship between surface indicators and subsurface structure remains limited, particularly in thick and well-developed volcanic terrains. Few studies have combined vegetation-based deformation with subsurface resistivity structure within a single framework. This study integrates unmanned aerial vehicle (UAV)-based LiDAR and electrical resistivity tomography (ERT) to characterize an active landslide in Kalisari, Central Java, Indonesia. ERT serves as a non-invasive method to map subsurface moisture and weak zones. UAV-LiDAR data were processed to generate a detailed terrain representation of the landslide. The UAV-LiDAR point cloud was used to extract surface morphology and vegetation-based deformation metrics, including tree stem-lean angle and azimuth. ERT surveys were conducted along multiple profiles across the landslide to delineate low-resistivity zones interpreted as saturated clay-rich materials. Results indicate that reactivation areas are clearly expressed by surface morphology and coherent stem-lean patterns. Although the overall morphology is rotational, recent movement is dominated by coherent lateral displacement. Stem-lean orientations align with the inferred movement direction (azimuth 45–100°). LiDAR-derived stem-lean estimates show good agreement with field measurements (R² = 0.87 for lean angle and ρ² = 0.85 for azimuth). ERT results also show that reactivation areas correspond to low-resistivity zones (3–30 Ω·m) that spatially coincide with surface cracks, flow-accumulation features, and stem-lean anomalies. This indicates that water infiltration through surface discontinuities recharges the weak saturated clay layer and makes the slope more prone to instability. Overall, integrating UAV-LiDAR and ERT provides a strong basis for linking surface deformation with subsurface structure and identifying landslide reactivation zones in volcanic landscapes.