Early rapid detection of ichthyotoxic algae based on fluorescent quaternion features via neural network

Abstract

We developed a fast fluorescent ichthyotoxic algae identification (FFIAI) model based on a specially designed convolutional neural network to predict if ichthyotoxic algae exist in water bodies by simply inputting the initial 3D fluorescence spectrum data of an algal testing sample. Discrete excitation-emission fluorescence spectra were used to characterize the fingerprint of each ichthyotoxic algae species and applied to the fluorescent quaternion representation model to form the fluorescent quaternion feature. We trained the FFIAI model with the fluorescent quaternion features of the training database. In our experiment, the well-trained FFIAI model can achieve 100 % accuracy in identifying six species of ichthyotoxic algae, including Amphidinium carterae, Karenia mikimotoi, Phaeocystis globose, Chattonella antique, Heterosigma akashiwo, and Prymnesium parvum. The corresponding detection limits are as low as 1800, 120, 3600, 60, 2000, and 2300 cells/mL, which meets the basic detection requirements of harmful algae blooms (HABs). Moreover, this model also performed well in mixed testing samples, with the identification accuracy of ichthyotoxic algae reaching 90 % when the ichthyotoxic algae are dominant in the mixture. The model could also recognize the ichthyotoxic algae correctly even with a relatively low concentration, with an overall identification accuracy of 75 %, demonstrating its potential for precise monitoring of ichthyotoxic algae in field studies.