@Article{rig.2025.070997,
AUTHOR = {Yan Eng Tan, Siti Nur Aliaa Roslan},
TITLE = {Advancing Sinkhole Susceptibility Mapping in Urbanised Karst Landscapes},
JOURNAL = {Revue Internationale de Géomatique},
VOLUME = {34},
YEAR = {2025},
NUMBER = {1},
PAGES = {777--791},
URL = {http://www.techscience.com/RIG/v34n1/64208},
ISSN = {2116-7060},
ABSTRACT = {Sinkholes, typically associated with karst landscapes, are emerging as significant geohazards in rapidly urbanising regions such as Kuala Lumpur, where human activities like land development, underground infrastructure, and groundwater extraction exacerbate subsurface instability. Despite their destructive potential, sinkholes remain under-monitored in Malaysia due to fragmented data and the lack of predictive spatial tools. This study aimed to develop a web-based, machine learning-driven framework for sinkhole susceptibility mapping to support public awareness, hazard mitigation, and geospatially informed urban planning. The framework was implemented using Google Earth Engine and Google Colab, focusing on Kuala Lumpur and parts of Selangor. Fourteen natural and anthropogenic control factors were derived from remote sensing and government datasets, including topography, lithology, groundwater depth, and proximity to infrastructure. High-and low-susceptibility zones were labelled based on sinkhole inventory and geological stability. Three models, Random Forest (RF), Artificial Neural Network (ANN), and One-Dimensional Convolutional Neural Network (1D CNN) were trained and compared. The RF achieved the highest predictive accuracy but showed signs of overfitting, the ANN produced sharper boundaries between classes, and the 1D CNN, while slightly less accurate, achieved the best ability to distinguish between high-and low-risk areas and generated smoother probability surfaces ideal for visual communication. Constrained by the limited sinkhole inventory and the scarcity of detailed subsurface datasets, this study is positioned as a prototype demonstration of cloud-based machine and deep learning–driven susceptibility mapping rather than a fully generalised model. The 1D CNN model was ultimately deployed in an interactive Google Earth Engine web application featuring toggleable layers and a click function to retrieve local sinkhole case information. This study demonstrates the potential of integrating remote sensing and deep learning for dynamic, interpretable, and publicly accessible urban geohazard mapping in karst-prone areas.},
DOI = {10.32604/rig.2025.070997}
}