AbdelkaderYS/SAE-Multisource-Niger

Multisource SAE for SDG Monitoring in Niger - DHS 2012 + GPS # Poverty Estimation in Niger with Satellite Imagery and Small Area Estimation > ### *Predict poverty across Niger's 64 departments from free satellite imagery, without costly household surveys!* > ### *Cibler les zones de pauvreté au Niger avec l'IA sur des images satellite, une alternative rapide et gratuite aux enquêtes ménages!* Combine DHS 2012 household surveys with Landsat 7 satellite imagery and a ResNet-18 neural network, then apply a Fay-Herriot small area model to estimate poverty at the department level (Admin 2, 64 departments) in Niger. **Best model** : FH-2 (urbanization + NDVI + CNN score) achieves a coefficient of variation (CV) of 11.9 percent. --- ## Key Results | Model | Predictors | AIC | CV (%) | gamma | | -------------- | ---------------------------------------- | ---------------- | -------------- | --------------- | | FH-0 | None (mean only) | 1532.5 | 93.7 | 0.916 | | FH-1 | Urbanization rate | 1414.0 | 17.8 | 0.709 | | **FH-2** | Urbanization + NDVI +**CNN score** | **1325.9** | **11.9** | **0.455** | - **CNN test R squared**: 0.69 (ResNet-18, 380 train / 96 test clusters) - **CNN test RMSE**: 49,904 (wealth index ranges from -70,000 to +465,000) - **CV threshold**: 11.9 percent is below 15 percent, meeting the EUROSTAT standard for reliable estimates - **Gamma**: 0.455 means the model contributes 54.5 percent of the weight, while the direct survey contributes 45.5 percent. This is a healthy balance for small area estimation. The CNN alone explains 69 percent of the wealth variance between clusters using only Landsat 7 pixels at 30 meter resolution (RGB bands). Adding these predictions to the Fay-Herriot model reduces the CV from 17.8 percent (urbanization alone) to 11.9 percent. --- ## How the ResNet Works ### Why ResNet-18 We use …