Ensemble and symbolic regression-based stability assessment of debris-flow barrier dams considering data scarcity

Background Debris-flow barrier dams are less studied than landslide dams but pose disproportionately high risks due to sudden breaching and severe downstream impacts. The scarcity of multi-dimensional observations, including morphological and hydrologic parameters, limits the performance and generalizability of existing stability assessment models, which are largely adapted from landslide dam studies and fail to capture the specific characteristics of debris-flow barrier dams. This data limitation also constrains feature extraction in conventional data-driven and machine learning approaches and increases the risk of overfitting. This study aims to develop an interpretable, data-driven framework for stability evaluation under data-scarce conditions. Results Using a framework that integrates physical model ensembles with symbolic regression, where inter-model complementarity is quantified through correlation and inconsistency analyses to guide feature construction, the proposed approach achieves an accuracy of 97.5% on the full dataset, representing an improvement of approximately 17.5% over conventional models. It effectively reduces false alarms while maintaining reliable classification performance. The model further identifies dam volume and lake volume as dominant controlling factors and determines critical threshold intervals for instability prediction. Conclusions This study proposes an integrated ensemble and symbolic regression framework that enhances predictive robustness and interpretability for debris-flow barrier dam stability assessment. The approach provides practical value for hazard evaluation and early warning and offers a potentially transferable framework for other geophysical problems characterized by data scarcity.