Improving fluoroprobe sensor performance through machine learning

Abstract

Phytoplankton, as a fundamental component of aquatic ecosystems, affects ecosystem dynamics and water quality in lakes. Accurate assessment of phytoplankton abundance and community structure is critical for understanding the health and functioning of lake ecosystems. To monitor phytoplankton dynamics efficiently, fluorescence sensors offer rapid and high-throughput estimates of chlorophyll-a concentrations, a widely used proxy for phytoplankton biomass. Implementation of multi excitation fluorescence probes, such as BBE FluoroProbe (FP), provides additional information on the taxonomic structure of the phytoplankton community based on their accessory pigments. However, applying the FP has faced challenges, leading to inaccuracies in phytoplankton assessment. FP limitations in the lake stem from the fact that the FP is calibrated based on only a few species and mostly on cultures, and that different taxonomic groups possess similar or the same pigments. Here, we use machine learning models to enhance assessment and identification of the major prevailing taxa by FP implementation in the well-studied Lake Kinneret (Sea of Galilee), Israel. We compared Extreme Gradient Boosting, Support Vector Regression (SVR) and Random Forest algorithms to assess community structure based on FP raw data. Phytoplankton species biomass data were collected concurrently using standard inverted microscope counts. The SVR algorithm outperformed the other algorithms in predicting phytoplankton abundance and composition in Lake Kinneret, based on the FP measurements. The SVR model presented lower mean square error in predicting phytoplankton biomass, and extended FP capabilities to identify also dinoflagellates, an important taxonomic group in Lake Kinneret. The approach presented here can be applied in other lakes, to overcome site-specific sensor limitations.

Keywords

• Machine learning
• Phytoplankton
• Fluoroprobe
• Freshwater
• Community
• Ecosystem