This study is at the forefront of integrating thermal imaging technology for the advanced detection of subterranean anomalies that pose potential risks. Utilizing the precision of infrared thermography, the research aims to delineate and identify these geological hazards. At the core of the analysis is a sophisticated convolutional neural network (CNN), which leverages deep learning algorithms to interpret complex thermal data. The methodology involves the deliberate creation of simulated geological hazards at varying depths, employing a variety of materials including air, water and soil. An advanced infrared camera is deployed to meticulously record the thermal variances over diurnal cycles, capturing the nuanced patterns of heat absorption and dissipation. Subsequently, the CNN undergoes rigorous training to classify the thermal images, considering the dimensions, stratification, and material composition of the subsurface features. The study achieves an impressive benchmark in classification accuracy, peaking at 71.3% in scenarios involving the detection of hazardous geological formations.
Sinkhole is the one of the highly impacting problem among natural hazard to human risk. To evaluate the condition of the road subsurface where sinkholes have occurred, Ground Penetrating Radar (GPR) is frequently utilized. GPR offers several advantages, including high utility, straightforward experimental methods, and the convenience of not necessitating additional procedures for investigation. Research on sinkholes, utilizing the practicality of GPR, can be summarized as follows: Batayneh et al. (2002) used GPR to identify areas prone to ground subsidence. Kruse et al. (2006) analyzed the characteristics of terrain where sinkholes occurred using GPR. Carbonel et al. (2014) studied the features of sinkhole-prone regions and aimed to understand the mechanisms behind sinkhole formation. Nadasi & Szabo (2023) detected sinkholes occurring at depths of 3 to 4 meters using GPR and created maps of sinkhole-prone areas by comparing their findings with neighboring regions. However, it's important to note that interpreting subsidence areas using GPR requires significant expertise and has limitations, such as reduced signal penetration depending on subsurface fluid content. To address these challenges and enhance sinkhole research, some studies have explored the use of infrared technology. Cho et al. (2016) analyzed the mechanism of temperature differences in subbase joints following joint construction. Hoai et al. (2019) conducted research on sinkhole detection using CNN algorithms, while Kalhor et al. (2021) examined temperature changes by creating various types of joints. Yavariabdi et al. (2023) trained CNN algorithms on satellite images for sinkhole detection and classification. The use of infrared techniques is essential to overcome these challenges. To this end, this research focuses on machine learning-based image analysis methods. This study aims to improve analysis reliability by applying a convolutional neural network (CNN) algorithm specialized in analyzing images captured with infrared technology to increase reliability for classification.
