Comparative analysis of various classification models for the early detection of marine microalgae using their fluorescence signatures

Microalgae species identification is crucial for the ecological monitoring of harmful algal blooms (HABs) since their frequency and severity have increased in the marine environment and pose significant threats to aquaculture. For this study, we selected three bloom-forming microalgae species- Alexandrium tamarense, Lingulodinium polyedra and Pseudo-nitzschia fraudulenta that are known to affect aquaculture farms. We analysed the spectral data from the algal organic matter (AOM) and chlorophyll regions to develop a multiclass classification model to identify marine microalgae from single and mixed algal samples. Additionally, we compared the performance of four machine-learning (ML) algorithms: support vector machine (SVM), convolutional neural network (CNN), random forest (RF) and K-nearest neighbour (KNN) in classifying these species using the three-dimensional (3D) spectral data from their pure cultures. Data augmentation techniques were applied to enhance the models’ performance, and we evaluated each model trained on the original and augmented data. Our results showed that the CNN trained on the augmented dataset achieved the highest accuracy of 93% for the single algal samples, outperforming the other models. Additionally, the CNN achieved a component accuracy of 0.88 and a Hamming loss of 0.1905 from the mixed algal sample, indicating the potential of the CNN algorithm in identifying these species from mixtures with similar spectral features. Overall, our results indicate that effective identification of marine microalgae can be achieved by combining AOM spectral signatures with ML algorithms. This approach can serve as a promising tool for managing HABs and aid in developing early warning systems for aquaculture farms.