Forest Fire Susceptibility Analysis with Remote Sensing Data and Machine Learning Algorithms using Region-based Datasets
POLISH JOURNAL OF ENVIRONMENTAL STUDIES, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.15244/pjoes/217446
- Dergi Adı: POLISH JOURNAL OF ENVIRONMENTAL STUDIES
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, BIOSIS, CAB Abstracts, Central & Eastern European Academic Source (CEEAS), Environment Index, Greenfile, Public Affairs Index, Academic Search Ultimate (EBSCO), Biomedical Reference Collection: Corporate Edition (EBSCO)
- Çanakkale Onsekiz Mart Üniversitesi Adresli: Evet
Özet
Forests are vital elements of terrestrial ecosystems and ensure the integrity and sustainability of natural factors such as soil, water, and climate. Forest fires are one of the disasters that cause deterioration of the ecosystem as well as social and economic impacts. In combating disasters, determining pre-disaster risks and reducing disaster damage is the first stage of disaster management. In this study, a dataset consisting of 679 & times;14 rows and columns derived from 312 forest fire points between 2017 and 2022 was prepared. For the dataset, 13 independent variables were mapped with remote sensing and geographic information systems techniques from satellite images and open-source data. Training and prediction datasets were created by extracting values for all variables in each pixel. Feature relevance was initially assessed using Mutual Information (MI), followed by model-specific interpretation using SHAP values. Model performance was evaluated using confusion matrices and ROC-AUC analysis. Machine learning algorithms, including Decision Trees, Multi-Layer Perceptron (MLP), Naive Bayes, Random Forest, and XGBoost, were trained using a dataset split into 80% training and 20% testing. At the end of the study, forest fire sensitivity maps were created with an accuracy rate of 96% using the Random Forest and XGBoost algorithms, which were the most powerful models. Susceptibility probabilities were rescaled to a percentage scale and classified into low, medium, and high categories using a quantile-based approach. Results indicate that distance to roads and population density are the most influential predictors, highlighting the dominant role of human activity in wildfire ignition.