Soil Condition Classification Based on Natural Water Content Using Computer Vision Technique

Natural water content affects many geotechnical parameters and geological properties of soils, which can reduce cohesion and friction, leading to potential failures in structures such as foundations, retaining walls, and slopes. Identification of the water content helps in designing effective drainage and water management systems to prevent flooding and erosion. In tunnel engineering, soil water content plays an important role as the stability of the tunnel face depends on it. This research solves the problem of classifying soil images depending on the natural water content by computer vision technology. First, laboratory soil tests were carried out, and the relationship between the amount of torque on the screw conveyor and the moisture content of the soil was established; photographs of the soil at different conditions were taken at each step of the experiment. Second, the resulting dataset after preprocessing was processed by convolutional neural network algorithms during deep learning; the transfer learning technique was used to obtain better results. As a result, seven algorithms were obtained that allow classifying the soil images, which can later be used to optimize the tunnel construction process. The best classification ability is demonstrated by the algorithm based on the DenseNet architecture (accuracy 0.9302 and loss 0.1980). The proposed model surpasses traditional approaches due to its increased automation and processing speed. Laboratory tests can be carried out only once for one type of soil in order to determine the boundaries of water content for classes labeling, after which only a cheap camera is required from the equipment to transmit new images for processing by the algorithm.