1. Recognition of bumblebee species by their buzzing sound
Mukhiddin Yusupov, Mitja Luštrek, Janez Grad and Matjaž Gams
Jožef Stefan Institute Department of Intelligent Systems

2. Bumblebees 35 species in Slovenia
Important pollinators, faster than honeybees, fly even in bad weather
Sometimes bred in captivity for pollination in greenhouses
One queen, a few hundred workers in a nest
Endangered by intensive farming resulting in lack of flowers and safe nesting places

3. Our task Build a classifier: Input: recording of bumblebee sound
Output: species and queen/worker type

4. Presentation structure
Related work
Dataset
Preprocessing
Feature extraction
Classification
Experimental results

5. Related work
No work on classification of bumblebees

6. Related work No work on classification of bumblebees Some work on:
General insect recognition for pest control
A small competition for the best distance metric for insect sounds
Birds
Frogs
...

7. Dataset 71 recordings of bumblebees in the wild Queens of 8 species
Workes of 7 species
15 classes in total

8. Preprocessing Manual segmentation Easy to see 3 relevant segments
Not so easy to see that only 20% buzzing

9. Preprocessing Manual segmentation Volume normalization
Pre-emphasis: boost high frequencies to emphasize buzzing over environmental noise
Easy to see 3 relevant segments
Not so easy to see that only 20% buzzing

10. Feature extraction (1)
Segment the signal into short windows

11. Feature extraction (1) Segment the signal into short windows
Mel-frequency cepstral coefficients (MFCC):
Fourier transform: time → frequency domain
Psycho-acoustic transformation of the frequency-domain data (appropriate binning etc.)
Use 13 coefficients (bins)

12. Feature extraction (2) Linear predictive coefficients (LPC):
Signal at a given time expressed as a linear combination of signals at previous times
Select coefficients so that the difference between the signal expressed this way and the real signal within the window is minimized
Use 10 coefficients

13. Feature extraction (2) Linear predictive coefficients (LPC):
Signal at a given time expressed as a linear combination of signals at previous times
Select coefficients so that the difference between the signal expressed this way and the real signal within the window is minimized
Use 10 coefficients
Compute mean coefficient values over all the windows in a recording

14. Classification
Three-fold cross validation (too few recordings per class for more folds)

15. Classification
Three-fold cross validation (too few recordings per class for more folds)
Three machine-learning algorithms:
J48 decision trees
Multi-layer perceptron (MLP)
Support vector machine (SVM)

16. Experimental results (1)
15 classes (species and queen/worker type) Two classes > 95 %, five > 80 %
Algorithm \ Features
MFCC
LPC
J48
56 %
MLP
60 %
SVM
64 %
57 %

17. Experimental results (2)
8 classes (species)
Algorithm \ Features
MFCC
LPC
J48
51 %
48 %
MLP
50 %
49 %
SVM
52 %

18. Experimental results (2)
8 classes (species) 2 classes (queen/worker type)
Algorithm \ Features
MFCC
LPC
J48
51 %
48 %
MLP
50 %
49 %
SVM
52 %
Algorithm \ Features
MFCC
LPC
J48
90 %
88 %
MLP
93 %
87 %
SVM
96 %
91 %

19. Conclusion
Recognition of bumblebees by their buzzing sound with accuracy 64 %
More data needed

20. Conclusion In the future:
Recognition of bumblebees by their buzzing sound with accuracy 64 %
More data needed
In the future:
Web application to classify unlabeled data
Accept submissions of new labeled data