Bias Correction Methods for Simulated Precipitation in the Brazilian Legal Amazon

Fabrício Daniel Dos Santos Silva^1*Helber Gomes¹Claudia Priscila Wanzeler Da Costa²Antonio Nogueira Neto²Ismael Guidson Farias de Freitas¹Mário Henrique Guilherme dos Santos Vanderlei¹Maria Cristina Silva¹Rafaela Lisboa Costa¹Jean Sousa dos reis¹Vania dos Santos Franco²Ana Paula Paes dos Santos²Ivan Saraiva³Rodrigo Lins da Rocha´Júnior⁴Jório Bezerra Cabral Júnior¹Helder José Farias da Silva¹Edmir Dos Santos Jesus²Douglas Batista da Silva Ferreira²Renata Goncalves Tedeschi²

• ¹Federal University of Alagoas, Maceió, Brazil
• ²Instituto Tecnologico Vale Desenvolvimento Sustentavel, Belém, Brazil
• ³Fundo Amazonia Centro Gestor e Operacional do Sistema de Protecao da Amazonia, Rio de Janeiro, Brazil
• ⁴SIMEPAR, Curitiba, Brazil

This study aimed to evaluate precipitation estimates over the Brazilian Legal Amazon (BLA) using high-resolution historical simulations from the MPI-ESM1-2-HR climate model, both before and after regionalization with the RegCM4.7.1 model. Continuous 32-year simulations (1981-2012) were compared against observed precipitation data on a regular 0.5° × 0.5° grid over the BLA. Six experiments were conducted: (1) MPI, comparing raw MPI-ESM1-2-HR precipitation with observations; (2) REG, comparing regionalized MPI-ESM1-2-HR precipitation via RegCM4.7.1 with observations; and (3-6) four experiments applying two bias correction methods, canonical correlation analysis (CCA) and principal component regression (PCR), to the MPI and REG out-puts, resulting in MPI-CCA, MPI-PCR, REG-CCA, and REG-PCR experiments. Monthly evaluations revealed very low average correlations (r) between the uncorrected simulations and observations: 0.008 for MPI and 0.013 for REG, with mean ab-solute errors (MAE) of 80 mm and 120 mm, and root mean square errors (RMSE) of 97 mm and 143 mm, respectively, indicating poor representation of observed climatology. However, the application of CCA and PCR substantially improved the simulations. MPI-CCA achieved r = 0.36, MAE = 43 mm, and RMSE = 54 mm, while REG-CCA reached r = 0.41, MAE = 42 mm, and RMSE = 53 mm. The best performance was ob-served with PCR: MPI-PCR showed r = 0.47, MAE = 40 mm, and RMSE = 51 mm, whereas REG-PCR obtained the highest accuracy with r = 0.52, MAE = 39 mm, and RMSE = 50 mm. These improvements were corroborated by Kling-Gupta Efficiency (KGE) analysis, reinforcing its value as a metric for precipitation simulation assessment. Among all months, REG-PCR achieved superior correlation and lower errors in eight out of twelve months (February, March, April, July, September, October, November, and December). MPI-PCR performed better in January, June, and August, while REG-CCA stood out only in May. These findings underscore the importance of bias correction, particularly PCR, in reducing uncertainties in future precipitation projections for the BLA. The results highlight the potential for applying PCR to model outputs to improve projections of climate extremes, thereby supporting strategic planning across multiple sectors in this critical region.