Abstract Empirical-statistical models of debris-flow are challenging to implement in environments where sedimentary and hydrologic triggering processes change through time, such as after a large earthquake. The flexible and adaptive statistical methods provided by machine learning algorithms may improve the quality of debris flow predictions where triggering conditions and the nature of sediment that can bulk flows varies with time. We developed a hybrid machine-learning model of future debris-flow volumes using a dataset of measured debris-flow volumes from 60 catchments that generated post-Wenchuan Earthquake (Mw 7.9) debris flows. We input topographic variables (catchment area, topographic relief, channel length, distance from seismic fault, and average channel gradient) and the total volume of co-seismic landslide debris into the PSO-ELM_AdaBoost machine-learning model, created by combining Extreme learning machine (ELM), particle swarm optimization (PSO) and adaptive boosting machine learning algorithm (AdaBoost). The model was trained and tested using post-2008 Mw 7.9 Wenchuan Earthquake debris flows, then applied to understand potential volumes of post-earthquake debris flows associated with other regional earthquakes (2013 Mw 6.6 Lushan Earthquake, 2010 Mw 6.9 Yushu Earthquake). We compared the PSO-ELM_Adaboost method with different machine learning methods, including back-propagation neural network (BPNN), support vector machine (SVM), ELM, PSO-ELM. The Comparative analysis demonstrated that the PSO-ELM_Adaboost method has a higher statistical validity and prediction accuracy with a mean absolute percentage error (MAPE) less than 0.10. The prediction accuracy of debris-flow volumes trigged by other earthquakes decreases to 0.11–0.16 (absolute percentage error), suggesting that once calibrated for a region this method can be applied to other regional earthquakes. This model may be useful for engineering design to mitigate the risk of large post-earthquake debris flows.

中文翻译：

---

估计潜在泥石流体积的混合机器学习模型

摘要 泥石流的经验统计模型在沉积和水文触发过程随时间变化的环境中实施具有挑战性，例如在大地震之后。机器学习算法提供的灵活和自适应的统计方法可以提高泥石流预测的质量，其中触发条件和沉积物的性质可以随时间变化。我们使用来自汶川地震后（Mw 7.9）泥石流产生的 60 个流域的实测泥石流体积数据集开发了未来泥石流体积的混合机器学习模型。我们输入地形变量（汇水面积、地形起伏、河道长度、距地震断层的距离、和平均通道梯度）和同震滑坡碎片的总体积进入 PSO-ELM_AdaBoost 机器学习模型，该模型是通过结合极限学习机 (ELM)、粒子群优化 (PSO) 和自适应增强机器学习算法 (AdaBoost) 创建的. 该模型使用 2008 年汶川地震后 Mw 7.9 泥石流进行训练和测试，然后应用于了解与其他区域地震（2013 Mw 6.6 芦山地震，2010 Mw 6.9 玉树地震）相关的震后泥石流的潜在体积。我们将 PSO-ELM_Adaboost 方法与不同的机器学习方法进行了比较，包括反向传播神经网络 (BPNN)、支持向量机 (SVM)、ELM、PSO-ELM。比较分析表明，PSO-ELM_Adaboost 方法具有更高的统计有效性和预测准确性，平均绝对百分比误差 (MAPE) 小于 0.10。由其他地震触发的泥石流体积预测精度下降到0.11-0.16（绝对百分比误差），表明该方法一旦对一个区域进行校准，就可以应用于其他区域地震。该模型可用于工程设计，以减轻大型震后泥石流的风险。