Efficiency of convolutional neural networks (CNN) based image classification for monitoring construction related activities: A case study on aggregate mining for concrete production
Seda Yeşilmen, Bahadır Tatar
Abstract
Monitoring construction activities is an important task for efficiency in construction site operations thus the topic received a fair amount of attention in the literature. Optimizing construction site operations by monitoring and detecting various tasks is dependent on the size of the construction field, which determines the tools that can be used for the job. A monitoring task can be performed with high efficiency through image classification algorithms by training the algorithms to detect construction machinery. If the area of monitoring is larger, such as the task of detecting construction related operations in a large infrastructural construction, using drone images might become inefficient. We aimed to take a first step towards a cost-efficient monitoring system specifically for construction activities that cover large territories. Consequently, satellite image classification has been performed for construction machinery detection in this study. We utilized different versions of well-established convolutional neural network architectures as backbone for the transfer learning method and their performances are evaluated. Finally, the best performing models are determined as DenseNet161 and ResNet101 with 0.919 and 0.903 test accuracies, respectively. DenseNet161 model was discussed in terms of its accuracy and efficiency in a case study to detect illegal aggregate mining activity through the basin of Thamirabarani River.