Search

Land cover (LC) is an important indicator to reach several of the targets under the Global Goals. Accurate global LC time-series are thus vital to monitor sustainable development. Although the number and quality of open-access, remotely sensed LC products is increasing, all products have uncertainties due to widespread classification errors. However, the relative magnitude of uncertainties among exiting LC products is largely unknown, which hampers their confident selection and robust use for sustainable development evaluation and planning. To close this gap, we quantified region-, time-period-, and coarse-LC class-specific data uncertainties for the 10 most widely used global LC time-series. To this end, we developed a novel multi-scale validation framework that accounts for differences in mapping resolutions and scale mismatches between the spatial extent of map grid cells and validation samples. We aimed for a fair validation assessment by carefully evaluating the quality of our validation samples with respect to landscape heterogeneity that LC products often fail to classify accurately. To address the issue, we supported the validation assessment with Landsat-based measures of cross-scale spectra similarity. The metric was computed by taking advantage of the full Landsat archive in Google Earth Engine. We base our assessment on more than 1.8 million globally integrated LC validation sites, where we mobilized around 2.8 million samples during the period 1980-2020 composed by hundreds of sampling effort of varied nature, from field surveys to crowdsourcing campaigns. Here, we will present the results of the assessment, providing insights on global and regional patterns of LC uncertainties. We found that no single product is more accurate over the others in mapping all LC classes, regions and time-periods. We will provide recommendations on the selection of fit-for-purpose LC time-series, and discuss future strategies for addressing their uncertainties in sustainable development evaluation and planning.