Identification and counting of Pacific oyster Crassostrea gigas larvae by object detection using deep learning

Abstract Natural seedling collection is widely used in the culture of various bivalve species. For successful natural seedling collection, collectors must be installed when larvae appear in the water column at a stage immediately before attachment. Aquaculture farmers generally identify target larvae by morphological features through microscopic examination in a time- and labor-expensive exercise, which also requires a level of expertise to ensure accurate larval identification. We develop a deep-learning-based object-detection technique that ultimately might reduce the time and effort required to accurately identify and count Pacific oyster larvae, render their identification more consistent, and negate the need for expertise. Images of plankton net samples collected in Matsushima and Sendai bays, Japan, were taken using a new photographic device with a CMOS image sensor. Images of oyster larvae identified by an expert were used to create a library of labeled images to train a deep-learning model, which proved to be 82.4% accurate in precision, 90.8% in recall, and 86.4% in F-measure. A further method for estimating larval shell height from the rectangular shape of oyster larval images is also developed. The standardized mean difference in shell height between measurements and estimates is 3.3%. This deep-learning model has the potential to significantly reduce the time and effort required to identify oyster larvae in plankton samples, and thereby costs of this exercise.