关键词: 声景分类, 被动声学监测, 声学指数, Mel频谱图, 可解释人工智能

AbstractIn complex soundscapes, multiple sound sources often coexist, and acoustic components overlap in temporal, spectral, and statistical dimensions, making it difficult for a single representation to capture all discriminative information. This study investigates dual-feature fusion and model interpretability for complex soundscape classification. Mel-spectrograms and acoustic-index sequences are first modeled with separate branches and then fused, and the framework is evaluated on nature-reserve soundscape, urban-park soundscape, and urban environmental sound datasets. LIME is used for the best spectrogram model, SHAP for the best sequence model, and a low-dimensional bottleneck combined with symbolic regression for the fusion model. Input-perturbation controls are applied to the spectrogram and sequence explanations, whereas surrogate R² and MSE are used to assess the fidelity of the symbolic proxy for the fusion model. Under the evaluated architectures and unified training conditions, Mel-spectrogram models show stronger standalone classification performance, while acoustic-index sequences computed as short-time statistics within 3-s clips provide complementary information. On CBL-data, fusion improves Micro-F1 and Macro-F1 over the best spectrogram model by 0.0560 and 0.0365, reaching 0.9065 and 0.7141, respectively. In contrast, fusion performs below the best spectrogram model on the urban-park dataset, indicating scene-dependent benefits. LIME-highlighted time-frequency regions show clear decision relevance. For SHAP, visually high-contribution time steps sometimes coincide with target acoustic events, but the equal-size random zeroing control does not show a stable advantage; this observation is therefore treated qualitatively. For the fusion model, symbolic regression yields high-fidelity proxies, and most class formulas reduce to linear combinations of bottleneck variables rather than explicit interactions. The resulting multi-level interpretation and validation workflow provides a methodological reference for reliability assessment in automatic soundscape recognition.

Key words: soundscape classification, passive acoustic monitoring, acoustic indices, Mel spectrogram, explainable artificial intelligence (XAI)