Basic data of sixty debris flows.

Debris flow is a sudden natural disaster in mountainous areas, which seriously threatens the lives and property of nearby residents. Therefore, it is necessary to predict the volume of debris flow accurately and reliably. However, the predictions of back propagation neural networks are unstable and inaccurate due to the limited dataset. In this study, the Cubic map optimizes the initial population position of the whale optimization algorithm. Meanwhile, the adaptive weight adjustment strategy optimizes the weight value in the shrink-wrapping mechanism of the whale optimization algorithm. Then, the improved whale optimization algorithm optimizes the final weights and thresholds in the back propagation neural network. Finally, to verify the performance of the final model, sixty debris flow gullies caused by earthquakes in Longmenshan area are selected as the research objects. Through correlation analysis, 4 main factors affecting the volume of debris flow are determined and inputted into the model for training and prediction. Four methods (support vector machine regression, XGBoost, back propagation neural network optimized by artificial bee colony algorithm, back propagation neural network optimized by grey wolf optimization algorithm) are used to compare the prediction performance and reliability. The results indicate that loose sediments from co-seismic landslides are the most important factor influencing the flow of debris flows in the earthquake area. The mean absolute percentage error, mean absolute error and R2 of the final model are 0.193, 29.197 × 104 m3 and 0.912, respectively. The final model is more accurate and stable when the dataset is insufficient and under complexity. This is attributed to the optimization of WOA by Cubic map and adaptive weight adjustment. In general, the model of this paper can provide reference for debris flow prevention and machine learning algorithms.泥石流（debris flow）是山区突发性自然灾害，严重威胁周边居民的生命与财产安全，因此精准可靠地预测泥石流体积具有重要现实意义。然而，受限于数据集规模，反向传播神经网络（back propagation neural network）的预测结果往往不稳定且精度不足。本研究首先利用三次映射（Cubic map）优化鲸鱼优化算法（whale optimization algorithm）的初始种群位置；同时引入自适应权重调整策略，优化鲸鱼优化算法收缩包裹机制中的权重参数。随后，将改进后的鲸鱼优化算法用于优化反向传播神经网络的最终权重与阈值。为验证所提模型的性能，本研究选取龙门山地震区60条震后泥石流沟谷作为研究对象。通过相关性分析，确定影响泥石流体积的4项主要因子，并将其作为模型输入开展训练与预测。选取4种方法进行对比以验证模型的预测性能与可靠性：支持向量机回归（support vector machine regression）、XGBoost、基于人工蜂群算法（artificial bee colony algorithm）优化的反向传播神经网络、基于灰狼优化算法（grey wolf optimization algorithm）优化的反向传播神经网络。研究结果表明，同震滑坡产生的松散沉积物是地震区泥石流发育的最关键影响因子。所提模型的平均绝对百分比误差（mean absolute percentage error）、平均绝对误差（mean absolute error）与决定系数R²分别为0.193、29.197×10⁴ m³与0.912。在数据集规模有限且场景复杂的条件下，所提模型具备更高的预测精度与稳定性，这得益于三次映射与自适应权重调整对鲸鱼优化算法的优化改进。总体而言，本研究构建的模型可为泥石流防灾减灾工作以及机器学习算法的应用提供参考依据。