ADistributeApproach forClassifyingAnuranSpeciesBasedonTheirCalls
Juan G. Colonna, Marco A. P. Cristo, Eduardo F. Nakamura{juancolonna, marco.cristo}@icomp.ufam.edu.br, [email protected]
IntroductionIn this work, we evaluate the performance of adistributed classification system in aWire-less Sensor Network for monitoring anurans.Our aim is to study how to take advantage ofthe collaborative nature of the sensor networkto improve the recognition of anuran calls. Toaccomplish this, we evaluate four low-cost tech-niques to combine three classifiers. We modelthe problem as an ensemble of classifiers andwe also propose and evaluate a rejection tech-nique on the a posteriori species probability vec-tor to discard confusing scenarios. In summary,our contribution consists in verifying two hy-potheses: (1) the decision provided by a sensorcommittee is better than the one provided by asingle sensor and (2) a rejection technique canreduce the misclassification rate. We found thatthe sensor committee with rejection is able to ef-fectively identify confusing scenarios, increasinggains over the isolated sensor to about 20%.
MethodologyTo test the first hypothesis, we compare the per-formance of the sensor committee with a singlesensor. To test the second hypothesis, we com-pare the error rate as the methods reject scenar-ios they identified as confusing. For the tests,we simulated scenarios where one or twoanuran species vocalize at the same timein an environment with attenuation and Gaus-sian noise. In addition, we analyze the impactof using three classification techniques and fourways of combining the classifiers.
MotivationAccounting amphibian populations, specificallyanurans (frogs or toads), is a common tool usedby biologists as an early indicator of environ-mental stress. The reason is that anuransare closely related to the ecosystem.
Employing automatic classification methodsand WSNs may help to estimate long-termchanges in amphibian populations.
ChallengeIn real situations, sensors will face many scenar-ios where random noises, other animal vocaliza-tions and even different species of anurans arepresent.
Wave form and spectrogram of two diferents species.
Providing reliable estimates is important foridentifying and discarding confusing sce-narios.
Proposed ApproachIn this work, an anuran call is an event that isdetected and processed by clusters of sensors.Thus, if we deploy L sensors in a certain area,each of them will have its own opinion aboutthe detected event. Then, sensors colaborateto combine these opinions, and identifydisagreements and differing estimates.
Combination Techniques1. The first combination strategy is a majority voting (MV), in which, the final decision is the
most common among the base classifiers.2. In our scenario, is reasonable to give more importance, or weight, to the classifiers who are
closer to the source due to the attenuation. Thus, we implemented a weighted majorityvoting (WMV), where the power of the received signal is used as weight.
3. The geometric average rule (GV): argmaxwj
∏Li=1 Pi(wj |x).
4. The arithmetic average rule (AV): argmaxwj
1L
∑Li=1 Pi(wj |x).
Although there are other combination techniques beyond the previously explained, we chose theseones due simplicity and low computational cost.
ResultsWe first compare the decision of combined sensors with the decision of a single sensor.This comparison is presented in Table 1, where error rates are used to assess the classifier perfor-mances. Each line in this table corresponds to the proportion of scenarios that were rejected (RR)according to their entropy.
Table 1: Error rates of Quadratic Discriminant Analysis (QDA)RR IS MV G(%) WMV G(%) GV G(%) AV G(%)0% 36.9 34.6 +8.1 34.2 +7.2 34.1 +7.7 33.8 +10.810% 32.9 31.8 +6.0 30.7 +7.8 34.1 -3.1 30.1 +9.020% 29.7 28.9 +2.4 28.6 +2.4 35.1 -22.0 26.1 +9.430% 28.3 28.9 +1.1 24.4 +15.2 35.1 -23.7 23.7 +18.740% 25.1 28.9 -11.6 22.3 +12.4 36.5 -43.4 20.8 +20.3
Gains (G) showed in bold face represent statistically significant(p < 0.05) differences to the baselines IS.
When we consider the situation where no scenario is rejected (RR = 0%), all combination strategiesinvolving the QDA classifier outperform the single sensor (IS). In general terms, QDA obtainedthe smallest error rates among the three classifiers we tested. Amongst the combinationstrategies, the AV was the best combining strategy. The results confirm that we can improvethe vocalization recognition by combining the sensor decisions. However, the best gain obtained wasmoderate (10.8%) due to the fact that the sensors provide little independent information since theyuse the same classifiers and almost the same input.As expected, the error rate for the isolated sensor decreases as it rejects the cases of high entropy(RR < 0%). The entropy was very effective in filtering confusing scenarios when usedalong with the arithmetic and weighted voting. This suggests these strategies were better forcapturing disagreements among the sensors that indicate multiple vocalizations. In such cases, thegreater the number of rejected cases, the greater the gain of the combined sensors over the singleone.
AcknowledgementsThe authors acknowledge the support granted by FAPEAM through process number 01135/2011and 2210.UNI175.3532.03022011 (Anura Project - FAPEAM/CNPq PRONEX 023/2009). We alsothank to professor Eulanda Miranda dos Santos.
