Categorization of Ecosystems Based on Soundscape Analysis: A Perspective from Image Classification

This article presents a novel approach to analyze the soundscape in ecosystems, in order to categorize them in terms of their acoustic properties, focusing on the characterization of four ecosystems through an image classification system which contain information of daily acoustic activity in the frequency range (1kHz-11kHz), for five consecutive months. Emphasis is placed on pre-processing of acoustic recordings and their transformation into contour images of normalized power spectral density (nPSD), which acts as a distinctive acoustic metric for each ecosystem. Within the classification system where a support vector machine (SVM) is implemented, the description of the images using SIFT descriptors by dense sampling (DSIFT) and its representation following a BoW scheme is highlighted. The results of the proposed methodology show a high system performance, above 97% accuracy.