Unveiling nonlinear ecological dynamics in marine soundscapes with explainable AI

In the context of the ongoing marine biodiversity crisis, marine soundscape monitoring has emerged as a promising and cost-effective approach for assessing ecosystem health through the analysis of interactions among biotic, abiotic, and anthropogenic factors. Although acoustic indices are widely employed to identify specific sound events derived from biophony, geophony, or anthropophony and their dominance within particular frequency ranges, these indices typically fail to capture the complex temporal interactions and variability introduced by biotic–abiotic relationships. Additionally, this complexity is further intensified by the exponential growth in acoustic datasets driven by advancements in autonomous monitoring technology. Addressing these limitations, this study proposes a novel, data-driven framework combining machine learning and explainable artificial intelligence (XAI) to quantify the multi-directional influence of various abiotic and biotic factors on marine soundscapes across multiple frequency bands. Specifically, I applied the eXtreme Gradient Boosting (XGBoost) algorithm combined with SHapley Additive exPlanations (SHAP) analysis to long-term acoustic data collected from the Taiwan Strait between 2014 and 2018. This analysis investigates the relative impact of abiotic factors (tidal level, lunar cycles, annual, seasonal, and diurnal variations, and temperature) and biotic factors (fish chorusing and snapping shrimp) across four distinct frequency ranges: 10–300 Hz (primarily geophony), 300–3000 Hz (dominated by fish chorusing), 3000–24,000 Hz (dominated by snapping shrimp), and the complete range of 10–24,000 Hz. Results demonstrate that SHAP clearly identifies the primary drivers within each frequency band and quantifies the magnitude of the contributions of multiple abiotic and biotic factors. Moreover, SHAP dependence plots highlight specific temporal trends and patterns linked to seasonal, diurnal, annual, temperature, fish chorusing, and snapping shrimp influences. This novel XAI-based analytical approach offers enhanced interpretability, providing a robust tool for exploring soundscape-environment relationships, thereby advancing marine conservation and ecosystem management efforts. • SHAP-XGBoost quantifies multi-driver biotic & abiotic effects on marine soundscape. • Interpretable ML links biophony to abiotic drivers with ranked effect sizes. • Identified seasonal, diurnal, and temperature-related acoustic patterns. • Integrated entropy-based acoustic indices to enhance ecological interpretability. • Actionable indicators and thresholds support targeted monitoring and policymaking.