Explainable Ensemble Learning for Stormwater-Drainage Failure Prediction under Nonstationary Extreme Precipitation: Integrating LiDAR, Land-Cover, Soil, and Infrastructure Data

Abstract

This study addresses the growing problem of stormwater-drainage failure under nonstationary extreme precipitation, where conventional rainfall-based or single-factor assessments may not adequately represent the combined effects of terrain, land cover, soil hydrology, infrastructure condition, and changing rainfall extremes. The purpose of the research is to develop and empirically evaluate an explainable ensemble-learning framework for predicting Stormwater-Drainage Failure Susceptibility (SDFS) by integrating LiDAR-derived terrain and surface characteristics (LDTSC), Land-Cover and Imperviousness Conditions (LCIC), Soil Infiltration and Hydrologic Characteristics (SIHC), Stormwater-Drainage Infrastructure Condition (SDIC), and extreme-precipitation variables. A quantitative, cross-sectional, explanatory-predictive case-study design was employed within a selected urban drainage catchment using purposive sampling, structured expert questionnaires, and multimodal geospatial and infrastructure data. Of 320 questionnaires distributed, 306 were returned and 294 were valid, while 48 drainage subcatchments formed the analytical spatial cases. Data analysis combined descriptive statistics, reliability and validity testing, Pearson correlation, multiple regression, Random Forest, XGBoost, LightGBM, ensemble learning, cross-validation, and SHAP-based explainability. SDFS recorded M = 4.21, SD = 0.52, while SDIC showed the strongest correlation with SDFS, r = .76, p < .001, followed by LCIC, r = .72, LDTSC, r = .68, and SIHC, r = .65, all p < .001. The regression model explained 70.3% of SDFS variance, R² = .703, adjusted R² = .699, F (4,289) = 171.42, p < .001, with SDIC the strongest predictor, β = .33. The final ensemble achieved 93% accuracy, .92 precision, .94 recall, .93 F1-score, and ROC-AUC = .97, outperforming the best individual model. SHAP analysis identified reduced drainage capacity, flow accumulation, impervious surface proportion, depression depth, and extreme rainfall intensity as the leading contributors to predicted failure. The findings imply that integrated, interpretable multimodal prediction can strengthen drainage-risk prioritization, infrastructure maintenance, urban flood preparedness, and climate-resilient stormwater planning for municipalities facing increasingly uncertain and spatially heterogeneous rainfall conditions.