Automatic Identification of Transformation in the Colombian Tropical Dry Forest Using GMM and UBM-GMM

Today, machine learning methods have become a powerful tool to help curb the effects of global warming by solving ecological questions. In particular, the Colombian Tropical Dry Forest (TDF) is an important ecosystem that is currently under threat due to deforestation generated by cattle, mining, and urban development since colonial times. One of the urgent challenges in this area is to understand the threatened ecosystems landscape transformation and forest degradation. Traditionally, environmental conservation experts measure these changes using transformation levels (high, medium, low). These levels have been obtained through direct observation, counting species, and measures of spatial variation through the time. Therefore, these methods are invasive to the study landscapes and require large amounts of time analysis. A proficient alternative to classical methods is the passive acoustic monitoring, as they are less invasive to the environment, avoid seeing the difficulty of species from isolated individuals, and help reduce the time of researchers at the sites. Even though too much data is generated, and computational tools have been required for their analysis. This paper proposes a new method to automatically identify the transformation in the Colombian TDF. The method is based on Gaussian Mixture Models (GMM) and Universal Background Model (UBM). In addition, it includes an acoustic indices analysis to select the most informative variables. The GMM proposal was tested in two local sites (La Guajira and Bolivar regions) and achieved an accuracy of 93% and 89% for each one, and it was obtained 84% with the general UBM model.

• Fecha de publicación: 2020
• Institución: Universidad Pedagógica y Tecnológica de Colombia
• Repositorio: RiUPTC: Repositorio Institucional UPTC
• Idioma: spa
• OAI Identifier: oai:repositorio.uptc.edu.co:001/14285
• Acceso en línea: https://revistas.uptc.edu.co/index.php/ingenieria/article/view/11752; https://repositorio.uptc.edu.co/handle/001/14285
• Palabra clave: acoustic index; ecoacoustics; gaussian mixture model; machine learning; ; maximum likelihood estimation; universal background model; ecoacústica; modelos de mezclas gausianas; índices acústicos; machine learning; estimación de máxima verosimilitud; modelo universal
• License: http://purl.org/coar/access_right/c_abf386