Rainfall prediction in eastern northeast brazil using machine learning and oceanic predictors

Abstract

This study proposes a machine learning (ML) approach to predict rainfall in the eastern sector of Northeast Brazil, a region characterized by significant climatic variability. Using binary classification, models were trained to determine whether four-month period (quadrimester) rainfall would be above or below the historical median. Predictors included oceanic and atmospheric variables (e.g., sea surface temperature, trade winds) identified by previous studies, combined with homogeneous rainfall groups. Data from 1982 to 2023 were divided into quadrimesters (April–July, August–November, December–March) and evaluated using Random Forest (RF), Logistic Regression (LR), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM). Nested cross-validation revealed that RF achieved the highest F1-score (0.671) and recall (0.799) when predicting the rainy quadrimester (April–July), demonstrating strong potential for identifying high-rainfall periods. Despite limited data and high variance, the results underscore ML’s viability for rainfall forecasting in the region, offering a baseline for future research with expanded datasets or advanced models.