Shrinkage Crack Detection in Expansive Soil using Deep Convolutional Neural Network and Transfer Learning

Abstract

The formation of shrinkage cracks is a natural phenomenon in expansive soils. The development of these cracks affects both the physical and mechanical properties of the soil. This paper proposes new procedures for predicting and detecting the formation of crack patterns in expansive soils, based on customized Convolution Neural Network (CNN) and transfer learning. A total of four different deep learning models are developed to detect the soil crack pattern by changing the convolution layers and hyper-parameters in the analysis. The novelty of the proposed detection methods lies in the use of customized CNN models in shrinkage crack detection for expansive soils. The customized CNN models are constructed by varying the number of convolution layers and the hyperparameters. The results show that the proposed CNN models provide very accurate results and are capable of detecting the presence of cracks in the soil with great accuracy. The best results are from one of the customized CNN models namely the Customized CNN Model 2 which consists of five convolution layers, three activation layers, one pooling layer, two fully connected layers, and a softmax layer. The results from this model are compared with other well-known approaches from the literature and are shown to provide improved results. Overall, the proposed deep learning methods developed in this paper produce excellent results in terms of the accurate detection of shrinkage soil cracks and can also be applied to other types of soil cracks.