An operational Early Warning Decision Framework For Debris FlowIntegrating Rainfall Thresholds and Seismic Signal Classification

Debris flows and landslides are frequently triggered by intense rainfall and are characterized by sudden onset and short warning lead times. Conventional early warning approaches that rely solely on rainfall thresholds are prone to false alarms or missed warnings due to spatial variability in rainfall and differences in actual slope conditions. To improve warning accuracy and operational applicability, this study proposes a novel early warning operational framework for debris flows that integrates rainfall thresholds, seismic monitoring, and near-real-time source classification into a multi-level, dynamic warning system. The proposed framework is implemented and evaluated in the Putunpunas River in Kaohsiung City, southern Taiwan, where a total of 46 documented debris-flow events were compiled and analyzed. Debris-flow occurrences were identified and confirmed through the combined use of riverine seismic signals and time-lapse camera observations, enabling reliable event detection and temporal validation. Based on reconstructed rainfall events, an empirical rainfall threshold was established using event duration (D) and effective cumulative rainfall (E),expressed as:

𝐸𝐷𝐹 = (14.1 ± 3.0)𝐷0.55±0.1

To assess whether a warning model trained on historical experience can successfully predict future debris-flow occurrences, this study further adopts a machine learning–based decision tree approach using the C5.0 algorithm to train the event classification model. This strategy allows objective evaluation of the predictive capability and generalization performance of the proposed integrated early warning framework under unseen event conditions, thereby enhancing its reliability and practical applicability for real-time debris-flow early warning operations. 

The proposed system first evaluates rainfall conditions using real-time precipitation data and applies three warning levels—alert, management, and action—corresponding to exceedance probabilities of 5%, 10%, and 20%, respectively, as an initial risk screening mechanism. When rainfall conditions exceed the defined thresholds, modules of seismic source detection and landslide monitoring (GeoLoc scheme) are simultaneously activated to detect potential landslides in real time. Furthermore, artificial intelligence (AI) based debris flow classifier is adopted to identify whether debris flow events have actually occurred. Compared to conventional rainfall threshold–based debris-flow early warning systems, our proposed approach enables real-time monitoring of upstream sediment supply associated with landslide occurrence and provides a secondary verification using riverine seismic signals.

This operational early warning framework enables to real-time assess rainfall threshold, landslide detection, and classify debris flow source, thereby enhancing the reliability and practical value of debris flow early warning and serving as a core component for future smart disaster prevention and real-time risk management systems. The framework was evaluated during Typhoon Fung-wong in November 2025. A warning was issued once rainfall exceeded the alert threshold based on real-time precipitation data, followed by activation of landslide monitoring and debris-flow detection modules. Using microtremor seismic signal analysis and AI-based event classification, the system verified event occurrence. During the event, only one out of six rainfall stations in the Putunpunas River exceeded the rainfall threshold, highlighting strong spatial variability in rainfall-induced hazard potential; nevertheless,the system was able to reflect actual hazard conditions in near real time through postevent verification and status updating,demonstrating its operational reliability.