Article Preview
TopIntroduction
The area of Computational Bioacoustic Scene Analysis has received increasing attention by the scientific community in the last decades (Stowell, 2018; Blumstein et al., 2011; Towsey, Truskinger, & Roe, 2015; Dong, Towsey, Zhang, & Roe, 2015; Li, Zhou, Zou, & Li, 2012). Such interest is motivated by the potential benefits that can be acquired towards addressing major environmental challenges including invasive species, infectious diseases, climate and land-use change, etc. Availability of accurate information regarding range, population size and trends is crucial for quantifying the conservation status of the species of interest. Such information can be obtained via classical observer-based survey techniques; however, these are becoming inadequate since they are a) expensive, b) subject to weather conditions, c) cover a limited amount of time and space, etc. To this end, autonomous recording units (ARUs) are extensively employed by biologists (Grill & Schlter, 2017; Ntalampiras, 2018a). This is also motivated by the cost of the involved acoustic sensors which is constantly decreasing due to the advancements in the field of electronics.
One of the first approaches employed for classifying animal vocalizations is described in (Mitrovic, Zeppelzauer, & Breiteneder, 2006). The authors extracted Linear predictive coding coefficients, cepstral coefficients based on the Mel and Bark scale, along with time-domain features describing the peaks and silence parts of the waveform. The classifier was a Support Vector Machine, while three kernels were considered, i.e. polynomial, radial basis function, and sigmoid. These were compared with nearest neighbor and linear vector quantization schemes. The specific dataset included sounds of four animal classes, i.e. birds, cats, cows, and dogs. The literature further includes several approaches which concentrate on specific species, classification of Australian anurans (Han, Muniandy, & Dayou, 2011), interpretation of chicken embryo sounds (Exadaktylos, Silva, & Berckmans, 2014), classification of insects (Noda, Travieso, Snchez-Rodrguez, Dutta, & Singh, 2016), etc. However, a systematic approach addressing the specific case of farm monitoring, is not present in the literature. This work wishes to cover exactly this gap (Figure 1).
Figure 1. The logical flow of the proposed method encompassing a) signal windowing, b) feature extraction, c) concept drift detection, d) statistical affinity calculation, e) ESN-based transfer learning, and f) update of the classification scheme
Indeed, the acoustic modality could provide complementary information to monitor the health as well as population of animals. For example, it could be used in combination with solutions such as (Kumar & Hancke, 2015; Nagpal & Manojkumar, 2016; Anu, Deepika, & Gladance, 2015) which record physiological parameters of the animals, such as rumination, body temperature, and heart rate with surrounding temperature and humidity. The valuable information that can be obtained via the acoustic modality could assist an overall assessment of the current status of the animals as well as the farm in general. More precisely, acoustic farm environment monitoring could assist in the following applications:
- •
Tracking of similar breed animals and parturitions
- •
Identification of specific animal(s) for several reasons (vaccination, medication, diseases, diet, etc.)
- •
Animal health monitoring
- •
Population monitoring
- •
Detect animals missing from the farm
- •
Intruder detection and identification