A Forest Acoustics – Temporal Frequency Convolution Neural Network Model for Detecting Illegal Logging Activities in Forest

Illegal logging has detrimental effects on the environment, including increased deforestation, which hampers carbon sequestration and leads to heightened carbon emissions. Consequently, there is an imperative to detect and prevent illegal logging to safeguard the environment. Detecting logging activities involves discerning complex environmental sounds. Environmental sounds such as forest acoustics are very complex given that they often have background noises such as wind making it difficult to classify them. This study concentrates on employing deep learning techniques for classifying forest acoustics, tackling the complexities inherent in environmental sounds. It specifically delves into temporal and frequency patterns of forest sounds, isolating sounds indicative of logging activities such as chainsaw sounds. The Temporal Frequency Convolution Neural Network model (TFCNN) is deployed to accurately recognize logging related acoustics while dealing with background noise during classification and data processing. The TFCNN model consists of two main branches, the harmonic spectrogram branch, and the percussive spectrogram branch. The outputs of both harmonic and percussive branches are concatenated to combine the information learnt from both. The model was trained using the FSC 22 dataset, which is tailored for forest - related analyses, making it an ideal choice. To tackle background noise during classification, attention mechanisms and feature representation techniques were employed. The model recorded an accuracy of 87%.