Machine Learning Maps Future Debris Flow Risks in Qinghai

Published in Geoenvironmental Disasters, the study integrates high-resolution precipitation projections from the Coupled Model Intercomparison Project Phase 6 (CMIP6) with environmental variables to predict spatial and temporal changes in hazard potential. The primary objective was to identify dominant conditioning factors and locate areas most sensitive to climate-driven disturbances, providing a technical basis for risk prevention in a region where debris flows frequently threaten local settlements and infrastructure.

The study focuses on Qinghai Province, situated in the northeastern sector of the Qinghai–Tibet Plateau. This area is characterized by complex disaster-forming conditions, with more than four-fifths of its land area consisting of plateaus and mountainous terrain above 3,000 meters. Historically, the province has experienced frequent debris flows, with 120 incidents recorded between 1990 and 2019, resulting in severe casualties and over 400 million in direct economic damages. The research team utilized a comprehensive inventory of 1,433 documented debris flow points to train their predictive models, ensuring the dataset represented events accumulated up to 2019.

To model susceptibility, the researchers selected thirteen environmental and climatic factors, including topographic parameters such as elevation and slope, hydrological conditions like river distance, and climatic variables such as annual and seasonal rainfall. After conducting correlation and Variance Inflation Factor analyses to address multicollinearity, the final predictor set retained twelve variables, including Digital Elevation Model (DEM), slope, road distance, and land use. These factors were processed into a uniform grid with a 1-kilometer spatial resolution to ensure consistency across the study area.

The team compared four machine learning classifiers: Random Forest, XGBoost, Support Vector Machine, and Logistic Regression. To avoid overestimating model performance due to spatial autocorrelation, they employed seven-fold spatial cross-validation. XGBoost demonstrated the strongest predictive performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.950. Statistical tests confirmed that XGBoost significantly outperformed the other models, leading the researchers to select it as the final model for generating susceptibility maps.

Interpretation of the XGBoost model using Shapley Additive Explanations (SHAP) revealed that road distance and elevation were the dominant predictors of debris flow susceptibility. Lower values of road distance and lower elevations were associated with higher predicted susceptibility. However, SHAP interaction analysis indicated that precipitation variables, while not the strongest individual predictors, exerted nonlinear amplification effects when interacting with topographic and human-disturbance factors. This suggests that the impact of rainfall on debris flow risk is significantly heightened in areas with favorable predisposing conditions, such as low-elevation valleys and road corridors.

The baseline susceptibility map showed a distinct east–west gradient, with high and very high susceptibility zones concentrated in the eastern and central parts of Qinghai. These zones are primarily located along river valleys and road corridors, forming elongated belt-like clusters. In contrast, the western and northern regions, which are dominated by high-elevation plateaus with sparse human activity, exhibited much lower susceptibility. This spatial pattern aligns with historical debris flow records, demonstrating the model’s ability to capture the main characteristics of hazard distribution in the region.

Future projections under three Shared Socioeconomic Pathways (SSPs) indicated that susceptibility changes are scenario-dependent. The combined proportion of high and very high susceptibility zones was projected to reach 8.5% of the study area under the high-emission SSP5-8.5 scenario in 2060. This represents an increase from the baseline period, where these zones accounted for 8.1% of the area. The largest high-risk area was projected under this high-emission scenario, with high and very high susceptibility zones covering 19,833 square kilometers and 39,723 square kilometers, respectively.

A transition matrix analysis quantified the spatial shifts between susceptibility classes from the baseline to the 2060 SSP5-8.5 scenario. The results showed that susceptibility escalation mainly occurred in areas already classified as moderate or high under baseline conditions. For instance, 4,664 square kilometers of baseline moderate susceptibility areas shifted to the high class, and 1,840 square kilometers shifted to the very high class. This indicates that intensified precipitation under high-emission scenarios may further amplify debris flow susceptibility in areas with complex terrain and existing vulnerability.

The study also conducted an exposure-weighted potential loss risk estimation to evaluate the possible impacts on population and buildings. By overlaying susceptibility maps with population density and building footprint data, the researchers found substantial overlap between future high-susceptibility zones and densely populated areas. This spatial overlap significantly amplifies the potential impacts of debris flows, suggesting that areas with concentrated human activities face compounded risks under future climate conditions, regardless of the specific emission pathway.

Despite the robust performance of the model, the authors acknowledged several limitations. The study did not explicitly incorporate dynamic factors such as land use changes and wind speed, which may lead to an underestimation of hazard potential. Additionally, the future susceptibility projections should be interpreted as scenario-based estimates rather than deterministic predictions, given the inter-model variability among the three CMIP6 global climate models used. The potential loss risk estimation also relied on static population and infrastructure data, limiting its ability to capture evolving socioeconomic risk patterns.

