Transformer-Based Climate Data Imputation for Economic Risk Assessment and Climate Data Sovereignty in Brazil

This study presents the development and evaluation of a Transformer-based framework for imputing missing climate data in Brazil using deep learning and climate reanalysis datasets. The research addresses the economic and scientific impacts of incomplete meteorological records on climate modeling, risk assessment, agricultural planning, and public policy. The proposed architecture combines ERA5 reanalysis data with observational records from the Brazilian INMET station network to reconstruct missing climate variables while preserving physical consistency and spatial coherence. The study evaluates Dense Neural Networks, regression methods, and Transformer-inspired models such as Pangu-Lite for reconstructing atmospheric fields and operational station-level imputation. The framework is integrated into a prototype national-scale climate intelligence infrastructure designed to standardize, harmonize, and distribute high-fidelity meteorological datasets through APIs and monitoring tools. Results demonstrate that neural network approaches outperform traditional interpolation methods in reconstructing variables such as temperature, humidity, pressure, precipitation, and solar radiation. The work highlights the importance of reliable climate datasets for economic forecasting, climate governance, agricultural insurance, adaptation planning, and climate risk management in Brazil.