Sensor arrays for acoustic monitoring of bird behavior and diversity. Preliminary results on source identification using unsupervised learning methods

Background/Question/Methods

In this study we are concerned with developing acoustic sensor arrays so that they will be useful for observing and analyzing bird diversity and behavior. We would like each sensor to see and “understand” part of the situation -- depending on its own location -- then to fuse their experiences with other such sensors to form a single, coherent understanding by the ensemble.

Toward that goal we have developed and tested sensor arrays that can identify their own location and sense bird vocalizations in real-world settings. We have developed filters to identify species (in some instances individual birds) and software tools to localize those individuals in natural environments. In this talk we will briefly touch on those topics, but focus on the issue of classification of bird songs using unsupervised machine learning methods.

The principal field site for our work has been the rainforest environment at the Reserva de la Biosfera Montes Azules, in Chiapas Mexico. The species of birds in our analysis have been the Barred Antshrike (Thamnophilus doliatus), Dusky Antbird (Cercomacra tyrannina),  Great Antshrike (Taraba major), and the Mexican Antthrush (Formicarius analis).

Bird songs were recorded from each of these birds during December 2006, by Martin Cody. The identification of each singer was inferred from timing and location.  Samples of 20 - 50 songs from each of the territories they occupied were included.  The sonogram of each song was measured for 20 traits, including length and maximum or minimum frequency at various parts of the song, and then represented by a vector.

Results/Conclusions

We explored with the use of Self-Organizing Maps for the classification of bird species/individuals from this dataset. The main goal has been to examine the scope in which unsupervised learning is capable of conferring meaningful categorization abilities and increasing autonomy to sensor arrays. We are currently directing efforts at identifying individuals and inferring their territories by these methods, with quite positive preliminary results. 

Overall, adaptive sensor arrays seem promising platforms for habitat monitoring applications. In the near future, our efforts will be directed towards enabling sensor arrays with increasing adaptability. To accomplish this we will build largely on the results presented in this talk.