https://doi.org/10.1002/esp.70093

Grain-size analysis offers insights into geological processes and landform dynamics.
Traditional grain-size sampling methods are labour intensive and offer limited spatial
coverage, posing challenges in paraglacial and periglacial environments characterized
by large spatial variability in sediment sizes. This study introduces a new workflow
that combines structure-from-motion, image segmentation and texture-based optical
granulometry techniques to estimate surface grain size in paraglacial and periglacial
environments efficiently. Utilizing high-resolution orthomosaics (ground sampling
distance 8 mm) and Cellpose, a deep-learning image segmentation model, the new
workflow achieves high-accuracy grain-size distributions (GSDs) with low errors.
These GSDs, along with lower resolution orthomosaics (ground sampling distance
30 mm), are used to train SediNet—a machine-learning framework—to predict GSDs
accurately from 340 340 pixel tiles. Tested across six alpine basins in the Canadian
Rockies and a rock glacier in Italy, the model demonstrates effectiveness and accu-
racy, promising advancements in geoscientific research and the understanding of
paraglacial and periglacial dynamics.

https://doi.org/10.1002/esp.70093