Predicting the thickness of alpine meadow soil on headwater hillslopes of the Qinghai-Tibet Plateau

Alpine meadow soils in tectonically active High Mountain Asia (HMA) are highly vulnerable, playing a critical role in the fragile ecosystems of Earth’s “Third Pole”. Understanding the formation and distribution of these soils is essential, yet the mechanisms governing their thickness remain unclear. To address this, we applied a multi-resolution (0.25–30.00 m) stochastic approach to predict soil thickness in a nested headwater catchment in the Lhasa River Basin, namely D5K (415 sites) and its headwater sub-catchment D5KH1 (330 sites). Using a linear mixed-effects model, we quantified uncertainty through within-site and between-site variance analysis, and compared the performance of deep learning (DL) with three traditional machine learning methods—random forest (RF), support vector machine (SVM), and artificial neural networks (ANN)—under three variable selection strategies: (1) all variables, (2) stepwise selection, and (3) a novel Boruta-based recursive method. Field investigations revealed that organic matter accumulation and freeze–thaw cycles are dominant pedogenic factors, producing a root-dense turf layer beneath a dark humus horizon. Freeze-thaw dynamics also contribute to geomorphological features such as landslides and stone stripes. Soil thickness varied significantly across topographic positions—valleys, sideslopes, and ridges (p < 0.001)—with within-site variance (SD = 11.06 cm) slightly exceeding between-site variance (SD = 9.58 cm). Fine-resolution digital elevation models (DEMs) effectively captured these variations, largely because they incorporate critical topographic factors such as the Topographic Position Index (TPI), which Boruta analysis identified as more influential on soil thickness than flow-related indices like the Topographic Wetness Index. Among the tested models, RF combined with the Boruta method and a 0.25 m resolution DEM provided the most accurate predictions, outperforming DL. These findings emphasize that the complexity of DL may not translate into superior performance for all applications, particularly in small dataset terrain-driven ecological studies.