Per capita expenditure prediction using model stacking based on satellite imagery

Abstract

One of the indicators for measuring poverty is per capita expenditure. However, collecting timely and reliable per capita expenditure data is quite challenging and expensive, as it requires collecting detailed household data directly. One way to deal with this issue is to use satellite image data processed by machine learning methods. This research proposes a method to predict the per capita expenditure of regencies or cities in Indonesia based on satellite imagery using machine learning techniques, such as k-nearest neighbors (KNN), random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM). The predictions are stacked to predict per capita expenditure using least absolute shrinkage and selection operator (LASSO) regression as the meta-learner. The model is trained on Google-Earth-based satellite imagery of Java Island, Indonesia, which provides more update field conditions compared to data collected from Statistics Indonesia (BPS). The research found that the stacked model outperforms the individual methods. However, the R2 criterion of the stacked method is comparable to that of RF, which is slightly higher than the others.