Título del libro: Lecture Notes In Networks And Systems
Título del capítulo: Exploring and Evaluating Multisource Remote Sensing for Poverty Mapping in Mexico: Integrating Geomedian and VIIRS Night-Time Light Data
ALLAN FRANCISCO RAMIREZ GONZALEZ; DELFINO VARGAS CHANES;
Autores externos:
Idioma:
Año de publicación:
2026
Palabras clave:
Band model; Convolutional neural network; Deep learning; Me-xico; Multi-Sources; Night time lights; Poverty mappings; Poverty prediction; Remote-sensing; Spectral band
Resumen:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.In recent years, the use of remote sensing data combined with deep learning has become increasingly popular for mapping socioeconomic indicators, benefiting from the growing availability of large-scale Earth observation datasets. This study examines the integration of remote sensing sources with varying numbers of spectral bands and resolutions to enhance the inference capabilities of convolutional neural networks (CNNs) for predicting multidimensional poverty in Mexico. We explore and evaluate two methods for combining daytime satellite imagery and nighttime lights to predict multidimensional poverty at the census track level (Basic Geostatistical Area, AGEB) in Mexico. To achieve this, we gather Landsat-8 imagery (comprising 6 spectral bands) and VIIRS annual composite nighttime lights (NTL) and combine them into a single numeric array in two distinct ways. The first method consists of a band composed entirely of a single NTL intensity AGEB statistic value correlated with the poverty rate. The second method integrates rescaled NTL pixel values as a seventh input band, spatially aligned with the Geomedian imagery. To see if proposed methods improve poverty inference, we train five CNN regression models using the 2020 Municipal Multidimensional Poverty measure from CONEVAL as reference values: (1) using Geomedian only, (2) Geomedian plus derived NTL value, (3) Geomedian plus rescaled NTL by Nearest Neighbor algorithm while (4) and (5) using same method but resampled by Bilinear and Cubic kernels. Predictions are generated at the AGEB level and aggregated to the municipal and state levels using weighted averages. The results show that the seven-band models outperform the six-band model. The best-performing model achieved an R2 of 0.76 at the municipal level, and up to 0.96 at the state level. Our research demonstrates that combining multi-source satellite data, including those with different spatial and spectral resolutions, can improve CNN performance for poverty prediction at different aggregation levels. This approach supports more accurate and granular socioeconomic analysis, contributing to data-driven public policy efforts in Mexico.
Entidades citadas de la UNAM:
ISBN:
9783032173201
Editorial:
Springer Science and Business Media Deutschland GmbH
