Table 2 Summary of recent literature on deep learning driven census-dependent intercensal population estimation. NOTE: Input Resolution indicates the spatial resolution of imagery input to model (i.e. after reprojection and resizing). %RMSE is percent root mean squared error, MAPE is mean absolute percent error.
From: Census-independent population estimation using representation learning
ROI | Input shape | Input resolution | Output resolution | Input data cost | Performance | |||
|---|---|---|---|---|---|---|---|---|
Validation | %RMSE | MAPE | \(R^2\) | |||||
Tanzania23 | \(32\times 32\times 8\) | 250 m (Landsat) | 8 km | Free | Train/test split | 51.5 | – | – |
USA1 | \(74\times 74\times 7\) | 15 m (Landsat) | 1 km | Free | Future census | – | 49.8 | 0.94 |
India24 | \(224\times 224\times 3\) | 20 m (Landsat) 20 m (Sentinel-1 radar) | 4.5 km (village level) | Free | Train/test split | 24.3 | 21.5 | 0.93 |
