The shape parameters of sand particles can be determined by either a Krumbein-Sloss chart or mathematical computation based on the particle image. However, these approaches are limited by user-dependent uncertainty and complicated algorithms with computational issues. This study explores the feasibility of predicting the shape parameters of sand particles via a well-trained deep learning model. Over 7000 sand particle images from six different sand types were used, and the corresponding shape parameters such as sphericity, roundness, slenderness, and circularity were computed. Inception-ResNet-v2 whose performance has been widely validated, as one of convolutional neural network-based model, used the pair dataset (i.e., image and each computed shape parameter) and comprised a pre-trained network for feature extraction and a regression output layer. Strategic data augmentation helped in increasing the number of training data to reduce the test loss efficiently. Several hyperparameters were carefully tuned to accomplish model optimization, while the generalization ability of the model was evaluated during training by using a validation dataset. The prediction results showed that the trained model yielded a highly accurate and precise prediction of the shape parameters, regardless of the image types. The predicted roundness had relatively scattered values in comparison with other parameters, presumably because of its indeterminate definition of the ground-truth. Thus, the proposed trained model enables accurate prediction of the shape parameters of sand particles solely based on an image containing the complete shape of the particle, without mathematical computation.
|Number of pages||11|
|Publication status||Published - 2022 Apr|
Bibliographical noteFunding Information:
This work was supported by the Land and Housing Institute (LHI) grant funded by the Korea Land and Housing Corporation, and the National Research Foundation of Korea (NRF) Grant funded by the Korea government (MSIT) (Nos. 2020R1A2C1014815, NRF-2021R1A5A1032433).
© 2022, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.
All Science Journal Classification (ASJC) codes
- Geotechnical Engineering and Engineering Geology
- Earth and Planetary Sciences (miscellaneous)