Image data from: Prototyping Deep Transfer Learning Classification of Holographic Plankton Imagery

Data used for the training and evaluation of Convolution Neural Networks (CNN) for the accompanying paper.

Paper Abstract:
The Continuous Plankton Recorder (CPR) is an autonomous plankton sampling platform towed by volunteer vessels along their standard routes, facilitating long-term, monthly sampling across extensive geographic areas. The integrated Continuous Plankton Recorder (iCPR) enhances the traditional CPR survey with new data by utilising the CPR sampler as a platform to deploy new technologies in modular self-powered packages, including high-resolution digital plankton imaging through a bespoke holographic camera. This paper presents the development of a computational pipeline for the preprocessing and classification of iCPR image data using a transfer learning method with Convolutional Neural Networks (CNNs). Prototype models trained on eight classes commonly observed by the iCPR were used to evaluate the performance of several pipeline configurations of CNN types and transfer learning methods. Additionally, novel image preprocessing methods using additional image output unique to holography to supplement standard intensity images to enhance classification performance were also explored. While results show incorporating additional hologram image information did not improve classification performance, high model accuracy and F1-score (>97%) was achieved with multiple combinations of CNNs and transfer learning methods. Findings from this research provides the groundwork for further expansion and development of the methodology used in the pipeline in moving towards operational classification of in-situ plankton images from routine iCPR deployments and subsequently, the ability to provide the CPR survey with a new generation of ecological metrics.