The trees in agroforestry plots create spatial heterogeneity of high interest for adaptation, mitigation, and the provision of ecosystem services. But to what distance, exactly, from the tree? We tested a novel approach, based upon geostatistics and Unmanned Aerial Vehicle (UAV) sensing, to infer the distance at which a single agroforestry tree affects the surrounding under-crop, to map yield, litter (i.e. stover) and compute crop-partial Land Equivalent Ratio (LERcp) at the whole-plot level. In an agro-silvo-pastoral parkland of semi-arid western Africa dominated by the multi-purpose tree Faidherbia albida, we harvested the pearl-millet under-crop at the whole-plot scale (ca. 1¿ha) and also in subplot transects, at three distances from the trunks. We observed that the yield was three times higher below the tree crown (135.6¿g m-2) than at a distance of five tree-crown radii from the trunk (47.7¿g m-2). Through geostatistical analysis of multi-spectral, centimetric-resolution images obtained from an UAV overflight of the entire plot, we determined that the 'Range' parameter of the semi-variogram (assumed to be the distance of influence of the trees on the Normalized difference vegetation index (NDVI)) was 17¿m. We correlated the yield (r2¿=¿0.41; RRMSE¿=¿48 %) and litter production (r2¿=¿0.46; RRMSE¿=¿35 %) in subplots with NDVI, and generated yield and litter maps at the whole-plot scale. The measured whole-plot yield (0.73¿t ha-1) differed from the one estimated via the UAV mapping by only 20 %, thereby validating the overall approach. The litter was estimated similarly at 1.05 tC ha-1 yr-1 and mapped. Using a geostatistical proxy for the sole crop, LERcp was estimated 1.16, despite the low tree density. This new method to handle heterogeneity in agroforestry systems is a first application. We also propose strategies for extension to the landscape level.