SeismicDF: A multi-regional benchmark dataset of seismic signals for advancing deep learning in debris flow warning

Accurate warning of debris flows is crucial for mitigating geohazard losses. Although warning methods combining environmental seismic signals with deep learning show immense potential, this field is currently in its infancy. The lack of comprehensive, standardized benchmark datasets severely limits the development, cross-domain application, and fair performance evaluation of deep learning models. To address this research gap, we officially open-source SeismicDF—a standardized benchmark dataset and end-to-end warning evaluation platform specifically designed for debris flow seismic signals. Key Features: Standardized Large-Scale Training Set: The system integrates 179 historical debris flow events from 9 global regions, constructing a large-scale training dataset containing 14,476 high-quality samples. Long-Term Field Evaluation Network: Continuous seismic waveform data exceeding 600 days are provided across three regions, covering 167 real-world debris flow events. This serves as a highly representative performance validation platform for real-time warning algorithms. End-to-End Warning Rule System: Bridging the gap between laboratory research and real-world deployment by tailoring physical and logical rules for real-time warnings in complex field environments. Comprehensive Warning Score: Introduces the "Comprehensive Warning Score" metric to help researchers intuitively quantify model performance and scientifically select the optimal warning architecture. Highly Extensible Model Benchmark: Built-in deep comparative benchmarks for CNN, ResNet, LSTM, and Transformer architectures, allowing researchers to integrate the latest custom network architectures with zero barriers. The release of SeismicDF aims to promote fair and transparent performance evaluation of deep learning methods within the global geoscientific community and accelerate the evolution of data-driven debris flow warning technologies. Yuanwei Song | Institute of Mountain Hazards and Environment, Chinese Academy of Sciences (IMHE, CAS) Contact: [e-mail] ################################################################################################################## 准确的泥石流预警对于减轻地质灾害损失至关重要。尽管结合环境地震信号与深度学习的预警方法展现出巨大潜力,但该领域目前仍处于起步阶段,且由于缺乏全面、标准化的基准数据集,严重制约了深度学习模型的开发、跨域应用及公平性评估。 为了填补这一科研空白,本项目正式开源 SeismicDF —— 针对泥石流地震信号设计的标准化基准数据集与端到端预警评估平台。 核心亮点: 标准化大规模训练集:系统整合了全球 9 个地区的 179 场历史泥石流事件,构建了包含 14,476 个高质量样本的大型训练数据集。 长时序野外评估台网:三个区域提供超过 600 天的连续地震波形数据,涵盖 167 场真实泥石流事件,旨在为实时预警算法提供极具代表性的性能验证平台。 全链路预警规则体系:打破实验室与现实的壁垒,量身定制了一套适用于野外复杂环境的实时预警物理与逻辑规则。 综合预警分数:首创“综合预警分数(Comprehensive Warning Score)”指标,帮助研究人员直观量化模型表现,科学选定最佳预警架构。 高度开放的模型基准:内置 CNN、ResNet、LSTM 和 Transformer 的深度对比基准,并支持研究人员零门槛接入最新的自定义网络架构。 SeismicDF 的发布旨在推动全球地球科学界对深度学习方法进行公平、透明的性能评估,加速由数据驱动的泥石流灾害事件预警技术的演进。 宋元伟 | 中国科学院成都山地灾害与环境研究所 (IMHE, CAS) 联系方式: [e-mail]