Table 3 Population model validation performance for (top) manually and automatically extracted features, along with (bottom) test performance of existing maps. See section “Results” for metric definitions. IQR is interquartile range of absolute errors. Bold indicates best in category. FT fine-tuning.
From: Census-independent population estimation using representation learning
Features used | \(R^2\) | \({\textsc {MeAPE}}\) | \({\textsc {MeAE}}\) (%) | \({\textsc {IQR}}\) | \({\textsc {AggPE}}\) (%) |
|---|---|---|---|---|---|
Hand-crafted features | |||||
Public only | − 0.22 | 57.8 | 5.05 | 8.60 | 21.5 |
Footprint only | 0.47 | 44.5 | 3.75 | 5.86 | 02.2 |
Public + Footprint | 0.46 | 48.8 | 4.63 | 5.11 | 07.6 |
Representation learning | |||||
Supervised | 0.20 | 54.7 | 5.36 | 7.97 | 02.0 |
Supervised (FT) | 0.33 | 52.9 | 4.72 | 6.23 | 05.5 |
SWAV | 0.34 | 51.6 | 6.60 | 4.33 | 00.8 |
SWAV (FT) | 0.41 | 46.9 | 5.83 | 4.35 | 03.5 |
DeepCluster | 0.26 | 50.3 | 4.60 | 6.03 | 06.7 |
DeepCluster (FT) | 0.13 | 62.5 | 5.98 | 8.32 | 06.8 |
Barlow Twins | 0.27 | 51.9 | 5.40 | 6.65 | 02.8 |
Barlow Twins (FT) | 0.39 | 44.0 | 3.91 | 6.32 | 01.1 |
Null model | |||||
None | − 0.12 | 76.45 | 7.57 | 10.0 | 01.7 |
Existing maps | |||||
GRID3 | 0.22 | 51.7 | 4.25 | 7.11 | 26.7 |
HRSL | − 0.12 | 70.7 | 5.04 | 7.94 | 46.8 |
WorldPop | − 0.41 | 86.8 | 5.85 | 8.18 | 77.9 |
