[H-DS07] Landslides and related phenomena

[HDS07-15] Comparison of YOLO and BASEGRAIN approaches for analysing sediments in debris flow channel deposits

*Samikshya Dahal¹, Fumitoshi Imaizumi², Shoki Takayama³, Tomoya Osada² (1.Gifu University, 2.Shizuoka University, 3.Kyoto University)

Keywords:Debris flow, Grain size, UAV photogrammetry, Machine learning, BASEGRAIN

Debris flow is a mixture of sediments and water that travels down steep slopes at high speeds, causing significant destruction. It transports sediments that range in size from millimetre-scale clay to meter-sized boulders, which are deposited in channels with their proportion varying spatially. Since the sediments in a channel influence future debris flows, research to find a solution for obtaining precise grain size distribution information is important. To achieve this, we assessed the effectiveness of two approaches in sediment analysis of debris flow channel deposits: Automatic object detection software BASEGRAIN and a convolutional neural network (CNN)-based model YOLOv8. We utilised orthomosaics and digital elevation models (DEMs) created through unmanned aerial vehicle structure-from-motion (UAV-SFM) photogrammetry to analyse the grain size and catchment topography of the debris flow channel. Results from BASEGRAIN and YOLOv8 were validated against field grid sampling results. BASEGRAIN software, which specialised in grain size analysis of fluvial gravel beds, was efficient in analysing boulders; however, it showed limitations in detecting sediments of debris flow deposits that vary significantly in shape, size, and geological composition. Debris flow sediments often have irregular shapes and include debris like dead leaves and wood, which limited BASEGRAIN’s automatic analysis. Manual removal of vegetation and wood was necessary for effective analysis. In contrast, YOLOv8 showed advantages in identifying angular sediments, classifying geological materials such as sandstone and shale, and detecting vegetation and wood, as these variations were used to train the model. The results indicated that the automatic object detection approach used by BASEGRAIN allows for the initiation of the sediment detection process without requiring model training. However, when addressing debris flow sediments with irregular shapes and varying sizes and materials, YOLOv8 demonstrated better sediment detection capabilities. These findings have significant implications for understanding the spatio-temporal variations in sediments within debris flow channels, also facilitating their integration into debris flow modelling.