Extended Abstract
Introduction
Natural hazards are hazards whose development and formation depend on natural events and humans can play a stimulating role in their occurrence. However, the characteristics of the natural environment are considered to be the basis for the formation and development of hazards. Earthquakes are among the most catastrophic natural hazards, causing widespread human casualties, injuries, and significant economic losses. Kermanshah Province is recognized as a high-seismic-risk region due to its location within the folded Zagros tectonic zone, the presence of several active faults, and a history of strong seismic events. Therefore, the identification and zoning of earthquake-prone areas, as well as seismic hazard analysis, are essential for effective crisis management, sustainable urban and rural development, reducing infrastructure vulnerability, and improving land-use planning strategies. The present study aims to develop a seismic hazard zonation map for Kermanshah Province using advanced machine learning techniques, specifically Random Forest (RF) and Support Vector Machine (SVM), and to evaluate and compare the predictive performance of these two models. The key innovation of this research lies in the integration of geological, seismological, and morphometric datasets within a Geographic Information System (GIS) framework, enabling the development of a data-driven approach for more accurate and reliable seismic hazard assessment.
Materials & Methods, Results
Initially, based on a comprehensive review of the literature, ten conditioning factors influencing seismicity were selected, including distance to faults, fault density, lithology, tectonic setting, slope, elevation, and selected morphometric indices. The corresponding thematic layers were prepared within a Geographic Information System (GIS) environment and transformed into the UTM Zone 38 coordinate system to ensure spatial consistency. Subsequently, 207 earthquake occurrence locations were extracted from existing seismic databases, and an equal number of non-occurrence (control) points were generated to represent the absence of seismic events. These presence–absence data, together with the conditioning factor layers, were integrated in a Python environment for implementing the Random Forest (RF) and Support Vector Machine (SVM) models.The dataset was divided into three subsets using a stratified sampling approach: 70% for training, 15% for validation, and 15% for testing. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) as the primary accuracy metric. In addition, feature importance analysis was conducted, and spatial aggregation of hazard values at the city level was performed using the Zonal Statistics tool to produce average seismic hazard maps for each administrative unit.
Discussion
The modeling results demonstrated that both the Random Forest (RF) and Support Vector Machine (SVM) models performed strongly during the training phase, each achieving an AUC value of approximately 0.959, indicating a high capability in distinguishing seismically susceptible areas. However, notable differences emerged during the testing phase. The RF model achieved an AUC of 0.906, demonstrating significantly higher predictive accuracy compared to the SVM model, which yielded an AUC of 0.823. This performance gap highlights the superior ability of the RF algorithm to capture complex and nonlinear relationships between tectonic and morphometric variables, its greater robustness against overfitting, and its improved generalization capability when applied to unseen data. The resulting seismic hazard zoning maps indicate that areas of high and very high hazard are predominantly concentrated near major active faults, in mountainous regions characterized by active folding, and within parts of the High Zagros zone. In contrast, low-hazard zones are generally located in plains and regions situated farther from active fault systems. The consistency of these spatial patterns with the geological and seismic characteristics of the region further confirms the reliability of the results. Feature importance analysis revealed that the most influential variable was the depth of past earthquakes, with an importance value of approximately 0.26. This was followed by fault density (0.24), which reflects the degree of tectonic deformation and the presence of active fault networks. The distance to faults was identified as the next most important factor, with an estimated contribution of 0.17, underscoring its critical role in controlling seismic hazard distribution. Furthermore, zonal statistical analysis indicated that counties such as Sarpol-e-Zahab, Paveh, and Ravansar exhibit significantly higher average seismic risk compared to other regions. Overall, the findings underscore the effectiveness of the RF algorithm in delineating seismically sensitive zones and producing more reliable hazard maps than the SVM model.
Conclusion
The results of the study show that the RF model provides better separation between medium, high, and very high risk classes compared to SVM and provides a clearer picture of priority zones for crisis management and land use planning by reducing the classification error rate. Finally, it can be said that the present study, by utilizing a combination of GIS and two powerful algorithms, RF and SVM, while confirming the general results of previous studies on the importance of geological and seismological factors in the formation of high-risk zones, has taken a step further and shown that the random forest model is a more accurate and reliable tool for earthquake risk zoning in the regional conditions of Kermanshah province. This result can be an important basis for urban and rural development planning, locating critical infrastructure, and formulating risk reduction and crisis management policies in one of the most seismic provinces in the country.