Influences de la méthode d'échantillonnage et de la présence d'une phase liquide sur l'estimation de la teneur en matière résiduelle d'un sol par analyse visuelle et intelligence artificielle

Abstract

The characterization of urban soils is crucial for their environmental management and rehabilitation planning. This research focuses on fill materials containing residual materials, specifically construction bricks. In the course of this work, ten sampling tests were conducted, combining ten configurations of sampling method, the presence of liquid, and, secondarily, the presence of fine particles. Through these ten tests, the goal was to study the representativeness of the configurations through granulometric and residual material concentration analyses for subsequent analysis by a convolutional neural network. The lots used for the tests were composed of gravelly soil particles smaller than 2.5 mm, with no particles smaller than 160 µm for tests without fine particles. These lots were artificially contaminated with particulate bricks at a mass concentration of 30%. The sampling methods employed were grab sampling, fractionated shoveling in 20 increments, and rotary separation. The liquid phases introduced during wet tests were water and canola oil, a safe substitute for petroleum products. This research confirmed the hierarchy of representativeness of the three studied sampling methods, with grab sampling being the least representative, followed by fractionated shoveling and rotary separation (only applicable to dry tests). The presence of a liquid phase significantly improved representativeness, especially for grab sampling. The presence of fines mainly affected the fundamental error, but this effect was minimal compared to segregation and grouping errors. However, soil and brick losses, including fine particles, slightly biased granulometric and brick quantification results, especially for wet tests. The results related to the image analysis highlighted the importance of rebalancing the photographic database used for training and validating the neural network to enable it to detect all brick concentrations, especially the most extreme ones. A diverse and extensive database required a complex neural network to converge. The adjustment of hyperparameters had specific effects on performance. The best network configurations generated losses around 0.012 and root mean square errors of around 0.035. Furthermore, the network generally identified false positives and false negatives similar to the sampling, thereby reinforcing confidence in the latter. However, the network's intrinsic error still added to the sampling error. A future objective would be to correct this intrinsic error.

Date	27 Nov 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Jean-Sébastien Dubé (Supervisor) & François Duhaime (Co-supervisor)