A machine learning framework for acoustic rainfall classification applied to two distinct soundscapes

Traditional rain gauges face coverage and maintenance limitations in tropical areas. Opportunistic acoustic sensing offers a scalable alternative; however, models often struggle with environmental adaptability and data imbalance. This study proposes a Machine Learning framework for rainfall classification across two contrasting soundscapes: Fortaleza's semi-urban and Amazonian forest. The pipeline extracts 102 features followed by site-specific dimensionality reduction. To handle data skewness, we apply undersampling and stochastic augmentation. Classifiers such as Random Forest, Extreme Gradient Boosting, and Support Vector Machine are evaluated. Results indicate that integrating acoustic feature engineering with rigorous dataset curation enables reliable rainfall classification in complex environments.