Shusuke Miyata

The integration of machine learning (ML) into debris flow prediction in volcanic areas, exemplified by the Gendol River watershed of Mount Merapi, offers transformative potential for hazard mitigation. This study aimed to develop real-time, computationally efficient ML models capable of integrating multi-source data, rainfall intensity of 25 mm/hour linked to 300 cm Debris Flow heights, antecedent precipitation, and geomorphological variables to predict debris flows with actionable lead times. Key objectives included optimizing prediction accuracy, minimizing the false positive rate to 18.2% for "Debris Flow" events, and enhancing model interpretability for deployment in data-scarce volcanic regions. Results demonstrated that ensemble methods and deep learning architecture outperformed traditional models, with Efficient Logistic Regression and Linear SVM achieving an accuracy of 82.35%, and Cosine KNN attaining a prediction speed of 272 observations per second. Critical predictors included temporal rainfall patterns (contributing more than 50% to flow initiation) and ash deposit thickness (with a 70% influence on decision-making). However, challenges persisted: imbalanced datasets of nine training instances for "Debris Flow" events led to misclassification rates of 100% for hybrid events like "Rainfall and Debris Flow," while models like Naive Bayes exhibited instability (accuracy dropping to 50%). Research gaps highlighted data scarcity for high-magnitude events, limited geographic transferability, and the absence of standardized evaluation metrics. Technical limitations included reliance on low-resolution remote sensing data, high computational costs for ensemble models requiring 10 operational cost units, and the opacity of neural networks, which hindered stakeholder trust. Despite these constraints, ML models achieved 85% accuracy in non-event recognition and 76.47% precision in Bagged Trees, offering scalable frameworks for early warning systems. The study highlights the importance of enriched datasets, adaptive algorithms, and interdisciplinary collaboration in transforming volcanic risk management from a reactive approach, ultimately safeguarding vulnerable communities through data-driven, life-saving predictions.

It is empirically known that debris flows are more likely to occur during and after volcanic eruptions. This study conducted numerical simulations to examine the effects of reduced infiltration capacity and increased sediment supply owing to volcanic activity on the rainfall threshold that causes debris flows. Rainfall runoff analysis was performed using various rainfall intensities as constant input conditions. The occurrence of debris flows was evaluated based on theoretical and experimental formulas. The mechanism by which sediment in river channels transforms into debris flow was discussed in terms of two processes: sediment sliding as a mass and sediment being eroded by surface flow. This study clarified the critical role of infiltration capacity in debris flow occurrence. When infiltration capacity is low, surface runoff becomes more dominant, easily entraining sediment and triggering debris flows—even during short or weak rainfall events. Quantitative analysis revealed that when infiltration capacity drops from 100 mm/h to 10 mm/h, the cumulative rainfall required to trigger debris flows decreases to only a few percent of the pre-eruption requirement. However, when volcanic activity supplies large volumes of sediment to river channels, the mechanism can shift to mass sliding. In such cases, debris flows may occur even under relatively weak but prolonged rainfall. The results indicate the importance of both sediment erosion and mass sliding processes in debris flow occurrence.

Volcano eruption during the snow season may lead to a drastic snowmelt owing to eruption heat, resulting in a mudflow. This phenomenon is called “snowmelt-type volcanic mudflow.” A snowmelt-type volcanic mudflow has more water when compared with an ordinary mudflow caused by rainfall such that this type of mudflow can become larger and has the potential to reach more remote areas with greater speed. It triggers tremendous damage in lower reaches in some cases. Therefore, adequate measures should be taken against snowmelt-type volcanic mudflow, and monitoring should be conducted in a volcanic area with snow. However, a snowmelt-type volcanic mudflow rarely occurs and as only a few examples are available, this type of mudflow is less understood. Accordingly, this study examines the generation process of snowmelt-type volcanic mudflow using experiments. The following two types of experiments were conducted. (1) The speed of snow melting is determined by mixing snow and heated gravel. (2) The process in which gravel comes flying and penetrates the snow layer is examined. These experiments have clarified the following points. The speed of melting snow varies depending on the contact condition of snow and heated gravel; the speed of melting snow increases where there is a mixture of snow and heated gravel; under the same condition of the amount of gravel, the smaller the grain size is, the greater is the speed of snow melting; and there is a correlation between the kinetic energy of the gravel flying and the mass of the snow pushed aside by penetration of gravel into snow layer. The findings revealed that 5 cm diameter gravel can melt snow most effectively.

