USGS FIREMON ve Akdeniz’e Özgü dNBR’leri Kullanarak Orman Yanma Şiddeti İçin Piksel ve Nesne Tabanlı Topluluk Algoritmaların Kullanımı
Taşkın Kavzoğlu, Alihan Teke, Elif Özlem Yılmaz · 2025
"Forest fires in the Mediterranean basin are a major threat to the sustainability of natural ecosystems and the protection of biodiversity. Forest fires affect not only the vegetation, but also the fauna and micro-organisms living in the ecosystem. Fires leading to the extinction of rare and endemic species accelerate the loss of biodiversity and disrupt the balance of ecosystems. This situation creates the need to effectively monitor and analyze the destruction caused by forest fires. Remote sensing technologies and satellite imagery play a critical role in assessing the severity and impact of forest fires. While pixel-based analysis evaluates the spectral characteristics of each pixel, object-based analysis considers homogeneous groups of regions in a broader context. In this study, these two remote sensing approaches were applied to Sentinel-2 satellite imagery for the Muğla fires of 12 and 14 July 2023. Using the pre- and post-fire imagery, burned area and severity analyses were performed using the latest ensemble-based machine learning algorithms: random forest, XGBoost, GBM and NGBoost. For pixel-based classification, the highest accuracy was obtained with the XGBoost algorithm with 90.95% On the other hand, the highest accuracy in object-based classification was calculated with the GBM algorithm with 85.94% The fire intensity analysis was modeled using the USGS FIREMON and the Mediterranean-based Differential Normalized Fire Intensity (dNBR) index values. The SHAP approach, one of the global explainable artificial intelligence approaches, was used to understand the decisionmaking mechanisms of the trained machine learning models and to determine the effectiveness of each feature/index within the model. According to SHAP results, spectral indices of BAI and NDVI in pixel-based classification and RVI in object-based classification were determined as the most effective factors in model predictions. The study highlights the importance of innovative methods for monitoring and managing forest fires, and points to the need for new strategies to conserve biodiversity."