Automatic Classification of Plankton from Digital Images

1
Automatic Classification of Plankton from Digital
Images
M Sieracki1, E Riseman2, W Balch1, M Benfield3,
A Hanson2, C Pilskaln1, H Schultz2, C Sieracki4,
P Utgoff2, M Blaschko2, G Holness2, M Mattar2, D
Lisin2, B Tupper1 Bigelow Laboratory for Ocean
Science1 Computer Vision Lab, U. Mass.
Amherst2 Louisiana State University3
Fluid Imaging Technologies4
Marine particles, including plankton and
non-living particles, play important roles in
ecosystem function and material flux in the
oceans. Digital imaging technology used in
instruments to study these particles can rapidly
produce huge archives of images that require
expert interpretation. Automated methods to
assist the expert interpret these images are
urgently needed. We are building automatic
classifier systems to work with the experts to
efficiently and accurately classify images of
marine particles. We will use images from
in-situ camera instruments (e.g. VPR) for
zooplankton and marine snow, an imaging-in-flow
system (FlowCAM) for phytoplankton, and digital
fluorescence microscopy for pico- and
nanoplankton. Experiments were conducted using
low resolution FlowCAM images of 13 classes of
phytoplankton from natural communities, and a
variety of image features and classifiers,
including classifier ensembles. These
preliminary tests yielded classification accuracy
of over 70, compared to published human expert
agreement of about 80. This indicates that
automated classification will be practical to
automate the majority of images. We intend to
develop a probabilistic approach to particle
enumeration, and to test the generality of our
classifiers across instrument types.
Classification Methods K-Nearest
Neighbors Decision Trees Naïve Bayes Ridge
Regression Support Vector Machines
Expert Classified Image Sets
1 Video Plankton Recorder (VPR)
Test Image Sets
Label 1
Experts manually classify particles
2 FlowCAM Imaging-in-flow
Label 2
3 Epifluorescence Microscopy
Label 3

• Conclusions
• Combinations of shape and texture features
  performed best
• Support Vector Machine classifier performed best
• Best accuracy was 73, approaching consistency
  rate of human experts (80)

Preliminary Results for FlowCAM Images

• Experiments
• 980 expert labeled FlowCAM images
• 780 total features
• 5 classifiers used

• Future Work
• Apply to other image types, more expert
  classified image sets
• Automated feature selection
• 3D FlowCAM (dual aspect angle images)
• Experiments with local image features
• Software tools for experts