Wood travelling in rivers is a major hazard to lives and infrastructures, because tons of wood material can travel nearing the speed of the flood flow. If post-event mapping, detection and numerical simulation have made important progress, detecting in-flow driftwood in all weather and at all times still presents several challenges. The present work aims to expand the capacity to detect in-flow wood by adapting the ground-penetrating radar electromagnetic method. The laboratory test was carried out over a water circulation flume using an 800 MHz nominal frequency antenna sampling at 100 Hz and a video camera set on top of the flume to measure the average velocity of the wood logs. A set of single wood logs of 20 cm in length travelled underneath the antenna. The GPR results have demonstrated that the method had the potential to detect moving wood, and that it could “see” underneath the water to the shallow flume floor. The experiments resulted in the ability to count wood travelling underneath the antenna, and the instantaneous velocity was obtained with velocities ranging from 0.307 to 0.352 m/s, which was slightly higher than the average velocity measured from video imaging. This difference is explained by the in-flow acceleration of the wood after its introduction into the flume.

Reducing the damage due to landslide dam failures requires the prediction of flood hydrographs. Although progressive failure is one of the main failure modes of landslide dams, no prediction method is available. This study develops a method for predicting progressive failure. The proposed method consists of the progressive failure model and overtopping erosion model. The progressive failure model can reproduce the collapse progression from a dam toe to predict the longitudinal dam shape and reservoir water level when the reservoir water overflows. The overtopping erosion model uses these predicted values as the new initial conditions and reproduces the dam erosion processes due to an overtopping flow in order to predict a flood hydrograph after the reservoir water overflows. The progressive failure model includes physical models representing the intermittent collapse of a dam slope, seepage flow in a dam, and surface flow on a dam slope. The intermittent collapse model characterizes the progressive failure model. It considers a stabilization effect whereby collapse deposits support a steep slope. This effect decreases as the collapse deposits are transported downstream. Such a consideration allows the model to express intermittent, not continuous, occurrences of collapses. Field experiments on the progressive failure of a landslide dam were conducted to validate the proposed method. The progressive failure model successfully reproduced the experimental results of the collapse progression from the dam toe. Using the value predicted by the progressive failure model, the overtopping erosion model successfully reproduced the flood hydrograph after the reservoir water started to overflow.

The complete understanding of the mechanisms controlling debris-flow initiation is still an open challenge in landslide research. Most debris-flow models assume that motion suddenly begins when a large force imbalance is imposed by slope instabilities or the substrate saturation that causes the collapse of the channel sediment cover. In the real world, the initiation of debris flows usually results from the perturbation of the static force balance that retains sediment masses in steep channels. These perturbations are primarily generated by the increasing runoff and by the progressive erosion of the deposits. Therefore, great part of regional early warning systems for debris flows are based on critical rainfall thresholds. However, these systems are affected by large spatial-temporal uncertainties due to the inadequate number and distribution of rain gauges. In addition, rainfall analysis alone does not explain the dynamics of sediment fluxes at the catchment scale: short-term variations in the sediment sources strongly influence the triggering of debris flows, even in catchments characterized by unlimited sediment supply.In this work, we present multi-parametric observations of debris flows at the headwaters of the Gadria catchment (eastern Italian Alps). In 2018, we installed a monitoring network composed of geophones, three soil moisture probes, one tensiometer and two rain-triggered videocameras in a 30-m wide steep channel located at about 2200 m a.s.l. Most sensors lie on the lateral ridges of this channel, except for the tensiometer and the soil moisture probes that are installed in the channel bed at different depths. This network recorded four flow events in two years, two of which occurred at night. Specifically, the debris flows that occurred on 21 July 2018 and 26 July 2019 produced remarkable geomorphic changes in the monitored channel, with up to 1-m deep erosion. For all events, we measured peak values of soil water content that are far from saturation (<0.25 at -20 cm, <0.15 at -40 cm, <0.1 at -60 cm). We derived the time of occurrence and the duration of these events from the analysis of the seismic signals. Combining these pieces of information with data gathered at the monitoring station located about 2 km downstream, we could determine the flow kinematics along the main channel.These results, although still preliminary, show the relevance of a multi-parametric detection of debris-flow initiation processes and may have valuable implications for risk management. Alarm systems for debris flows are becoming more and more attractive due the continuous development of compact and low-cost distributed sensor networks. The main challenge for operational alarm systems is the short lead-time, which is few tens of seconds for closing a transportation route or tens of minutes for evacuating settlements. Lead-time would significantly increase installing a detection system in the upper part of a catchment, where the debris flow initiates. The combination of hydro-meteorological monitoring in the source areas and seismic detection of channelized flows may be a reliable approach for developing an integrated early warning - alarm system.