Assessment of Burned Area Products During the Severe Fire Season in the Xingu River Basin, Brazilian Amazon

This study evaluated the performance of four burned area products (AQ1Km, FireCCIS311, MCD64A1, and VNP64A1) using the reference datasets MapBiomas Fire and the Global Annual Burned Area Map (GABAM) for 2024. The two 30 m reference datasets are generated using different automated classification approaches, including deep learning in MapBiomas Fire and Random Forest in GABAM, allowing the sensitivity of product evaluation to alternative machine-learning-based reference frameworks to be examined. The Xingu River basin is a heterogeneous tropical landscape characterized by extensive forest ecosystems, protected and Indigenous territories, agricultural frontiers, and recurrent fire activity. Accuracy metrics derived from confusion matrices (contingency table metrics) were combined with analyses of burned scar geometry, burned patch size, land use and land cover, low-resolution grid regression, and Kernel Density Estimation (KDE) based on active fire hotspots. Results showed substantial differences among products in their ability to represent burned scars under tropical conditions. FireCCIS311 achieved the best overall performance, presenting the highest Intersection over Union (IoU) values (52.67% against MapBiomas Fire), better geometric coherence, and stronger spatial regression consistency (r2 = 0.88). AQ1Km showed higher sensitivity and a stronger association with active fire hotspots, but also greater commission errors (CE) and spatial overestimation. MCD64A1 and VNP64A1 presented intermediate performance, with higher omission rates, particularly in forested and fragmented areas. All products showed low accuracy for small, burned patches, while performance improved for larger and more continuous scars. The results also highlighted the strong influence of the reference dataset on validation outcomes, with markedly better agreement observed with MapBiomas Fire than with GABAM. These results show that product uncertainty can alter the spatial evidence used to characterize wildfire impacts. Regionally validated and complementary burned area products can therefore provide a more reliable basis for forest fire monitoring, post-fire assessment, and operational fire management in heterogeneous tropical landscapes.