Combining satellite imagery and a GAN-based data augmentation method for poverty estimation

Explicit poverty data are crucial for formulating precise poverty alleviation policies. Existing methods for estimating poverty with remote sensing data are restricted by limited training samples. This study presents a poverty estimation model using high-resolution remote sensing (HRRS) data enhanced through a generative adversarial network (GAN). The method combines night-time light (NTL) data with Demographic and Health Surveys wealth index (WI) data to generate NTL-WI proxy labels representing poverty levels. A two-step training process is used to train the GAN model, enabling it to generate HRRS image datasets corresponding to specified poverty levels and expanding the available training data for poverty estimation. A convolutional neural network and a ridge regression model are then employed to generate poverty estimates. Results from Senegal, Tanzania, and Rwanda show R² values of 0.70, 0.60, and 0.51, respectively, outperforming the non-augmented baseline values of 0.58, 0.47, and 0.41, as well as models using only NTL or WI labels. Furthermore, an HRRS dataset is constructed comprising approximately 300,000 images of various poverty levels, and asset wealth maps are generated for the three countries. These outcomes highlight advancements in GAN-based data augmentation for poverty estimation, practical tools for detailed mapping, and a foundation for targeted poverty policies.