Advances in forest carbon mapping have the potential to greatly reduce uncertainties in the global carbon budget and to facilitate effective emissions mitigation strategies such as REDD+ (Reducing Emissions from Deforestation and Forest Degradation). Though broad-scale mapping is based primarily on remote sensing data, the accuracy of resulting forest carbon stock estimates depends critically on the quality of field measurements and calibration procedures. The mismatch in spatial scales between field inventory plots and larger pixels of current and planned remote sensing products for forest biomass mapping is of particular concern, as it has the potential to introduce errors, especially if forest biomass shows strong local spatial variation. Here, we used 30 large (8-50 ha) globally distributed permanent forest plots to quantify the spatial variability in aboveground biomass density (AGBD in Mg ha-1) at spatial scales ranging from 5 to 250 m (0.025-6.25 ha), and to evaluate the implications of this variability for calibrating remote sensing products using simulated remote sensing footprints. We found that local spatial variability in AGBD is large for standard plot sizes, averaging 46.3% for replicate 0.1 ha subplots within a single large plot, and 16.6% for 1 ha subplots. AGBD showed weak spatial autocorrelation at distances of 20-400 m, with autocorrelation higher in sites with higher topographic variability and statistically significant in half of the sites. We further show that when field calibration plots are smaller than the remote sensing pixels, the high local spatial variability in AGBD leads to a substantial "dilution" bias in calibration parameters, a bias that cannot be removed with standard statistical methods. Our results suggest that topography should be explicitly accounted for in future sampling strategies and that much care must be taken in designing calibration schemes if remote sensing of forest carbon is to achieve its promise.