A Study on Prediction of Volcanic Mudflows on Miyake Island Using Machine Learning

<p>1. Introduction</p><p>Miyake Island erupted five times in 2000, producing volcanic mudflows. As a result, mudflows were deposited in residential areas around the island, and bridges on the road that circles the island collapsed. However, the occurrence of these volcanic mudflows and the distribution of the damage may vary depending on the deposition of volcanic ash and the intensity and spatial distribution of precipitation. Therefore, it is important to identify the topographic locations where mudflows are likely to occur and accumulate so that countermeasures can be taken against future volcanic mudflows. For these reasons, this study uses decision trees, a type of machine learning, to predict the occurrence of volcanic mudflows on Miyake Island, and to simulate mudflows to study the arrival time and extent of volcanic mudflows.</p><p></p><p>2 Estimation of the location of volcanic mudflow generation and deposition using decision trees</p><p>Eight types of topographic characteristics (elevation, slope, plan curvature, profile curvature, erosion height, uneroded height, sediment transport index, and wetness index) were calculated using a 1-meter resolution DEM acquired by aircraft laser survey as explanatory variables for the decision tree, and the descriptive statistics of each topographic characteristics were calculated on a 50-meter mesh. The descriptive statistics of each geomorphic characteristiccs were aggregated on a 50-meter mesh. The occurrence and depositional areas of volcanic mudflows, which were determined from aerial photographs taken in November 2000, were converted to the 50-meter mesh described below and used as objective variables.</p><p>As a result, a decision tree with a correctness of 68.85% and a kappa coefficient of 0.528 was obtained from the weighted confusion matrix. Figure 1 shows the obtained tree and a map of the ends of the tree.</p><p></p><p>Figure 1 shows the distribution of the tree ends as follows. Erosion 1 is located near the summit of Oyama, where a large amount of volcanic ash was deposited by the 2000 eruption. On the east side of Mt. Oyama, mudflows generated by Erosion 1 join mudflows generated by Erosion 2 and Erosion 3 and flow down to the coast along the valleys of Erosion 2, Erosion 4 and Erosion 6. The mudflows that flowed down to the coast overflowed the channel of sediment 4 in the flat area near the coast and deposited in sediment 5. On the west side of Oyama, volcanic mudflows generated by erosion 1 are deposited in the meshes of sediments 1 and 3 (near the former village farm). From erosion 2 and 3, mudflows are generated by secondary erosion of the mudflows deposited in sediments 1 and 3 and by erosion of the volcanic ash deposited on the west side of Mt. The mudflows that flowed down to the coast overflowed the channel of sediment 4 in the flat area near the coast and were deposited in sediment 5.</p><p></p><p>3 Calculating the arrival time and arrival range of mudflows using a mudflow model</p><p>Numerical simulations of volcanic mudflows were conducted to determine the arrival time and arrival area of volcanic mudflows. The aforementioned 1-meter DEM was downscaled to 5-meter resolution, and calculations were performed in three areas of Miyakejima (Kamitsuki area, Tsubota area, and Yukeihama and Village Farm areas). In the Kamitsuki area, the mudflow generated near the top of the caldera reaches Kamitsuki in about one hour. In the Kamitsuki area, the mudflow generated near the top of the caldera takes about one hour to reach Kamitsuki, and several erosion control dams along the downstream valley are found to delay the arrival of the mudflow. In Yukeihama and Village Farmarea, the mudflow reaches the former Village Farm within 20 minutes after the occurrence of the mudflow, and from there it reaches the Yukeihama in about 40 minutes.</p><p>View PDF for the rest of the abstract</p>