Speaker
Description
Traditional rain gauges are often challenged in terms of coverage and maintenance, especially in tropical urban and remote areas. In this context, opportunistic acoustic sensing using, e.g., wildlife recording devices or even cell phones, could offer a scalable alternative to complement them and alleviate such challenges. However, acoustic-based rainfall detection, classification, and estimation models struggle with methodological adaptability across diverse environments and climatological events' imbalances. In this work, we propose and evaluate a signal processing framework for rainfall classification in two contrasting soundscapes: a semi-urban university campus and an Amazonian rainforest. The framework exhaustively computes a vector with 149 acoustic metrics readily available from off-the-shelf signal processing software. Then, an empirical optimization is performed to select a small subset of metrics, which are used to train Stochastic Gradient Descent, Logistic Regression, XGBoost, and Random Forest classifiers. To mitigate data skewness, we apply a strict 1:1 global class balance among rainfall categories, as well as apply stochastic augmentation. The classifiers are evaluated considering a standard 5-class precipitation scale and a condensed 3-class scheme. Random Forest operating on the 3-class structure significantly outperformed all other algorithms overcoming both urban and forest noise and achieved peak accuracies of 83.03% (semi-urban) and 92.15% (rain forest) on the globally balanced datasets, whereas the 5-class baseline scored 71.57% and 79.14%, respectively. The results demonstrate that acoustic sensing and machine learning can provide reliable rainfall detection in complex, poorly gauged environments.