Rapid visual assessment of soil physical properties using image processing and convolutional neural networks

Date of Publication

7-2025

Document Type

Master's Thesis

Degree Name

Bachelor of Science in Civil Engineering (Honors) - Ladderized

Subject Categories

Civil Engineering

College

Gokongwei College of Engineering

Department/Unit

Civil Engineering

Thesis Advisor

Erica Elice S. Uy
Joenel G. Galupino

Defense Panel Chair

Mary Ann Q. Adajar

Defense Panel Member

Miller DL. Cutora
Jomari F. Tan

Abstract (English)

The assessment of soil physical properties is a fundamental step in geotechnical engineering, as these properties influence the design and stability of various structures. In this study, image processing techniques and a convolutional neural network (CNN) model was applied to rapidly assess soil physical properties. An image-based tool was developed to establish particle size, gradation characteristics, moisture content, Atterberg limits, and soil classification according to the Unified Soil Classification System (USCS) and AASHTO Soil Classification System.


Soil samples were subjected to standard ASTM procedures to determine reference values for validation, while images were captured in a controlled setup with calibration for distortion and scaling. The particle size was estimated on loosely packed particles using a clustering approach in terms of sieve size where the particles were retained, achieving an accuracy of 94.53% when the maximum dimension was used for classification. It was observed that the proposed tool shows potential use but is limited up to particles retained on a No. 40 sieve.

A CNN based on Light Soil-Net by Pandiri et al. (2024) was used to develop single output models for moisture content, modal particle size, and USCS classification. The model achieved 85.50% accuracy in moisture estimation (0% to 25%), 100.00% accuracy in modal particle size classification (4.75 mm to 0.425 mm), and 87.00% accuracy in USCS classification across nine classes. These results demonstrate the potential of image-based soil property estimation, but it remains limited to the samples used.

A multi-output CNN model was also developed for regression and classification purposes. The root mean square error (RMSE) varied by property, with larger errors in broader range parameters such as liquid limit. Only %Sands and %Fines had R² values above 0.70. The classification accuracy ranged from 53.59% to 60.20%, lower than the accuracy of the single output models. Despite this, Welch’s t-tests showed consistent performance across datasets. However, the model failed to classify images with at least 80% accuracy. While the multioutput CNN model showcased significant potential for use in real-world applications, further refinement and larger datasets are recommended to improve accuracy and extend applicability to broader soil classes and conditions. Overall, this study highlighted the potential use of rapid visual assessment of soil physical properties through image-based analysis.

Abstract Format

html

Abstract (Filipino)

None

Abstract Format

html

Language

English

Format

Electronic

Keywords

Soil structure

Upload Full Text

wf_yes

This document is currently not available here.

Share

COinS