Deep Learning-Based Multi-Label Classification for Forest Soundscape Analysis: A Case Study in Shennongjia National Park

Forest soundscapes contain rich ecological information that reflects the composition, structure, and dynamics of biodiversity within forest ecosystems. The effective monitoring of these soundscapes is essential for forest conservation and wildlife management. However, traditional manual annotation methods are time-consuming and limited in scalability, while commonly used acoustic indices such as the Normalized Difference Soundscape Index (NDSI) lack the capacity to resolve overlapping or complex sound sources often encountered in dense forest environments. To overcome these limitations, this study applied a deep learning-based multi-label classification approach to long-term field recordings collected from Shennongjia National Park, a typical subtropical forest ecosystem in China. The model automatically classifies sound sources into biophony, geophony, and anthrophony. Compared to the NDSI, the model demonstrated higher precision and robustness, especially under low-signal-to-noise-ratio conditions. While the NDSI provides an efficient overview of soundscape disturbances, it demonstrates limitations in differentiating geophonic components and detecting subtle variations. This study supports a complementary “macro–micro” analytical framework that enables capturing broad, time-averaged soundscape trends through the NDSI, while achieving fine-grained, label-specific detection of biophony, geophony, and anthrophony through the multi-label classification model. This integration enhances analytical resolution, enabling the scalable, automated monitoring of complex forest soundscapes. This study contributes a novel and adaptable approach for real-time biodiversity assessment and long-term forest conservation.