A semi-automated image analysis procedure for in situ plankton imaging systems

Plankton imaging systems are capable of providing fine-scale observations that enhance our understanding of key physical and biological processes. However, processing the large volumes of data collected by imaging systems remains a major obstacle for their employment, and existing approaches are designed either for images acquired under laboratory controlled conditions or within clear waters. In the present study, we developed a semi-automated approach to analyze plankton taxa from images acquired by the ZOOplankton VISualization (ZOOVIS) system within turbid estuarine waters, in Chesapeake Bay. When compared to images under laboratory controlled conditions or clear waters, images from highly turbid waters are often of relatively low quality and more variable, due to the large amount of objects and nonlinear illumination within each image. We first customized a segmentation procedure to locate objects within each image and extracted them for classification. A maximally stable extremal regions algorithm was applied to segment large gelatinous zooplankton and an adaptive threshold approach was developed to segment small organisms, such as copepods. Unlike the existing approaches for images acquired from laboratory, controlled conditions or clear waters, the target objects are often the majority class, and the classification can be treated as a multi-class classification problem. We customized a two-level hierarchical classification procedure using support vector machines to classify the target objects (< 5%), and remove the non-target objects (> 95%). First, histograms of oriented gradients feature descriptors were constructed for the segmented objects. In the first step all non-target and target objects were classified into different groups: arrow-like, copepod-like, and gelatinous zooplankton. Each object was passed to a group-specific classifier to remove most non-target objects. After the object was classified, an expert or non-expert then manually removed the non-target objects that could not be removed by the procedure. The procedure was tested on 89,419 images collected in Chesapeake Bay, and results were consistent with visual counts with >80% accuracy for all three groups.