A Lightweight Convolutional Neural Network for Environmental Sound Recognition Toward Smart Monitoring in Mountain Forest Parks
With the development of smart ecological conservation and intelligent scenic-area management, acoustic monitoring has become an increasingly important means of ecological perception in forest parks. However, existing environmental sound recognition models generally entail substantial computational demands that are difficult to satisfy on low-power field devices; meanwhile, publicly available sound datasets are mostly collected in urban environments and differ markedly from the complex multi-source soundscapes of mountain forest parks. To address these problems, this study constructs a localized environmental sound dataset for Zhangjiajie National Forest Park and proposes a lightweight convolutional neural network (L-CNN) for environmental sound recognition oriented toward smart monitoring. Taking mel spectrograms as input, the model builds upon the PANNs-CNN10 architecture and integrates an efficient channel attention mechanism and a temporal attention module. Comparative experiments against six representative audio classification models show that the L-CNN achieves an accuracy of 96.74% and a macro-averaged F1 score (Macro-F1) of 96.74% on the nine-class recognition task, while reducing the parameter count from 5.23 × 106 to 3.15 × 106 (a 39.8% reduction) and the model size from 19.9 MB to 11.51 MB (a 42.2% reduction) relative to PANNs-CNN10. The results demonstrate that the proposed model can substantially reduce model complexity while maintaining competitive recognition accuracy, thereby providing technical support for automated acoustic monitoring and ecological perception in mountain forest parks.
Authors
- De-Bing Zhuo (ORCID: https://orcid.org/0000-0002-8059-048X)
- Zhan Hu (ORCID: https://orcid.org/0009-0008-1875-9601)
- Shuowen Li
- Binzhao Guo
- Zheqian He
Institutions
- Jishou University (CN)
Publication Details
- Journal
- Sensors
- Published
- 2026-11-20
- DOI
- https://doi.org/10.3390/s26227337
- Primary Topic
- Music and Audio Processing
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- Jishou University