Application of a convolutional neural network to improve automated early warning of harmful algal blooms.

Abstract

Continuous monitoring and early warning together represent an important mitigation strategy for harmful algal blooms (HAB). The coast of Texas experiences periodic blooms of three HAB dinoflagellates: Karenia brevis, Dinophysis ovum, and Prorocentrum texanum. A plankton image data set acquired by an Imaging FlowCytobot over a decade of operation was used to train and evaluate two new automated image classifiers. A 112 class, random forest classifier (RF_112) and a 112 class, convolutional neural network classifier (CNN_112) were developed and compared with an existing, 54 class, random forest classifier (RF_54) already in use as an early warning notification system. Both 112 class classifiers exhibited improved performance over the RF_54 classifier when tested on three different HAB species with the CNN_112 classifier producing fewer false positives and false negatives in most of the cases tested. For K. brevis and P. texanum, the current threshold of 2 cells.mL-1 was identified as the best threshold to minimize the number of false positives and false negatives. For D. ovum, a threshold of 1 cell.mL-1 was found to produce the best results with regard to the number of false positives/negatives. A lower threshold will result in earlier notification of an increase in cell concentration and will provide state health managers with increased lead time to prepare for an impending HAB.