Comparative Analysis of Deep Learning and Machine Learning Models for Burned Area Estimation Using Sentinel-2 Image: A Case Study in Muğla-Bodrum, Turkey

As one of the main natural disasters affecting the natural cycle of vegetation, wildfires cause deterioration of natural ecosystems and the destruction of forests. Every year, a large quantity of forest is destroyed by wildfires all over the world. Therefore, precise and timely information about the extent of the burned areas is critical for monitoring wildfire damage. The use of remote sensing and deep learning methods offers great benefits to improve the reliability and efficiency of burned area detection. The main purpose of this paper is to demonstrate the efficacy and usability of a deep learning model (i.e., Convolutional Neural Networks-CNNs) for burned area mapping and compare its predictive performance with a novel machine learning algorithm (i.e., Extreme Gradient Boosting-XGBoost) using various state-of-the-art accuracy metrics. In this paper, the Mugla-Bodrum fire that occurred in July 2021 situated on the Mediterranean coast of Turkey was evaluated. To investigate the effects of the spectral indices (i.e., normalized burn index-NBR, differenced normalized burn index-dNBR, and normalized difference vegetation index-NDVI), two datasets consisting of different combinations (i.e., Sentinel-2 original spectral bands and spectral bands together with spectral indices) were employed in analyses. The findings of burned area delineation indicated that the CNN model produced 97.53% overall accuracy whilst the XGBoost algorithm achieved an overall accuracy of 93.14% with the use of spectral indices as ancillary data. A similar performance difference was also observed when the predictions were analyzed with the area under the curve metric.