Particle size analysis is one of the most frequently requested tests in geotechnical analysis laboratories. This test determines the distribution of particle sizes in a soil (particle size distribution). Traditionally, sieving and sedimentometry are used for particle size analysis of granular soils and fine soils, respectively. Although reliable, these techniques have the disadvantage of being time-consuming. Existing image analysis tools such as Wipfrag reduce this analysis time. However, they are limited to large particles and use segmentation approaches that are sensitive to image resolution.
This thesis proposes a deep learning-based approach to estimating particle size from photographs in order to overcome the limitations of traditional methods and current image analysis software. To this end, a dataset of synthetic, segmented images was generated using the Unity game engine to reproduce realistic soil textures. Conditional entropy and an object segmentation model (GrainSegNet) were used to evaluate their visual realism. Subsequently, a convolutional neural network (GrainSizeNet) was developed to analyze the influence of input image types (top/bottom views, with or without segmentation) and the contribution of the generated synthetic images compared to traditional data augmentation methods. Finally, GrainSizeNet was adapted to analyze photographs of split spoon samples (SS) taken during the standard penetration test in order to distinguish five soil categories from the USCS (Unified Soil Classification System) and predict the mass fraction of fine particles (diameter less than 0.075 mm) by comparing direct and hierarchical learning.
Conditional entropy reveals a similarity between synthetic and real images, and GrainSegNet's evaluation on real images highlighting satisfactory performance on laboratory photographs (average IoU of 0.60). However, performance decreases in more complex contexts (average IoU of 0.26), particularly for materials with high particle density or finer particles. These results highlight the relevance of synthetic images while showing the need for further adaptations for environments that are very different from those simulated.
The GrainSizeNet variants, evaluated with different input data configurations, show similar performance, with Mean Absolute Errors (MAE) ranging from 2.90% to 3.66% calculated on the mass percentages retained on the 11 control sieves of the BNQ 2501-025 standard, indicating that a single top view image is sufficient to ensure a reliable estimate of particle size distribution. Furthermore, the integration of synthetic images improves the generalization of the model: the MAE was reduced from 11.54% (dataset composed solely of real images) to 6.19%.
When applied to field data, the model produces promising results. Hierarchical strategies offer the best performance for classification (precision = 69%) and for estimating the proportion of fines (MAE = 15.3%). Despite these encouraging results, GrainSizeNet remains dependent on the quantity and quality of the data available for training. Furthermore, its performance improves as it is fed with larger and more diverse datasets.
Simo Tenekam, H. (Author),
Duhaime (Supervisor),
Dubé (Co-supervisor) &
Toews (Co-supervisor),
17 Jun 2026Student thesis: Doctoral thesis › Doctorate in Engineering: Engineering