a Training data (left maps) and predictions using a single featurization of daytime imagery (right maps). Insets (far right) marked by black squares in global maps. Training sample is a uniform random sampling of 1,000,000 land grid cells, 498,063 for which imagery were available and could be matched to task labels. Out-of-sample predictions are constructed using five-fold cross-validation. For display purposes only, maps depict ~50 km × 50km average values (ground truth and predictions at ~1 km × 1 km). b Test-set performance in the US shown for 12 variables from the 2015 American Community Survey (ACS) conducted by the US Census Bureau31. Income per household (HH) (in purple) is also shown in Figs. 2 and 3, and was selected as an outcome for the analysis in those figures before this ACS experiment was run. c Both labels and features in MOSAIKS are linear combinations of sub-image ground-level conditions, allowing optimized regression weights to be applied to imagery of any spatial extent (Supplementary Note 2.9). MOSAIKS thus achieves label super-resolution by generating skillful estimates at spatial resolutions finer than the labels used for training. Shown are example label super-resolution estimates at 2 × 2, 4 × 4, 8 × 8, and 16 × 16, along with the original 1 × 1 label resolution (See Supplementary Fig. 12 for additional examples). Systematic evaluation of within-image R2 across the entire sample is reported in Supplementary Note 2.9 for the forest cover task.