Predicting the post-wildfire mudflow onset using machine learning models on multi-parameter experimental data

Post-wildfire mudflows are increasingly hazardous due to the prevalence of wildfires, including those on the wildland-urban interface. Upon burning, soil on the surface or immediately beneath becomes hydrophobic, a phenomenon that occurs predominantly on sand-based hillslopes. Rainwater and eroded soil blanket the downslope, leading to catastrophic debris flows. Soil hydrophobicity enhances erosion, resulting in post-wildfire debris flows that differ from natural mudflows in intensity, duration, and destructiveness. Thus, it is crucial to understand the timing and conditions of debris-flow onset, driven by the coupled effects of critical parameters: varying rain intensities (RI), slope gradients, water-entry values, and grain sizes (D50). Machine Learning (ML) techniques have become increasingly valuable in geotechnical engineering due to their ability to model complex systems without predefined assumptions. This study applies multiple ML algorithms: multiple linear regression (MLR), logistic regression (LR), support vector classifier (SVC), K-means clustering, and principal component analysis (PCA) to predict and classify outcomes from laboratory experiments that model field conditions using a rain device on various soils in sloped flumes. While MLR effectively predicted total discharge, erosion predictions were less accurate, especially for coarse sand. LR and SVC achieved good accuracy in classifying failure outcomes, supported by clustering and dimensionality reduction. Sensitivity analysis revealed that fine sand is highly susceptible to erosion, particularly under low-intensity, long-duration rainfall. Results also show that the first 10 minutes of high-intensity rain are most critical for discharge and failure. These findings highlight the potential of ML for post-wildfire hazard assessment and emergency response planning.

💡 Research Summary

The paper investigates post‑wildfire debris‑flow initiation by experimentally reproducing key hydrological and geomorphological conditions in laboratory flumes and applying a suite of machine‑learning (ML) techniques to the resulting dataset. Thirty‑six controlled tests were performed with two hydrophobic‑layer configurations (H‑Top, where the water‑repellent layer is exposed at the surface, and H‑Sub, where it is buried beneath a thin hydrophilic layer). Four input variables were systematically varied: median grain size (D₅₀), water‑entry value (ψ), slope angle (δ), and rain intensity (RI).

For the H‑Top configuration, multiple linear regression (MLR) was used to predict total water discharge (TD) and total erosion (TE). Feature selection, guided by a correlation matrix, removed contact angle and friction angle, leaving D₅₀, ψ, δ, and RI. The MLR model achieved high predictive performance for TD (R² = 0.96 training, 0.72 testing, MSE = 14.5 → 75.3) but performed poorly for TE (R² = 0.87 → 0.45, MSE = 655 911 → 1 070 667), indicating over‑fitting, especially for coarse sand where linear assumptions break down.

For the H‑Sub configuration, the goal shifted to binary classification of “infinite failure” versus “no failure.” First, K‑means clustering identified two natural groups, and principal component analysis (PCA) revealed that the first two components were dominated by RI and ψ, followed by D₅₀ and δ. Both logistic regression (LR) and support‑vector classifier (SVC) then achieved >90 % accuracy in separating the two classes, with decision boundaries clustering in the high‑RI, high‑ψ region.

Sensitivity analysis highlighted the interaction between grain size and rain intensity. Fine sand (D₅₀ = 0.2 mm) is especially vulnerable under low‑intensity, long‑duration rainfall, whereas coarse sand shows relative resilience even at higher intensities. Crucially, the first 10 minutes of high‑intensity rain (RI ≥ 70 mm h⁻¹) dominate both discharge volume and failure likelihood, underscoring the importance of early‑stage monitoring in real‑world emergency response.

The study’s contributions are threefold: (1) generation of a high‑quality, controlled experimental dataset that isolates key variables often confounded in field studies; (2) demonstration that a combination of regression, classification, clustering, and dimensionality‑reduction techniques can jointly elucidate the physical drivers of post‑fire debris flows; and (3) provision of actionable insight that the initial rain period is the critical window for hazard mitigation.

Limitations include the modest sample size, which restricts model generalizability, and the reliance on linear regression for erosion prediction, which cannot capture complex non‑linear interactions. Future work should expand the experimental matrix, incorporate non‑linear algorithms such as random forests, gradient boosting, or deep neural networks, and integrate field observations to develop hybrid models capable of real‑time risk forecasting for post‑wildfire landscapes.