Sound-based logging detection using deep learning

Abstract

Illegal logging represents a major environmental issue which impedes the stability of forest ecosystem and supports climate change, flooding, soil erosion, weeding of habitats, extinction of animal and plant species. This paper proposes a method for mitigating the logging issue by automatically detecting the sound of logging activities. More specifically, we propose to detect the chainsaw sound using deep learning. Two deep learning approaches were considered, one based on multilayer perception (MLP) and the other based on convolutional neural network (CNN). As inputs to our models, we used time, frequency and time-frequency audio features. For this research, we collected two datasets of audio signals. First dataset, downloaded from YouTube, is used for training and validating the proposed models. Second dataset, which we recorded in a real environment, is used for testing of the proposed models. The experiments have shown that the CNN-based approach outperforms the MLP-based one, with a sound classification accuracy of 94.96% on the first dataset and 88.87% on the second dataset.