The findings suggest that climate change may intensify debris flow risk in Qinghai Province, particularly in transitional zones between moderate and high susceptibility. Integrating machine learning with climate projections provides a reliable technical basis for identifying climate-sensitive areas. The researchers emphasize that incorporating climate-adaptive susceptibility information into land use planning is essential for reducing future debris flow risks. Prioritizing monitoring and early-warning systems in road corridors, river valleys, and settlement-intensive areas could help protect vulnerable communities and infrastructure from the increasing threat of debris flows.

Geographically, Qinghai extends approximately 1,240.6 kilometers from west to east and 844.5 kilometers from north to south, covering roughly 722,300 square kilometers. The terrain is shaped by two principal mountain systems oriented east–west and north–south, with lower mountains in the east contrasting with broad plateaus and basins in the west. Qinghai is influenced by three climatic systems: continental monsoon, inland arid, and plateau alpine regimes. This produces an arid environment marked by hypoxia, intense solar radiation, low air temperatures, strong winds, and annual rainfall generally below 400 millimeters that decreases from southeast to northwest.

To build a balanced sample set for susceptibility modeling, the authors generated 1,464 non-debris-flow locations to match the 1,433 documented disaster sites. The sampling procedure applied strict spatial separation to prevent clustering: a minimum spacing of 10,000 meters was enforced between non-hazard points, and any points within 13,000 meters of documented debris flow occurrences were excluded. All spatial raster layers were projected to the WGS 1984 UTM Zone 47 N coordinate system. Historical climatic records incorporated thirty-year averages of annual rainfall, seasonal rainfall during July and August, and July rainfall alone, which was isolated because preliminary analyses identified July as Qinghai’s wettest month. Exposure data incorporated a 100-meter gridded population layer derived from China’s seventh national census alongside a vector building footprint dataset.

Future precipitation data came from a 1-kilometer monthly dataset for China covering 2021 to 2100, which applied delta-based statistical downscaling to combine coarse global climate models with high-resolution baseline observations. To reduce uncertainty, the researchers calculated a pixel-by-pixel ensemble mean across three CMIP6 models: EC-Earth3, GFDL-ESM4, and MRI-ESM2-0 under the SSP1-1.9, SSP2-4.5, and SSP5-8.5 emission pathways. Mean annual ensemble precipitation ranged from 334.85 millimeters under SSP5-8.5 in 2060 to 366.44 millimeters under SSP1-1.9 in 2060. Coefficients of variation across the three climate models ranged between 10.98% and 18.96%, with the highest inter-model divergence occurring under SSP1-1.9 in 2030 and the lowest under SSP1-1.9 in 2060.

Model hyperparameters were optimized using Bayesian optimization for XGBoost and grid search for random forest, support vector machines, and logistic regression through five-fold cross-validation. Under seven-fold spatial cross-validation, XGBoost achieved a mean accuracy of 0.895 and a Kappa of 0.707. A Friedman test confirmed statistically significant performance differences among the four algorithms, with a test statistic of 18.943 and a p-value of 0.000281. Pairwise Wilcoxon signed-rank tests demonstrated that XGBoost significantly outperformed logistic regression, random forest, and support vector machines, with p-values of 0.0156 across all comparisons. In the SHAP interaction analysis, the strongest interaction emerged between elevation and road distance, followed by elevation and river distance. Annual rainfall and elevation formed the third strongest pair with a mean absolute interaction value of 0.0606, demonstrating that precipitation effects are modulated by topography.

Baseline evaluations showed that the very low susceptibility category covered 589,117 square kilometers, or 84.3% of the province, while low and moderate zones accounted for 4.2% and 3.4%, respectively. For risk estimation, the team assigned relative weights of 0.9, 0.7, 0.5, 0.3, and 0.1 to very high, high, moderate, low, and very low zones before multiplying them by population and building densities. At the local scale in eastern high-susceptibility zones, the 2060 SSP1-1.9 scenario exhibited the largest area of high building loss risk, surpassing the SSP2-4.5 and SSP5-8.5 pathways. Under the 2060 SSP5-8.5 projection, 576,748 square kilometers remained very low, while 32,552 square kilometers persisted in the very high category. To refine future assessments, the authors suggested coupling machine learning with physically based process models and incorporating dynamic socioeconomic projections.

Subject of Research: Machine learning-based assessment of future debris flow susceptibility under CMIP6 precipitation scenarios in Qinghai Province

Article Title: Machine learning-based assessment of future debris flow susceptibility under CMIP6 precipitation scenarios in Qinghai Province, China

Article References: Lü, Q., Yang, K., Zhang, W., Wang, L., & Wang, L. (2026). Machine learning-based assessment of future debris flow susceptibility under CMIP6 precipitation scenarios in Qinghai Province, China. Geoenvironmental Disasters, 13(1), Article 40. https://doi.org/10.1186/s40677-026-00393-5

Image Credits: AI Generated

DOI: 10.1186/s40677-026-00393-5

Keywords: debris flow, machine learning, CMIP6, climate change, Qinghai Province, susceptibility mapping, XGBoost, geohazard, risk assessment, precipitation projections, Machine, learning-based