Analyzing Poverty Dynamics through Time Series: A Wavelet Transform Approach

Abstract

Reducing global poverty is a central goal of the Sustainable Development Goals (SDGs), requiring timely, high-resolution data to monitor socioeconomic changesespecially in low- and middle-income countries where traditional survey data is sparse. Recent advances in machine learning (ML) and Earth observation (EO) data have enabled new approaches to poverty estimation. However, many existing models rely on aggregated annual or multi-year imagery, overlooking intra-annual variation that is particularly relevant in agriculturally driven economies. This study explores the potential of integrating intra-annual vegetation dynamics captured through Normalized Difference Vegetation Index (NDVI) with annual multi-spectral data to enhance poverty prediction. An unsupervised approach using wavelet transforms is proposed to summarize temporal NDVI signals, allowing models to retain essential seasonal patterns while reducing dimensionality and noise. Experiments were conducted in a simulated environment using nighttime light intensity as a proxy for wealth, and further evaluated against the Demographic and Health Survey (DHS) dataset across Africa. The results show that incorporating selected wavelet-derived NDVI features significantly improves model accuracy over baseline methods, even in the presence of missing data. These findings highlight the critical role of intra-annual temporal information and demonstrate the value of wavelet-based summarization for scalable, robust poverty estimation using satellite imagery