Machine Learning Framework for Hydrological Drought Forecasting in Brazilian Basins with Diverse Climates

Drought is a multifaceted natural hazard characterized by complex mechanisms, diverse contributing factors, and slow onset, affecting food, water, energy, and ecosystem security. Brazil, like many regions worldwide, has faced significant drought challenges over the past decade, impacting basins that play a critical role in water supply, hydropower generation, and agriculture. This study explores the application of Machine Learning (ML) algorithms and Two-variate Standardized Index (TSI) to forecast drought conditions at 3- and 6-month time scales.

In this study we employ Support Vector Regression (SVR) and Multilayer Perceptron Artificial Neural Networks (ANNs), using as predictors univariate indices and climate indices representing climate modes of variability that influence Brazil's precipitation and drought regimes. Our methodology includes feature selection through Recursive Feature Elimination, lagged correlations, and statistical evaluation using the Mean Absolute Error (MAE), Mean Square Error (MSE) and Coefficient of Determination (R²).

Results demonstrate that both SVR and ANN models effectively predict drought conditions, with R² varying between 0.71 and 0.91, MRS less than 0.2 and MAE not exceeding 0.35, for key indices at 3- and 6-months lags. The strong predictive performance underscores the potential of ML to address challenges in drought forecasting, enabling proactive water resource management and mitigation in regions vulnerable to hydrometeorological extremes.