On the Fine-tuned VGG16 for Poverty Estimation using Satellite Images

Poverty can be defined as the inability of a person or a household to satisfy their primary needs. Most countries use household consumption as a standard measurement of population wealth. However, a comprehensive door-to-door data collection such as a survey is time-consuming, costly, and vulnerable to human error. Passively collected data such as nighttime and daytime satellite images can be used to estimate poverty indicators. This research is aimed to estimate financial-based poverty (per capita expenditure) of 40 regencies in Central Java and the Special Region of Yogyakarta in 2019. The fine-tuned VGG16 is proposed to classify the night light luminosity. We add image augmentation to enrich the training images and the proposed architecture can classify daytime images with 0.767 accuracy. We extract daytime images into features using base VGG16 and fine-tuned VGG16. The extracted features are used to train multiple machine learning models with repeated 10-fold cross-validation. Regression results using these two extracted features are compared. Nonparametric mean test suggests no difference between these two features. The combination of fine-tuned VGG16 and Random Forest generates the best model with median RMSE of 1393.599. The successful work of this approach is used to monitor poverty on a monthly basis using daytime satellite imagery data.