Data Study Groups are week-long events at The Alan Turing Institute bringing together some of the country’s top talent from data science, artificial intelligence, and wider fields, to analyse real-world data science challenges. Plankton image classification Cefas (The Centre for Environment, Fisheries, and Aquaculture Science) is an agency of Defra (the Government’s Department of Environment, Food and Rural Affairs) and world leading experts in marine and freshwater science. Research at Cefas aims to tackle the serious global problems of climate change, marine litter, overfishing, and pollution to secure a sustainable future for marine ecosystems. The Cefas Endeavour, a multi-disciplinary research vessel, collects millions of plankton images during its surveys through the Plankton Imager (PI) system: a high-speed imaging instrument which continuously pumps water, takes images of the passing particles, and attempts to identifies the zooplankton organisms present (Figure 1). Images have varying shapes and sizes with a highly-skewed distribution towards smaller particles/images. Of these, over 80 percent can be classified as detritus (e.g., sand, seaweed fragments, microplastics) which are traditionally removed by-eye before any analysis, leaving the remaining plankton images to be manually labelled. The challenge dataset consisted of 58,791 TIF (Tag Image File Format) images of individual objects detected and segmented in imagery collected on the RV Cefas Endeavour research vessel using the PI system. Approximately 17,000 of these images are of individual zooplankton. The plankton images had previously been manually classified by experts into two main categories: Copepods, small or microscopic aquatic crustacean of the large taxonomic class Copepoda (see Figures 25 and 26), and Non- Copepods (see Figures 23, 24, 27, 28), for all other plankton not belonging to the Copepoda class. The experts also categorised these images further into 38 species classes. This expert manual classification allowed challenge participants to verify the accuracy of the automated classification methods explored. The number of images varied greatly between the 38 classes, ranging from 4000 images to 10 images per class. Challenge participants therefore had to decide how to address this imbalance in order to produce a model that could be useful and accurate classifications of plankton. The remaining 40,000 images consisted of individual pieces of detritus (see Figures 29 and 30). These images were of other objects collected by the RV Cefas Endeavour PI system such as sand, seaweed, or microplastics. Manual removal of these images has been shown to be a significant bottleneck in the analysis of imagery collected using the PI. Therefore as an additional challenge, participants had the opportunity to explore automated sorting of images into plankton and detritus in order to facilitate application of plankton classification models to imagery collected from the PI in real time without pre-processing to remove these erroneous objects.