Performance validation

Aquatic Imager performance validation was conducted on zooplankton data from a 7-minute transect collected around Indian Key in Tampa Bay, Florida, on October 24, 2023.

Classifications assigned by application of our proprietary software were compared to validated classifications assigned by our zooplankton taxonomist. The method involves running a classification using the original (unvalidated) images, determining the accuracy of that run by comparing with the validated images, then adding additional validated images to the training library and conducting a reclassification using the additional information. Ten runs were conducted for each taxon.

For the six taxa displayed above, N = the number of images in the training library for the taxon. The images included in the training library are derived from samples collected from various areas around Tampa Bay and Southeast Florida, not just from the October 24, 2023, sample data included in these plots. The change in accuracy with each sample run represents the change in classification success of the unvalidated images during each run compared with the known (validated) classifications (the ‘confusion matrix’ outcome). The ‘percent of sample added’ reflects the number of images added to the training library for each run relative to the total number of images (N) in the sample (not in the training library). ‘Images in the training library’ is the new number of images in that taxon’s training library following the addition of new validated images.