Geospatial Deep Learning for Estimating Socioeconomic Well-Being in Developing Regions: An Analysis of Burundi’s Poverty Projection

Rural poverty remains a challenge in developing nations primarily due to limited socioeconomic (SE) data, restricting effective policy-making and humane projects. The substantial progress in Remote sensing satellite technology and open data initiatives have enhanced accessibility to spatial information. This unique ability is crucial for understanding the social problems in specific locations, facilitating targeted and effective solutions. Thus, proposing a novel deep learning approach that integrates high-resolution satellite images with Demographic Health Survey (DHS) data to extract critical SE indicators such as wealth index, access to water, electricity, cellphones, education, and health results. Employing Transfer Learning (TL) strategy with Inception V3 and VGG16 architectures that utilizes nighttime (NT) light intensity images as a reliable data substitute which enhances poverty prediction by learning features from imagery, and fine-tune Inception V3 model for improved performance. The fully convolutional neural network (CNN) not only predicts NT lights from daytime images but also captures essential features such as landscapes and man-made constructions like farmlands, buildings, and highways/roads without any additional monitoring. As a result, the model not only aligns well but also outperforms on-site survey data, achieving improved accuracy in mapping poverty.