Deciphering the many maps of the Xingu, an assessment of deforestation from land cover classifications at multiple scales

Remote sensing is an invaluable tool to objectively illustrate the rapid decline in habitat extents worldwide. The many operational Earth Observation platforms provide options for the generation of land cover maps, each with unique characteristics, as well as considerable semantic differences in the definition of classes. As a result, differences in baseline estimates are inevitable. Here we compare forest cover and surface water estimates over four time periods spanning three decades (1989-2018) for [~]1.3 million km2 encompassing the Xingu river basin, Brazil, from published, freely accessible remotely sensed classifications. While all datasets showed a decrease in forest extent over time, we found a large range in the total area reported by each product for all time periods. The greatest differences ranged from 9% (year 2000) to 17% of the total area (2014-2018 period). We also show the high sensitivity of forest fragmentation metrics (entropy and foreground area density) to data quality and spatial resolution, with cloud cover and sensor artefacts resulting in errors. We further show the importance of choosing surface water datasets carefully because they differ greatly in location and amount of surface water mapped between sources. In several of the datasets illustrating the land cover following operationalization of the Belo Monte dam, the large reservoirs are notably absent. Freshwater ecosystem health is influenced by the land cover surrounding water bodies (e.g. Riparian zones). Understanding differences between the many remotely sensed baselines is fundamentally important to avoid information misuse, and to objectively choose the most appropriate dataset for conservation, taxonomy or policy-making. The differences in forest cover between the datasets examined here are not a failure of the technology, but due to different interpretations of forest and characteristics of the input data (e.g. spatial resolution). Our findings demonstrate the importance of transparency in the generation of remotely sensed datasets and the need for users to familiarize themselves with the characteristics and limitations of each chosen data set.