Burned Forest Area Mapping From Post-Fire Sentinel-2 Imagery Using Object-Based Machine Learning Classification

Abstract

Forest fires cause serious damage not only to the ecosystem in the forest but also to social and economic life. Rapid detection of burned areas with remote sensing methods is important both to determine the current damage and to evaluate the economic and ecological losses caused by the fire and to create rapid response plans. This study presents an approach to identify and map burned forest areas using an object-based random forest (RF) machine learning (ML) classification method using only post-fire Sentinel-2 imagery on the Google Earth Engine (GEE) platform. In addition to original spectral bands of Sentinel-2 (B2, B3, B4, B8, B11, B12), mid-infrared burn index (MIRBI), normalized burn ratio 2 (NBR2), burn area index (BAI) and normalized difference vegetation index (NDVI) bands were calculated and included as additional bands in the Sentinel-2 image. Prior to object-based classification, image segmentation was carried out using the Simple Non-Iterative Clustering (SNIC) algorithm. Training samples were selected on the GEE platform and object-based classification with the RF algorithm was applied to four study areas (Marmaris – MR, Kavaklıdere – KV, Manavgat – MG, Çanakkale - CK) in Türkiye where forest fires have occurred in recent years. The results showed high performance with an overall accuracy of 93.5% in MR, 97.7% in CV, 94.8% in MG and 96.5% in CK with the object-based RF classifier. In addition, the spatial and temporal transferability of the object-based RF algorithm was evaluated based on two study areas (MG and CK) and the RF model transferability provided an overall accuracy of 87.5% in MR, 94.8% in CV, 93.6% in MG and 96.8% in CK. The results show that burned forest areas can be successfully detected by object-based classification method using cloud-based GEE platform from Sentinel-2 images with a uni-temporal post-fire imagery approach and the potential of developing a transferable object-based classification model for mapping burned forest areas.