Methodology for Generating, Augmenting, and Validating an Audio Dataset for Classifying Fire and Forest Sounds

This paper proposes a methodological framework for building a binary audio dataset for the automatic classification of fire sounds and forest ambience. Two operational recordings, one for the fire class and one for the forest class, are used strictly as seed data for controlled segmentation and augmentation. The workflow includes mono conversion at 16 kHz, amplitude normalization, segmentation into 5 s windows with 2 s overlap, low-intensity stochastic augmentation, and the systematic logging of the generated samples. The study also explains why augmented data are appropriate for training and internal validation, while final performance claims must remain reserved for testing on independent, standardized real recordings.