Mono-temporal and multi-temporal approaches for burnt area detection using Sentinel-2 satellite imagery (a case study of Rokan Hilir Regency, Indonesia)

Accurate and up-to-date information about the burnt area is important in estimating environmental losses, prioritizing rehabilitation areas, and determining future planning strategies. The publicly available medium resolution optical Sentinel-2 satellite data provides a practical and effective solution for burnt area detection. In this study, we proposed two different approaches using mono-temporal and multi-temporal Sentinel-2 satellite imagery to detect burnt areas in Rokan Hilir Regency, Indonesia. The multi-temporal approaches utilized two different ensemble machine learning algorithms (Random Forest and XGBoost) and used six composite spectral indices of the differenced Normalized Burn Ratio (dNBR), differenced Normalized Burn Ratio 2 (dNBR2), differenced Normalized Difference Vegetation Index (dNDVI), differenced Soil Adjusted Vegetation Index (dSAVI), differenced Char Soil Index (dCSI), differenced Burnt area Index for Sentinel-2 (dBAIS2), and differenced Mid-infrared Burn Index (dMIRBI) as model inputs. The burnt areas are labeled by combining hotspots with confidence intervals above 95%, fire spots, and change detection methods. The XGBoost model achieved the best performance with an F1 score of 0.97 and an accuracy of 96%. Furthermore, we use the SHapley Additive exPlanations (SHAP) to quantify the contribution of each feature as well as its correlation with the target class. The dNBR, dMIRBI, and dNBR2 indices contribute the most to the XGBoost model. In comparison, this study also investigates and compares a mono-temporal approach with One-dimensional Convolutional Neural Network (CNN-1D) architecture and the performance obtained is slightly better than both machine learning models. Overall, both mono-temporal and multi-temporal approaches satisfactorily detect the burnt area.