ForestX-Net: a novel acoustics events recognition model for the forest environment using hybrid deep learning architecture

Recognizing the acoustics events involves the identification and classification of auditory objects within various environments. For the forest ecosystems, it plays a vital role in domains such as monitoring biodiversity, detecting illegal activities, managing environmental threats and wildlife preservation. In this study, a novel framework which is named as ForestX-Net was proposed to classify forest acoustics events using a hybrid deep learning approach. The dataset comprises 10 distinct sound classes which are fire, rain, thunderstorm, helicopter, axe, chainsaw, gunshot, footstep, frog, and wolf howl, with 74 samples per class, recorded under authentic forest conditions. Spectrogram representations of these audio signals were extracted and employed as inputs to a pre-trained ResNet-18 model. Feature embeddings from ResNet-18 yielded a 740x512 feature matrix, which was subsequently utilized as input to a Multilayer Perceptron (MLP). The proposed architecture achieved an exact test accuracy of 92.57%, demonstrating its effectiveness in distinguishing acoustically diverse sound events.