Deep learning-based predictive models of land subsidence and collapsed pipes in Razavi Khorasan Province, Iran
Narges Kariminejad,Adel Sepehr,Mohammad Kazemi Garajeh,Arman Ahmadi,Atoosa Gholamhosseinian
DOI: https://doi.org/10.1007/s12145-024-01349-9
2024-06-14
Earth Science Informatics
Abstract:Land subsidence (LS) and pipe collapse (PC) as the major types of geomorphologic hazards lead to noticeable changes in landscape alterations, land damage, loss of soil and water, surface erosion, and sediment buildup in affected areas. To overcome this, the susceptibility to LS and CP was investigated using three deep learning convolutional neural network (DL-CNN) architectures, including Res-Net, AlexNet, and VGG-Network. We used various predictor variables, and then, trained and tested our DL-CNN models using ReLu, Cross-Entropy, and Adam as activation, loss, and optimization functions, respectively. Our findings showed that DL-CNN models achieved an overall accuracy of 0.9836, 0.9721, and 0.9642 for the Res-Net, AlexNet, and VGG-Network, respectively, for CP sensitivity detection. In addition, the Res-Net, AlexNet, and VGG-Network with an overall accuracy of 0.9698, 0.9654, and 0.9519, respectively, showed satisfying performances for LS detection. We also applied univariate summary statistics, including L(r), the pair correlation function (g(r)), and the O-ring function (O(r)), to investigate the spatial pattern and distribution of CP and LS. The L(r) function graph showed that the spatial patterns of CP and LS were clustered across all the investigated distance scales. The value of this function fell outside the Monte Carlo range, indicating that the accumulation of CP and LS at the mentioned distance scale was statistically significant. The results of the O(r) function for the distribution pattern of CP in the study area indicated that this phenomenon was mostly distributed next to each other, implying the facilitating effect of CP on the creation and expansion of each other across all the investigated distance scales. Similarly, the univariate function g(r) also showed the dispersed distribution of subsidence LS at all distances next to each other. In summary, the results of this research revealed that much of the study area was susceptible to CP and LS. The proposed methodology and findings of this study would be useful for land managers, stakeholders, and researchers.
geosciences, multidisciplinary,computer science, interdisciplinary applications