The widely adopted debris flow susceptibility (DFS) assessment methods rely on numerous debris flow hazard points. However, in field investigations, the hazard points obtained are often limited due to constraints imposed by the temporal and spatial scales of the study area. Consequently, conducting DFS assessments with a small amount of debris flow points—namely, a small sample debris flow database—has become a critical challenge that urgently needs to be addressed. This study proposes a novel framework integrating small-sample oversampling techniques with the Random Forests model from machine learning algorithms for DFS evaluation, establishing a systematic “debris flow sample collection–SMOTE sample extension–Random Forests susceptibility assessment” workflow. The introduced oversampling technique effectively addresses issues of insufficient accuracy, poor generalization capability, and slow convergence caused by limited samples. To validate the reliability and superiority of this framework, it was implemented for DFS assessment in the mountainous regions of central and northwestern Right-Wing Middle Banner of Horqin. Random Forests Model is introduced for longitudinal comparison, and traditional Information Content Model is also used for comparative analysis. Evaluation results demonstrate that debris flows in the study area primarily occur in valley regions and show a significant correlation with human activities. Model performance was quantitatively evaluated using Receiver Operating Characteristic curves and Kappa coefficients, confirming the superior performance of the proposed framework. The findings suggest that oversampling techniques enable satisfactory DFS assessments, with the newly established framework proving particularly suitable for geological hazard evaluation using a small-sample debris flow database. This advancement holds substantial implications for disaster prevention and mitigation strategies.
Ma et al. (Sat,) studied this question.