Marine soundscape forecasting: A deep learning-based approach

Abstract

Advancements in autonomous monitoring technology over the past decade have led to the widespread use of marine soundscape monitoring to assess marine environments. These environments are rapidly changing and exhibit complex temporal patterns and trends across different frequencies, influenced by biotic and abiotic factors as well as extreme events. This variability introduces a high degree of unpredictability. Despite the rapid development of anomaly detection algorithms and deep-learning models for forecasting, their application to marine soundscapes remains unexplored. This study investigates the use of the unsupervised learning-based isolation forest (iForest) technique to detect anomalous events in marine soundscapes that cause sudden changes in sound levels. Additionally, it evaluates the potential of deep-learning models for estimating trends and forecasting soundscapes while identifying the factors that influence their accuracy. To address these questions, I used marine passive acoustic monitoring data collected from the Taiwan Strait in 2017. The iForest method identified a higher number of anomalies (∼17) in the lower frequency range (10–500 Hz) with a precision of 75 %, primarily due to typhoons, cold bursts, and flooding. In contrast, precision was around 50 % in the mid (500–3000 Hz) and high (3000–24,000 Hz) frequency ranges, where most anomalies resulted from sudden changes in the acoustic behaviors of fish and shrimp, respectively. To analyze trends in marine soundscapes at different temporal scales—annual, seasonal, and diurnal—the anomaly-informed NeuralProphet model was employed. Results showed that NeuralProphet effectively captured annual and seasonal trend changes compared to the traditional singular spectrum analysis method. Beyond NeuralProphet, I also tested two recently developed state-of-the-art forecasting models—time-series dense encoder (TiDE) and neural hierarchical interpolation for time series (NHiTS)—to predict marine soundscapes. In the seven-day ahead seasonal forecasting task, the NHiTS model outperformed both TiDE and NeuralProphet. The deep-learning forecasting models produced more accurate predictions in the mid (500–3000 Hz) (MAE ∼0.4–1) and high (3000–24,000 Hz) (MAE ∼1.5–3) frequency ranges, where seasonal acoustic activity from fish and shrimp strongly influenced sound levels. In contrast, forecast accuracy declined in the lower frequency range (10–500 Hz) (MAE ∼4–8), where sound levels are more stochastic due to anthropogenic and meteorological influences. The findings of this study highlight the potential of deep-learning models for forecasting and trend estimation in marine soundscapes. These models not only improve our understanding of the conditions under which trends change but also enhance our ability to anticipate anomalies and forecast failures. This capability could provide researchers and policymakers with a powerful tool for monitoring transitions and deviations across different temporal scales, ultimately contributing to the conservation and management of marine ecosystems.