Spatial resolution and temperature matter for mapping Cerrado wetlands and dry grasslands

Wetlands in the Brazilian Cerrado play key roles in regional carbon and water cycles but remain poorly mapped due to their patchy distribution and seasonal variability. Therefore, knowing where and when they occur is urgently needed. To address this gap, we evaluated how spatial resolution and inclusion of thermal (on top of traditional multispectral) data affected wetland vs. dry grassland mapping accuracy using Unoccupied Aerial Vehicle (UAV) imagery. Additionally, we investigated variable importance and how including topography and vegetation patch size as post-processing constraints improved accuracy. We used multispectral and thermal data with resolutions ranging from 0.10 to 1.50 m to train and validate Random Forest models across two seasons. Mapping accuracy increased with pixel size up to 1.0 m, declining at coarser resolutions. Incorporating land surface temperature (LST) significantly improved classification, increasing accuracy by 4.2 to 7.3 percentage points depending on the season. Grassland type classification was primarily driven by the Normalized Difference Vegetation Index (NDVI) and LST, with the latter being especially discriminant in the wet season. Accuracy was further improved by incorporating ancillary data, reaching up to 94% in the wet season. When compared with state-of-the-art land cover maps for Brazil, our drone-based results reveal a wetland extent more than four times larger in the study area than previously reported, underscoring the widespread underestimation of these ecosystems. These findings highlight the value of combining UAV-based multispectral and thermal data for identifying and monitoring Cerrado wetlands, providing essential information to guide conservation efforts in this threatened ecosystem.