University of Guilan Caspian Journal of Environmental Sciences 1735-3033 23 2 2025 04 01 Uncovering the hidden patterns of fire risks: A cluster analysis approach (K-Medoids and FCM) for Hyrcanian Forest in Iran 481 493 7977 10.22124/cjes.2024.7977 EN Shaghayegh Zolghadri Department of Forestry, Faculty of Natural Resources, University of Guilan, Iran Mehrdad Ghodskhah Daryaei Department of Forestry, Faculty of Natural Resources, University of Guilan, Iran Kamran Nasirahmadi Assistant Professor, Faculty of Chemistry, University of Science and Technology, Mazandaran Esmaeil Ghajar Department of Forestry, Faculty of Natural Resources, University of Guilan, Iran Journal Article 2024 08 31 Forest fires have become a significant environmental concern in the Hyrcanian Forest, causing extensive loss of vegetation and posing a threat to biodiversity. Accurate prediction of high-risk fire locations is crucial for effective forest management. In this study, we developed and evaluated a clustering-based model using a multilayer perceptron artificial neural network with an error backpropagation training procedure to model fire risk potential in Saravan Forest, Guilan Province, North Iran. To optimize generalization, the model utilized two unsupervised clustering-specific procedures, namely Fuzzy C-Means and k-Medoids. The primary focus of our study was on the model's ability to predict potential fire risk locations, which is essential for forest fire prevention and control. The input criteria included recorded fire incidents, distances to farmland, roads, rivers, air pressure, solar radiation, slope, aspect, wind speed, and percentage of canopy cover density. The results showed that the procedure of the two algorithms used in this study in allocating potential fire hazard points is highly similar, differing mainly in the methodology employed for data center allocation. According to the results, the RMSE, R², and MSE for the model used in this study are respectively equal to 0.2861, 99.38, and 0.01919, which indicates the reliability of the model. Moreover, according to the Confusion matrix analysis table's results, FCM was slightly better than K-medoids in terms of its predictive accuracy. This model demonstrated high accuracy in predicting fire hazards, showing promising potential for forest fire prediction using clustering-based models. Additionally, our model exhibited superior performance compared to other clustering techniques for identifying potential fire hazard sites. Our developed clustering-based model provides valuable insights for forest managers to identify locations at fire risk, enabling more efficient resource allocation and preventative measures. This approach can significantly improve forest fire management and reduce ecological damage caused by wildfires. https://cjes.guilan.ac.ir/article_7977_aa0ac1b21d0f849b4091619980af1349.pdf