Global agroforestry maps remain uncertain, partly because agricultural land masks used to identify farmland trees have never been systematically validated against ground-truth data. Existing approaches rely on coarse-resolution land-use products and discrete classifications that poorly capture the continuum of tree–crop integration, directly affecting large-scale estimates of climate mitigation and ecosystem services provided by agroforestry. Here, we present the first systematic evaluation of how different land-use masks affect agroforestry mapping accuracy, combining agricultural land masks derived from three 10-meter land-use/land-cover products with a global sub-meter resolution tree canopy map. Validating each land mask-tree canopy combination against 14,162 georeferenced ground-truth observations that distinguish agroforestry from non-agroforestry trees, we find that the choice of agricultural land mask alone drives a 45-percentage-point shift in accuracy. The ESRI-based mask achieves 77.2 % accuracy, outperforming ESA WorldCover (69 %), Dynamic World (68 %), and the Global Pasture Watch (GPW)- based map (32 %) across all biomes tested. Applying the best-performing combination, we estimate that 86 million hectares of agricultural land in Europe and 36 million hectares in the Sahel host at least 1 % tree cover, with sparse-canopy systems (<10 % tree cover) dominating both regions (32 Mha and 21 Mha, respectively). These estimates diverge sharply from the most cited agroforestry maps, which overestimate agroforestry extent by up to 186 % in Europe and underestimate it by 80 % in the Sahel. Together, these findings demonstrate that both land-use mask selection and tree-detection resolution are major, previously unquantified sources of uncertainty in global agroforestry assessments and call for a shift away from coarse-resolution gridded data toward tree-centric, gradient-based approaches capable of resolving individual trees across the agriculture–forest continuum.