Predicting per capita expenditure using satellite imagery and transfer learning: A case study of east Java province, Indonesia

Collecting poverty data through the National Socio-Economic Survey (SUSENAS) demands significant time, costs, and human resources. To enable more efficient policy-making, predicting the poverty rate before the release of Statistics Indonesia (BPS) data is essential. This research compares day and night satellite images to predict per capita expenditure in East Java, Indonesia, which has the highest number of poor people. The satellite images are processed using a transfer learning approach that employs a pretrained Convolutional Neural Network (CNN) model with VGG-16 architecture as a feature extractor. These extracted features are then used as independent variables to predict East Java's per capita expenditure using Support Vector Regression (SVR) with RBF and polynomial kernels. The findings indicate that night images are more reliable than day images, with the best model being a combination of transfer learning and the SVR polynomial kernel using night images. The prediction mapping aligns well with the unmodeled night image, demonstrating the effectiveness of this approach in predicting per capita expenditure.