Classificação Supervisionada de Áreas Queimadas do Cerrado Utilizando Atributos Espectrais Provenientes de Séries Temporais do Sensor WFI

The Brazilian Cerrado evolved under natural presence of fires, but human-driven wildfires threaten the conservation of the ecosystems of this biome, as they are more intense, diffuse, and frequent. Consequently, the monitoring of fire events is an important instrument for environmental management. Exist two main approaches of wildfire studies considering remote sensing technologies and applications: the detection of fire foci (or active fires) and the mapping of burned areas. The objective of this article is to analyze the feasibility of using imagery time series from the WFI (Wide Field Imaging Camera) sensor on board the CBERS-4, CBERS-4A, and AMAZONIA-1 satellites for mapping burned areas observed in 2020, 2021 and 2022 in the Chapada dos Veadeiros National Park, by using Random Forest for supervised classifications. The time series comprises 235 images integrated into a regular grid, resulting in six independent datasets (NIR, BAI, EVI, GEMI, NDVI, and NDWI) utilized for training, validation, and test. Overall, we observed that NDVI yielded the lowest values of the accuracy metrics adopted and that when assessing the performance of the annual datasets, 2021 delivered the best results, followed by 2020 and 2022. Generalization tests conducted on annual datasets applying multitemporal models, i.e., models containing samples from all three years, produced IoUs above 70% in 2020 (excluding EVI and NDVI) and 2021 (excluding NDVI), and above 60% in 2022 (excluding NDVI).