Determination of burned areas using Sentinel-2A imagery and machine learning classification algorithms

Abstract

We aimed to determine the spatial extent of burned areas using remote sensing (RS) data and machine learning methods. It is often difficult, time-consuming and costly to collect in-situ data after fires; therefore, RS is used in determining burnt regions. We selected the Manavgat district of Antalya province as the study area due to the major forest fires occurred in 2021. We used pre-post Sentinel 2A images due to the ability of Sentinel in burned area mapping, fire density and damage determination, and being openly available. Then we implemented indices to determine the changes caused by fires. The indices are Normalized Burned Ratio (NBR), Normalized Vegetation Index (NDVI), Relative differenced Normalized Burn Ratio (RdNBR), Relativized Burn Ratio (RBR), Burned Area Index (BAI), and Modified Soil Adjusted Vegetation Index (MSAVI), Soil Adjusted Vegetation Index (SAVI). Afterwards, we utilized Random Forest (RF) Algorithm, Support Vector Machine (SVM), and Classification Regression Tree (CART) for the Machine Learning (ML) classification. We used the Google Earth Engine (GEE) platform to obtain satellite images and apply indices and ML based classification. Results illustrated that, RF was the most accurate algorithm with 98.57% overall accuracy and SVM has the lowest overall accuracy with 86.19% for the region.