Large-scale estimation of surficial sediment size in alpine landforms using UAV photogrammetry and machine learning.

Accurate estimation of surficial sediment size in alpine landforms such as talus slopes, rock glaciers, and moraines is crucial for understanding geomorphologic processes and predicting the potential impact of natural hazards. Traditional methods for measuring sediment size in these environments can be time-consuming and labor-intensive. Additionally, they are usually applied to selected areas and are rarely used to cover larger areas, making the development of more efficient approaches essential. This study presents a new method for estimating large-scale surficial sediment size based on unmanned aerial vehicle (UAV) photogrammetry and combining SediNet and PebbleCountAuto image analysis methods. SediNet is a configurable machine-learning framework for estimating either (or both) continuous and categorical variables from a photographic image of clastic sediment. SediNet can achieve subpixel resolutions because the dimensions of the grains aren't being measured directly. However, site-specific sediment sizes are necessary to train this model. On the other hand, PebbleCountAuto does not require any site calibration by using segmentation methods to delimitate the grains automatically and provide a full grain-size distribution. Our proposed methodology trains the SediNet model using the sediment sizes outputs of the PebbleCountAuto method. Our study area is the upper part of the Lake O'Hara watershed in the Canadian Rockies, composed of talus slopes and a large ice-cored moraine. We performed two types of UAV flights; high-altitude flights (~100 m height) to cover the whole study area with medium-to-high resolution orthomosaic (pixel resolution 3 cm) and low-altitude flights (~25 m height) at smaller patches to achieve high-resolution orthomosaic (pixel resolution 5-8 mm). First, the sediment size was estimated in the high-resolution patches with the PebbleCountAuto method. Then, these results were used to train the SediNet model and generate a large-scale sediment size estimation. Our results show that this combination of methods is a reliable and efficient approach for accurately estimating sediment sizes in alpine landforms. The use of UAV photogrammetry allows for the rapid and cost-effective collection of high-resolution imagery, while the combination of SediNet and PebbleCountAuto provides robust estimates of sediment size over large areas. This new method can improve our understanding of geomorphologic processes and hazard assessment in these environments.