Environmental threat classification from mel-spectrograms using an augmented 2D CNN with sub-second CPU inference
Illegal logging threatens biodiversity, carbon storage and ecological stability, yet ranger patrols and camera networks are costly, spatially limited and rarely provide timely alerts. This paper presents ForestEar, a proof-of-concept framework that classifies 2-second audio clips as Safe, Illegal Activity or Environmental Risk using a two-dimensional convolutional neural network (CNN) applied to log-mel spectrograms. In the absence of a labelled forest-threat corpus, classes from the public ESC-50 and UrbanSound8K datasets were remapped as proxy labels, and stochastic augmentation was applied to minority classes. On a held-out test partition, a binary Safe-versus-Illegal-Activity evaluation ( n = 338) reached 92.9% accuracy, with recall of 0.95 (Safe) and 0.90 (Illegal Activity). A three-class configuration that adds the Environmental Risk class is demonstrated in a preliminary evaluation, which identified Environmental Risk as the most difficult class. Per-clip inference took approximately 250 ms on a laptop CPU. Because the results derive from proxy benchmark data, a single data split that was not grouped by source recording, and single training runs, they demonstrate feasibility under benchmark conditions rather than field performance. Validation on forest recordings and on embedded hardware remains future work.
Authors
- K. Ragavan
- Lanka Sree Chathurya
- P.N. Bhargav Teja
Institutions
- Vellore Institute of Technology University (IN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1038/s41598-026-75350-1
- Primary Topic
- Music and Audio Processing
- Type
- article
- Field-Weighted Citation Impact
- 0.00