| Abstract [eng] |
Accurate and reliable poverty statistics are essential for policy makers to develop effective programs and interventions to reduce income inequality and improve social assistance. However, conventional methods of acquiring socioeconomic data are often time-consuming, resource-intensive, and expensive. Specifically, the Lithuanian Official Statistics Portal offers only annual indicators of at-risk-of-poverty. Access to monthly data could lead to a deeper understanding of poverty dynamics and facilitate forecasting. Convolutional deep learning architectures are increasingly applied to various computer vision problems, including satellite imagery, where aggregated nightlight luminosity is of particular importance as a proxy of socioeconomic activities. Previous studies have shown that remote sensing techniques offer a viable method for assessing poverty levels. In particular, the utilization of convolutional neural networks (CNNs) to analyze satellite imagery has demonstrated potential in forecasting the luminosity of nighttime lights, which can then be used to determine the underlying poverty levels. This study leverages the NASA Black Marble product suite, which offers cloud-free satellite imagery retrieved from Visible Infrared Imaging Radiometer Suite (VIIRS) sensor. These Night-time Light (NTL) images are corrected for various factors including atmospheric conditions, terrain, lunar BRDF effects, thermal and straylight interferences. Thus, they effectively capture the presence of human activity through artificial lighting. During the experiment, monthly NTL images spanning from 2013 to 2023 are retrieved from NASA LAADS DAAC archives and prepared for CNN training: interpolated, normalized and subsequently cut into 56 × 56, 112 × 112 and 224 × 224 pixel tiles. Next, data augmentation is applied by shifting the NTLs by 50 % in both x and y directions which effectively generates additional images and expands the training set respectively to 54780, 15840 and 5016 images. Experiment reuses 6 convolutional pre-trained models (NASNet, ResNet50, ResNet101, DenseNet101, DenseNet201, and EfficientNetb0) to obtain image embeddings, which are further employed in elastic net type regularized regression and random forest models. The tested solution, which relies solely on nighttime satellite data, shows potential for accurate and more nuanced forecasts of poverty. The best model for monthly poverty predictions, incorporating ResNet50 CNN architecture and 56 × 56 size tiles, yielded test R2 – 0.510, MDAE – 1.960 and MAPE – 9.811. |
