Deep Learning Based image segmentation for expensive soil desiccation crack recognition and qualification

Ling, Hui Yean (2025) Deep Learning Based image segmentation for expensive soil desiccation crack recognition and qualification. Master dissertation/thesis, UTAR.

PDF Download (12Mb)

Abstract

Expansive soils undergo significant volume changes due to moisture fluctuations, which lead to desiccation cracks formation that affect soil properties and engineering performance, compromising the safety of geo structures. The analysis of these cracks was essential for mitigating their impact; however, traditional quantification methods were labour intensive and imprecise, highlighting the need for more robust and automated techniques. This study investigated the feasibility and effectiveness of image-based techniques using advanced deep learning algorithms to quantify desiccation cracks in expansive soils. The objectives of the study included designing soil desiccation experiment setup for desiccation crack image acquisition, evaluating crack imaging analysis based on deep learning algorithms, and quantifying desiccation cracks through image processing techniques. Laboratory experiments were conducted using a custom-built image acquisition tool to capture crack images under simulated soil desiccation conditions. Crack images obtained were processed and annotated to produce a dataset of 820 images for the training and testing of deep learning models. Deep learning models, including U-Net, Res-UNet, and DeepLabv3+ with pre-trained backbones such as MobileNetV2, ResNet-18, ResNet-50, and Xception, were trained and evaluated along side a traditional Otsu's thresholding method as the baseline for crack detection and segmentation. The evaluation considered segmentation performance using evaluation metrics (precision, recall, F1 score, IoU), computational efficiency, and crack geometrical parameters quantification (surface crack ratio, crack width, crack length, and crack segment). Results demonstrated that DeepLabv3+ variants consistently outperformed other methods, with MobileNetV2 backbone offering the best balance of computational efficiency, segmentation accuracy, and robustness across case-wise performance conditions. Compared to traditional approaches, deep learning models, particularly with DeepLabv3+ variants, produced more reliable crack segmentation masks, thus enabling more accurate quantification of crack geometrical parameters, as demonstrated by lower error rates. This study validates the effectiveness of deep learning based segmentation methods for automated soil crack recognition and quantification, contributing to engineering applications with improved methodologies for analysing desiccation behaviour in expansive soils. Keywords: Civil engineering, Photographic processing, Quantitative methods, Automation, Deep Learning

Actions (login required)