A Step Toward Instrument-Agnostic Plankton Classification in the Shared Label Space of IFCB and FlowCam Data

The large range of marine plankton cell and/or body sizes has necessitated the emergence of various automated imaging devices, each tailored to image a different part of the plankton size spectrum. Variations in the format and resolution of the resulting image data have led to the development of bespoke, often instrument-specific automated classifiers. The need to train and validate multiple classifiers complicates working with image data from multiple cameras, and hinders efforts to study the full plankton community. In this study, we discuss the challenges associated with building a multi-instrument classifier, including handling domain shifts between cameras and class imbalances within the individual datasets. We then present a scalable, flexible frameworktowards instrument-agnostic classification, which we explicitlybound to the shared label spaceof two cameras: an Imaging FlowCytobot (IFCB) and a FlowCam. The model was trained in a two-step process, focusing first on individual datasets comprising 145 IFCB and 193 FlowCam categories, and then on a combined dataset of 236 categories. To take advantage of multimodal fusion, we integrated image features with 29 metadata attributes for FlowCam; for IFCB, only image data (with simple synthetic attributes) were available, so fusion was not applied. Within shared label spaces, the results demonstrate fewer low-performing categories (F1< 0.5) when moving from single-instrument training to the combined model, alongside a 4.5% average F1 improvement among the hardest classes, with the multi-instrument model achieving 84.6% test accuracy, and improved representation of rare taxa. These findings indicate that pooling heterogeneous imagery can enhance generalization across instruments for overlapping taxa. However, true generalization would require additional datasets and richer cross-modal alignment (e.g., attention-based conditioning). Our results underscore the potential for scalable cross-platform solutions to advance ecological research, resource management, and environmental conservation efforts.