Burn Severity Assessment with Different Remote Sensing Products for Wildfire Damage Analysis

In recent decades, rapidly increasing forest fires have become a significant threat to the forest environment and rural communities. The average annual land affected by wildfires from 1997 to 2018 reached 10350 ha in Türkiye. In order to mobilize forestry protection and post-wildfire recovery plans, earth observation satellites have become the key component due to their wide range of data and vision capacity. In this study, a classification-based burn severity assessment was planned created on single post-wildfire satellite images from the Southern Mediterranean Region of Türkiye which has a quite complicated topography. The classification algorithm was trained to classify images into four classes: unburned forest area, low severe burned forest area, moderate severe burned forest area and high severe burned forest area. The classification results compared with differenced Normalized Burn Ratio (dNBR). Various remote sensing products were taken into consideration during generating the methodology. For minimizing fieldwork and understanding the study area characteristics, aerial photos of 0.25 m spatial resolution were analyzed and used for train/test points collection; 11519 train and 400 test points have been selected. Sentinel-2 were used as input data. Classification algorithm selected as Random Forest. Overall accuracy, kappa coefficient, precision, recall and F-score parameters have been calculated for accuracy assessment. As a result, F-scores of 0.9, 0.77, 0.71 and 0.85 were obtained from Sentinel-2 for unburned forest area, low severe burned forest area, moderate severe burned forest area and high severe burned forest area, respectively. Corresponding F-scores of 0.85, 0.47, 0.63 and 0.76 were calculated from the dNBR.