10:00 AM - 10:15 AM
[STT34-05] Development of automatic patrol, automatic photography and automatic extraction by unmanned aerial vehicles (UAVs) in case of volcanic eruptions
★Invited Papers
Keywords:UAV, Volcano Disaster, automation, AI
Safe investigation by unmanned aerial vehicles (UAVs) is expected because flying manned aircraft over volcanic craters during volcanic eruptions is risky. Since 2016, the Ministry of Education, Culture, Sports, Science and Technology's Next Generation Volcano Research Promotion Project has been investigating "Developing a method for real-time understanding of volcanic disasters using unmanned aerial vehicles (drones, etc.)" as sub-theme 1 of Issue D.
In the research of the Next Generation Volcano Research Promotion Project, we have mainly taken a large number of images with UAVs and created three-dimensional models using SfM-MVS (Structure from Motion / Multi-view Stereo) technology. The three-dimensional data obtained can be used to produce a three-dimensional red map and topographic interpretation to determine the distribution area of volcanic ejecta. In addition, the thickness and volume of volcanic ejecta such as lava flows can be estimated by differential analysis before and after eruption.
On the other hand, we are also investigating a method that does not create a three-dimensional model, but uses a method used in the field of infrastructure to automatically patrol predetermined points, take images and extract changes. Fixed-point observation is important for recording changes in the location of fumaroles and geothermal activity during volcanic eruptions. However, it is difficult to observe the same point from the sky for each observation, so UAVs are expected to be used for this purpose. We conducted a demonstration experiment using the Nishiyama crater group formed by the eruption of Usu volcano in 2000. The equipment used was a Matrice300RTK (airframe) and H20T (camera), and automatic patrol and photography were conducted using the "live mission function". The live mission function records the aircraft's position information, camera direction, zoom and other shooting operations during the first flight, and automatically takes pictures along the same route and at the same angle of view during the second and subsequent flights. This allows images to be taken from virtually the same position during the day and at night.
The manufacturer's technological development has made it possible to achieve automatic patrolling and photography, which was not possible at the start of the research, and we are working to improve the analysis of the acquired images. One of the challenges in creating a three-dimensional model is the loss of accuracy due to smokes in the image. As the direction of smokes changes with wind direction, it is possible to obtain an image without smokes by flying the UAV at different times. In order to quickly extract images with smokes, we investigated automatic extraction using machine learning on the vertical images taken (Mori et al., 2023). By training on images taken after the 2016 Mt. Aso eruption using images of smokes, clouds, factory smoke, and wildfire smoke, we were able to extract the majority of smokes. Automatic extraction was performed on oblique photographs taken on Mt. Usu using this learning model, but extraction accuracy was low, partly due to the difference between vertical and oblique images, so improving accuracy on oblique images will be a future challenge. Improving the accuracy for oblique images will be a future challenge. The location (latitude, longitude and elevation) of the extracted smokes can be immediately identified by importing them into the oblique photogrammetric system (Fig. 1). In the future, we would like to further develop the automatic extraction of smokes from images and work on a method to automatically extract the location of bombs arrival point and the distribution area of the lava flow.
In the research of the Next Generation Volcano Research Promotion Project, we have mainly taken a large number of images with UAVs and created three-dimensional models using SfM-MVS (Structure from Motion / Multi-view Stereo) technology. The three-dimensional data obtained can be used to produce a three-dimensional red map and topographic interpretation to determine the distribution area of volcanic ejecta. In addition, the thickness and volume of volcanic ejecta such as lava flows can be estimated by differential analysis before and after eruption.
On the other hand, we are also investigating a method that does not create a three-dimensional model, but uses a method used in the field of infrastructure to automatically patrol predetermined points, take images and extract changes. Fixed-point observation is important for recording changes in the location of fumaroles and geothermal activity during volcanic eruptions. However, it is difficult to observe the same point from the sky for each observation, so UAVs are expected to be used for this purpose. We conducted a demonstration experiment using the Nishiyama crater group formed by the eruption of Usu volcano in 2000. The equipment used was a Matrice300RTK (airframe) and H20T (camera), and automatic patrol and photography were conducted using the "live mission function". The live mission function records the aircraft's position information, camera direction, zoom and other shooting operations during the first flight, and automatically takes pictures along the same route and at the same angle of view during the second and subsequent flights. This allows images to be taken from virtually the same position during the day and at night.
The manufacturer's technological development has made it possible to achieve automatic patrolling and photography, which was not possible at the start of the research, and we are working to improve the analysis of the acquired images. One of the challenges in creating a three-dimensional model is the loss of accuracy due to smokes in the image. As the direction of smokes changes with wind direction, it is possible to obtain an image without smokes by flying the UAV at different times. In order to quickly extract images with smokes, we investigated automatic extraction using machine learning on the vertical images taken (Mori et al., 2023). By training on images taken after the 2016 Mt. Aso eruption using images of smokes, clouds, factory smoke, and wildfire smoke, we were able to extract the majority of smokes. Automatic extraction was performed on oblique photographs taken on Mt. Usu using this learning model, but extraction accuracy was low, partly due to the difference between vertical and oblique images, so improving accuracy on oblique images will be a future challenge. Improving the accuracy for oblique images will be a future challenge. The location (latitude, longitude and elevation) of the extracted smokes can be immediately identified by importing them into the oblique photogrammetric system (Fig. 1). In the future, we would like to further develop the automatic extraction of smokes from images and work on a method to automatically extract the location of bombs arrival point and the distribution area of the lava flow.