Daily rainfall prediction in Lagos using logistic regression, random forest, and support vector machine with web development

Abstract

Accurate rainfall prediction is critical for agricultural planning, flood-risk management, and urban infrastructure resilience, particularly in tropical coastal cities such as Lagos, Nigeria, where unpredictable rainfall events cause significant socioeconomic disruption. This study evaluates three machine learning (ML) classification models---logistic regression (LR), random forest (RF), and support vector machine (SVM)---for daily binary rainfall prediction using a 22-year meteorological dataset of 8,314 observations sourced from Visual Crossing. Two features were engineered from the raw data: daily temperature range and a seasonal indicator. To address the asymmetric cost of false negatives in a tropical rainfall context, Youden's J statistic was applied to optimise the decision threshold of each model, prioritising sensitivity over the conventional 0.5 default. RF achieved the strongest overall performance, recording an accuracy of 76.28%, sensitivity of 79.27%, F1-score of 76.44%, and area under the receiver operating characteristic curve (AUC-ROC) of 0.8454, outperforming SVM (AUC-ROC: 0.8215) and LR (AUC-ROC: 0.8001). Humidity, cloud cover, dew point, and visibility emerged as the most consistent predictors across models, while moon phase and wind speed showed negligible importance in all three classifiers. All three trained models were deployed in an interactive R Shiny web application, enabling non-technical users, including farmers, planners, and policymakers, to obtain real-time rainfall predictions from meteorological inputs. This study provides an end-to-end rainfall prediction pipeline tailored to the Lagos environment and demonstrates the practical value of coupling ML with accessible deployment frameworks for climate decision-making in developing countries.