TY - JOUR
T1 - CNN-Based Intelligent Method for Identifying GSD of Granular Soils
AU - Zhang, Pin
AU - Yin, Zhen Yu
AU - Chen, Wen Bo
AU - Jin, Yin Fu
N1 - Funding Information:
This research was financially supported by the Research Grants Council (RGC) of Hong Kong Special Administrative Region Government (HKSARG) of China (Grant No. 15220221).
Publisher Copyright:
© 2021 American Society of Civil Engineers.
PY - 2021/12/1
Y1 - 2021/12/1
N2 - Different from conventional methodology, this study presents an intelligent approach to fast identify the grain-size distribution (GSD) of granular soils using a convolutional neural network (CNN) under a deep learning framework. A database including 279 images of granular soils with their GSDs is first created. Then, the framework of the CNN is tailored to identify GSD. The CNN-based model is trained to predict 11 grain sizes corresponding to 1%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100% of granular soils passing (i.e., d1, d10, ..., d100) using 80% of images, followed by the model testing using the rest 20% of images. By feeding an image of a soil sample into the proposed CNN-based model, the GSD can be predicted within several seconds. The predicted GSD exhibits excellent agreement with the measured one with an average error of 2.29% on the testing sets. It can be concluded that the proposed CNN-based model successfully provides a new intelligent way to fast, accurately, and conveniently identify the GSD of granular soils through images of soils.
AB - Different from conventional methodology, this study presents an intelligent approach to fast identify the grain-size distribution (GSD) of granular soils using a convolutional neural network (CNN) under a deep learning framework. A database including 279 images of granular soils with their GSDs is first created. Then, the framework of the CNN is tailored to identify GSD. The CNN-based model is trained to predict 11 grain sizes corresponding to 1%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100% of granular soils passing (i.e., d1, d10, ..., d100) using 80% of images, followed by the model testing using the rest 20% of images. By feeding an image of a soil sample into the proposed CNN-based model, the GSD can be predicted within several seconds. The predicted GSD exhibits excellent agreement with the measured one with an average error of 2.29% on the testing sets. It can be concluded that the proposed CNN-based model successfully provides a new intelligent way to fast, accurately, and conveniently identify the GSD of granular soils through images of soils.
KW - Grain-size distribution
KW - Gravels
KW - Image recognition
KW - Neural networks
KW - Sand
UR - http://www.scopus.com/inward/record.url?scp=85115810914&partnerID=8YFLogxK
U2 - 10.1061/(ASCE)GM.1943-5622.0002214
DO - 10.1061/(ASCE)GM.1943-5622.0002214
M3 - Journal article
AN - SCOPUS:85115810914
SN - 1532-3641
VL - 21
JO - International Journal of Geomechanics
JF - International Journal of Geomechanics
IS - 12
M1 - 04021229
ER -