This is a supporting webpage to our paper published in Journal of Insect Behavior.

The Inexpensive Sensors 
  • The left figure is one of the cages we used to gather data for this project. 
  • The right figure is a logical version of the setup with the components annotated.
  • More details about this sensor can be found in paper SIGKDD Demo: Sensors and Software to allow Computational Entomology, an Emerging Application of Data Mining (PDF here).


Sensor Data Processing

We used a sliding window that slides through the sensor recording to detect flying insect sound. When an insect flying sound is detected in the sensor recording, the signal corresponding to the flying sound is first extracted from the original recording; then, a noise filter is applied to the signal to remove the background noise; finally, the signal is saved into a one-second long audio clip by centering the insect signal and padding with 0s elsewhere.

  • The top figure is an example of a one-second audio clip containing a insect flying sound generated by the sensor. The sound was produced by a female Culex. stigmatosoma
  • The middle figure shows how the insect sound is saved(archived) in memory for future use. 
  • The bottom figure is the frequency spectrum of the insect sound computed using DFT. 
  • You can download this example data from here.
Listen to the Insect Sound:
  • The sound in the top figure:
    The gadget spec URL could not be found
  • The sound in the middle figure:
    The gadget spec URL could not be found





Insect Classification
  • A Toy Example
  • The task of this example is to classify three species of insects, Cx. stigmatosoma ♀, Ae. aegypti ♀, and Cx. tarsalis ♂. 
  • The top figure shows the histograms of wingbeat frequency of each species.
  • The bottom figure is the Gaussian curves that fit the histograms.
  • You can download the data from here.
  • The code to reproduce the figures is here.
  • The instruction on how to run the code is here.
Classification Accuracy using Different Features
Features Used in Classification wingbeat frequency alone frequency spectrum alone frequency spectrum and flight activity circadian rhythm
Classification Accuracy 81.87% 87.57% 95.23%
  • The table shows the classification performance using different classifiers / features.
  • The code to reproduce the results is here.
  • The instruction on how to run the code is here.

  • Insect Classification with Increasing Number of Species
In this experiment, we demonstrate the effectiveness of our classifier by applying it to datasets with an incrementally increasing number of species, and therefore increasing classification difficulty. We began by classifying just two species of insects; then at each step, we added one more species (or a single sex of a sexually dimorphic species) and used our classifier to classify the increased number of species. 

  • The top figure is the histograms of wingbeat frequencies of all ten species of insects used in this experiment. 
  • The bottom figure is the Gaussian curves that fit the histograms.
  • You can download the data from here.
Classification Accuracy using Different Features
Step Species Added Classification Accuracy        Step Species Added Classification Accuracy
1 Ae. aegypti (male) N/A        6 Cx. quinquefasciatus (male) 92.69%
2 Musca domestica 98.99%        7 Cx. stigmatosoma (female) 89.66%
3 Ae. aegypti (female) 98.27%        8 Cx. tarsalis (male) 83.54%
4 Cx. stigmatosoma (male) 97.31%        9 Cx. quinquefasciatus (female) 81.04%
5 Cx. tarsalis (female) 96.10%        10 Drosophila simulans 79.44%
  • The table shows the classification accuracy measured at each step and the relevant class added.
  • The code to reproduce the results is here.
  • The instruction on how to run the code is here.

  • A Case Study: Sexing Mosquitoes
The insect used here is Ae. aegypti mosquitoes. The task is to separate the females from the males. In addition, we can easily tune a threshold in our classifier to minimize the total cost when the mis-classification cost is asymmetric(for example, misclassifying a female to be a male costs much more than the reverse).
  • The figure shows the sex discrimination accuracies with different numbers of training data using our proposed classifier and the wingbeat-frequency-only classifier.
  • You can download the data from here
  • The code to reproduce the results is here.
  • The instruction on how to run the code is here.
  • The left table is the confusion matrix of the sex Ae. aegypti mosquitoes when the misclassification cost is symmetric (decision threshold is 0.5).
  • The right table is the confusion matrix of sexing the same mosquitoes, but misclassifying a female as a male costs much more than the reverse ( and the decision threshold to be classified as female is tuned to be 0.1 here).
  • The code to reproduce the results is here.
  • The instruction on how to run the code is here.


--------------------------------------------------------------------------------------------------------------------
Last edited by Yanping Chen (http://www.cs.ucr.edu/~ychen053/)