On Acoustic Monitoring of Rainfall Intensity

Audio-based environmental monitoring is gaining ever-increasing interest in the last decades facilitating a wide range of applications. An emerging task concerns the automatic estimation of rainfall intensity based on the respective acoustic activity. This work proposes an audio processing and modelling pipeline tailored to the requirements of the specific task. More precisely, we a) preprocessed the audio signals through filtering prior to feature extraction, b) integrated meteorological data as auxiliary features, c) explored different FFT window lengths considering the stationary characteristics of the available data, and d) constructed an ensemble model by stacking multiple transformer-based regressors. Importantly, during this analysis, we employed a publicly available dataset, i.e. SARID, adopting a standardized experimental protocol enabling reliable comparison of different approaches. Finally, the optimised model ensemble achieved a noticeable increase over the state of the art. Last but not least, the implementation of the described experimental pipeline is available at https://www.kaggle.com/code/imemine/ ensemble-model-for-rain-intensity-estimation.