Unsupervised Learning for Spectral Data Analysis as a Novel Sensor for Identifying Rodent Infestation in Urban Environments

This 2017 IEEE SENSORS paper applies unsupervised learning to reflectance spectra of mouse urine on floor surfaces, compared with human urine and coffee. It is a laboratory test of spectral sensing for rodent detection.

Presented at the 2017 IEEE SENSORS conference in Glasgow by Omar Costilla-Reyes, Zachary Coldrick, and Bruce Grieve, the paper comes from Dr. Costilla Reyes’s appointment as an Agri-Sensor Research Associate at the University of Manchester from August 2016 to August 2017. That project asked whether hyperspectral imaging could detect rodent activity through appropriate chromophores, the light-absorbing molecules that give a material its spectral signature, at a sensitivity sufficient for rapid intervention. Rodent urine is known to fluoresce, and the study uses spectral measurements of that signature to tell urine from other liquids. The paper appears as an appendix of his PhD thesis, which acknowledges funding from the EPSRC Impact Acceleration Account and rodent urine samples supplied by Rentokil Ltd.

The thesis appendix describes reflectance measurements: a fibre-optic probe and a spectrometer record how much light a dried spot of liquid reflects at each wavelength. Samples were placed on ceramic tiles, a carpet, and filter paper, as stand-ins for floors where rodents may be found. Unsupervised learning looks for groups in data without being told the labels in advance. Here K-means clustering assigns each spectrum to the nearest of a set of cluster centres, while t-SNE and multidimensional scaling are manifold techniques that project long spectra into two dimensions so the groups can be seen. The experiments clustered mouse urine across the different surfaces, then compared mouse urine with human urine and filtered coffee, which look alike to the eye, and finally repeated that comparison with Gaussian noise added to the signals.

The thesis reports clusters that separated the samples cleanly in the first two experiments with the best initialization, and the noise experiment examined how the choice of clustering initialization shapes the grouping when the signals are degraded. Laboratory measurements on controlled samples let the spectral signature of mouse urine be studied directly, and the clear separation from human urine and coffee shows that this signature carries enough information for an unsupervised method to find it without labels. Rodent infestation is costly and hard to monitor in large buildings, so a sensor that reads a spectrum in place of a visual inspection is attractive. The authors propose carrying the approach into a real urban setting, such as a warehouse floor, as the natural next step.

It shares with Dr. Costilla Reyes’s gait-analysis papers the use of pattern recognition on raw sensor signals, and it was made available on IEEE Xplore in December 2017 alongside his IEEE SENSORS 2017 paper on age-sensitive differences in walking tasks.