Rainfall Projections for the Brazilian Legal Amazon using Recurrent Neural Networks

Abstract

Rainfall in the Brazilian Legal Amazon (BLA) is vital for climate and water resource management. This research uses spatial downscaling and validated rainfall data from the National Water and Sanitation Agency (ANA) to ensure accurate rain projections with artificial intelligence. Recurrent Neural Networks (RNN) with Long Short-Term Memory (LSTM) were employed to forecast rainfall from 2012 to 2020. The RNN model showed strong alignment with observed patterns, accurately predicting rainfall seasonality. However, median comparisons revealed fair approximations with discrepancies. The Root Mean Square Error (RMSE) ranged from 6.7mm to 11.2mm, and the coefficient of determination (R²) was low in some series. Extensive analyses showed low Wilmott agreement and high Mean Absolute Percentage Error (MAPE), highlighting limitations in projecting anomalies and days without rain. Despite challenges, the study lays a foundation for future advancements in climate modeling and water resource management in the BLA.