LAPS Plankton Detector: A User-Friendly Computer Vision Tool for Automatic Plankton Identification

In-situ imaging systems are becoming a widespread tool to study spatial distributions and temporal dynamics of both phyto- and zooplankton communities. Such systems generate a massive volume of data, demanding the development of computer vision and machine learning techniques to assist in the automatic image classification process. In this study, we present the software suite named “LAPS Plankton Detector (LPD)”, designed as a user-friendly interface for ecologists who wish to automatically classify particles and organisms from raw plankton images generated from both in-situ and benchtop instruments. This freeware interface provides a comprehensive pipeline of typical computer vision tasks, with different configurations, from ROI detection and image segmentation to final automatic classification based on feature extraction (Random Forest and SVM) and Convolutional Neural Networks. Several trained classifiers were implemented by varying the number and size of classes, set of features, data augmentation methods, and CNN architectures. The LPD was applied to analyze time-series plankton image datasets acquired in the coastal area of Ubatuba (SE Brazil) and off Oshima Island (Japan), using different imaging instruments. Depending on the total number of classes, classifier accuracy varied from 75% to 94% using Random Forest, and from 85% to 96% with Deep Learning algorithms, in both cases considering cross validation approaches. However, when classifiers were applied to predict unknown plankton image samples collected over long-term time series, accuracy declined over time to 72 - 89%, demanding a continuous update of both training set and classifiers. Such limitation is likely to decrease with the acquisition of additional plankton images, particularly for the less-abundant taxa. We conclude that the LPD provides a full set of computer vision resources for routine automatic plankton identification tasks.