Uma abordagem para monitoramento de anuros baseada em processamento digital de sinais bioacústicos

Abstract

Wildlife monitoring is often used by biologists and ecologists to acquire information about animals and their natural habitats. In survey programs, specialists collect environmental information to infer about animal population status and their variations over time. The main goal of such programs is to identify environmental problems in early stages. However, acquiring the necessary data for this purpose is a manual work and must be carried out by groups of experts in areas of di cult access during long periods of time. In this context, Wireless Sensor Networks (WSNs) are useful alternatives to alleviate the manual work. Such networks are made up of small sensors with transmission, storage, and local processing capabilities. These networks enable bioacoustic methods for automatic species recognition to be embedded in the sensor nodes in order to automate and simplify the monitoring task. Since animal sounds usually provide a species ngerprint, it can be used to recognize the presence or absence of a target species in a site. Accordingly, in this thesis, we present an approach that combines machine learning methods, WSNs and bioacoustic signal processing techniques for wildlife monitoring based on animal calls. As a proof-of-concept, we choose anurans as the target animals. The reason is that anurans are already used by biologists as an early indicator of ecological stress, since they provide relevant information about terrestrian and aquatic ecosystems. Our solution integrates four fundamental steps: noise ltering and bioacoustic signal enhancement, automatic signal segmentation, acoustic features extraction, and classi cation. We also consider the WSNs limitations, trying to reduce the communication and processing load to extend the sensors' lifetime. To accomplish with the restriction imposed by the hardware, we represent the acoustic signals by a set of low-level acoustic descriptors (LLDs or features). This representation allows us to identify speci c signal patterns of each species, reducing the amount of information necessary to classify it. The adverse environmental conditions of the rainforest pose additional challenges, such as noise ltering. We developed a ltering method based on Singular Spectrum Analysis (SSA). This choice was based on several comparisons with other ltering methods. The SSA method has additional advantages: it is non-parametric, it adapts to the di erent input signals, and it has an equivalent