Machine Learning Integrating Climate Data with GRACE Mission Data for the Reconstruction of Terrestrial Water Storage Anomalies in Brazil

• Ano de defesa: 2025
• Autor(a) principal: SANTOS, Lucas de Siqueira
• Orientador(a): GONÇALVES, Rodrigo Mikosz
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Pernambuco
• Programa de Pós-Graduação: Programa de Pos Graduacao em Ciencias Geodesicas e Tecnologias da Geoinformacao
• País: Brasil
• Palavras-chave em Português: Terrestrial water storage; Machine learning; GRACE; Climate change
• Link de acesso: https://repositorio.ufpe.br/handle/123456789/67178

Terrestrial water storage (TWS) is a critical component of the hydrological cycle, directly influencing water security, energy production, and climate resilience in Brazil. Although the country has abundant freshwater resources, their uneven spatial distribution combined with the growing impacts of climate change exposes both the population and the economy to hydrological risks. The Gravity Recovery and Climate Experiment (GRACE) missions have provided valuable insights into TWS variability since 2002; however, their limited temporal coverage constrains long-term analyses. To overcome this limitation, this research reconstructed terrestrial water storage anomalies (TWSA) for Brazil’s 12 major river basins from 1985 to 2002, integrating GRACE data with climatic variables (precipitation, soil moisture, temperature, and teleconnection indices) and anthropogenic indicators derived from land use and land cover (LULC) data. Two machine learning models—Random Forest (RF) and Long Short-Term Memory (LSTM)—were implemented and compared to assess performance, interpretability, and suitability for GRACE-TWSA reconstructions. Results indicate natural seasonality throughout the year, with vegetation and climate indices emerging as highly influential predictors of TWSA, while anthropogenic factors affect anomalies differently across basins, particularly in areas dominated by agriculture and livestock activities (such as cotton in the Amazon Basin and perennial crops in the São Francisco Basin). Both RF and LSTM achieved satisfactory performance, though LSTM was able to reconstruct the time series for only a few basins, while RF provided greater interpretability of variable contributions. The Mann-Kendall test applied to the RF-reconstructed TWSA series revealed significant long term decreasing trends in the Uruguay, Parnaíba, São Francisco, and East Atlantic basins, underscoring Brazil’s vulnerability to water stress under future climate scenarios. By extending GRACE-derived observations, this study advances understanding of how climate and LULC influence TWSA variability and provides evidence to support public policies for sustainable water resource management in Brazil.