A geomorphic process-informed machine learning strategy for interpretable soil thickness mapping in the subtropical hilly regions of Southern China

Abstract Soil thickness exhibits strong spatial heterogeneity in hilly regions, and its predictive mapping is hindered by the complexity of geomorphic processes and limited model interpretability. We develop an interpretable, process-informed framework integrating variable construction, multi-algorithm comparison, uncertainty quantification and slope-zone interpretation, and apply it in Xunwu County, subtropical China. Using 513 profiles and 35 variables spanning terrain, climate, erosion, vegetation, spectral and composite information, we compared five machine learning models—Random Forest, quantile regression forest (QRF), XGBoost, CatBoost and LightGBM—and selected the final model on held-out test data. Process information entered through process-informed features (three terrain core variables and six interaction composites) and slope-stratified interpretation and uncertainty evaluation. The composites supported interpretation, but removing them changed the mean R² from 0.418 to 0.443 (paired 95% CI − 0.015 to + 0.060), so their predictive benefit remains unproven. CatBoost achieved the highest mean R² under repeated random splits (0.418 ± 0.056; RMSE 22.94 cm; MAE 16.65 cm), followed by XGBoost (0.416 ± 0.065); the difference was not significant. This accuracy reflects local interpolation within the sampled domain: under spatial blocking and leave-quadrant-out validation, R² was − 0.84 to − 0.91 and − 0.03, respectively. The 90% QRF intervals achieved 85.4% coverage, increasing to 87.7% after conformal calibration, with a mean width of 70.4 cm. The applicability domain covered 92.9% of the area. SHAP showed that climate and topographic position dominated the gentle and transitional zones, whereas slope, spectral variables and vegetation were more prominent in the steep zone. These results support local, slope-stratified soil thickness mapping and interpretation rather than prediction in spatially independent areas.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-17
DOI
https://doi.org/10.1038/s41598-026-80775-9
Primary Topic
Soil Geostatistics and Mapping
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A geomorphic process-informed machine learning strategy for interpretable soil thickness mapping in the subtropical hilly regions of Southern China

Guo Xi, Ji Huang, Hongyu Wang, Jiayue Wan et al.
2 links
Scientific Reports
Soil Geostatistics and Mapping
article

A geomorphic process-informed machine learning strategy for interpretable soil thickness mapping in the subtropical hilly regions of Southern China

Guo Xi, Ji Huang, Hongyu Wang, Jiayue Wan, Meiying Zheng
article en

Abstract

Abstract Soil thickness exhibits strong spatial heterogeneity in hilly regions, and its predictive mapping is hindered by the complexity of geomorphic processes and limited model interpretability. We develop an interpretable, process-informed framework integrating variable construction, multi-algorithm comparison, uncertainty quantification and slope-zone interpretation, and apply it in Xunwu County, subtropical China. Using 513 profiles and 35 variables spanning terrain, climate, erosion, vegetation, spectral and composite information, we compared five machine learning models—Random Forest, quantile regression forest (QRF), XGBoost, CatBoost and LightGBM—and selected the final model on held-out test data. Process information entered through process-informed features (three terrain core variables and six interaction composites) and slope-stratified interpretation and uncertainty evaluation. The composites supported interpretation, but removing them changed the mean R² from 0.418 to 0.443 (paired 95% CI − 0.015 to + 0.060), so their predictive benefit remains unproven. CatBoost achieved the highest mean R² under repeated random splits (0.418 ± 0.056; RMSE 22.94 cm; MAE 16.65 cm), followed by XGBoost (0.416 ± 0.065); the difference was not significant. This accuracy reflects local interpolation within the sampled domain: under spatial blocking and leave-quadrant-out validation, R² was − 0.84 to − 0.91 and − 0.03, respectively. The 90% QRF intervals achieved 85.4% coverage, increasing to 87.7% after conformal calibration, with a mean width of 70.4 cm. The applicability domain covered 92.9% of the area. SHAP showed that climate and topographic position dominated the gentle and transitional zones, whereas slope, spectral variables and vegetation were more prominent in the steep zone. These results support local, slope-stratified soil thickness mapping and interpretation rather than prediction in spatially independent areas.

Scientific Reports
Ministry of Agriculture and Rural Affairs (CN), Jiangxi Agricultural University (CN)
National Key Research and Development Program of China
Life on land
Soil Geostatistics and Mapping
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon