Optical granulometry coupled with deep learning offers a rapid alternative to traditional mechanical sieving for soil characterization. This thesis presents the development of an open dataset, named Photogranulometry, comprising 12,714 high-resolution images of 316 soil samples annotated with their laboratory particle size distribution curves. To process this high visual dimensionality under weak supervision, an architecture based on attention-based multiple instance learning (ABMIL) and a DINOv2 feature extractor is introduced. This method analyzes photographs as grouped tiles to predict the global particle size distribution curve without explicit segmentation. The model achieves a median weighted mean absolute error of 4.06%. The study demonstrates the viability of the approach on moist samples, which eliminates the laboratory drying step and enables in-situ geotechnical analysis.
Plante St-Cyr, T. (Author),
Duhaime (Supervisor) &
Dubé (Co-supervisor),
31 Jul 2026Student thesis: Master's thesis › Master in Engineering: Construction Engineering