Interpretable multi-morphology and multi-scale microalgae classification based on machine learning

Abstract

The multi-morphology and multi-scale mixed microalgae are widely distributed in natural and artificial systems. There is an urgent need to develop an efficient approach to classify the mixed microalgae for natural water system monitoring and microalgae bioprocesses, such as wastewater treatment, carbon dioxide capture and prevention of harmful algal blooms. The numerical feature datasets of pure and mixed cultures of multi-morphic microalgae with a size range between 5 and 500 μm are established in the study. A large number of input features increases model complexity and computational costs, and the feature space dimension was reduced from 24 dimensions to 11 dimensions using the Pearson coefficient matrix and principal component analysis to reduce the impact of unimportant factors. Research indicates that the classification performance of the ensemble model is significantly better than that of the linear and nonlinear models. The average F1_score of the random forest optimized by grid search classified pure and mixed microalgae are 0.952 and 0.943, respectively, which are 2.2 % and 1.0 % higher than those without optimization. The Shapley Additive exPlanations theory and the ensemble model are combined to analyze the critical factors for microalgae classification, and the texture features play a crucial role in all the numerical features of microalgae images.