Machine learning for cyanobacteria inversion via remote sensing and AlgaeTorch in the Třeboň fishponds, Czech Republic.
Journal:
The Science of the total environment
PMID:
38971250
Abstract
Cyanobacteria blooms in fishponds, driven by climate change and anthropogenic activities, have become a critical concern for aquatic ecosystems worldwide. The diversity in fishpond sizes and fish densities further complicates their monitoring. This study addresses the challenge of accurately predicting cyanobacteria concentrations in turbid waters via remote sensing, hindered by optical complexities and diminished light signals. A comprehensive dataset of 740 sampling points was compiled, encompassing water quality metrics (cyanobacteria levels, total chlorophyll, turbidity, total cell count) and spectral data obtained through AlgaeTorch, alongside Sentinel-2 reflectance data from three Třeboň fishponds (UNESCO Man and Biosphere Reserve) in the Czech Republic over 2022-2023. Partial Least Squares Regression (PLSR) and three machine learning algorithms, Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were developed based on seasonal and annual data volumes. The SVM algorithm demonstrated commendable performance on the one-year data validation dataset from the Svět fishpond for the prediction of cyanobacteria, reflected by the key performance indicators: R = 0.88, RMSE = 15.07 μg Chl-a/L, and RPD = 2.82. Meanwhile, SVM displayed steady results in the unified one-year validation dataset from Naděje, Svět, and Vizír fishponds, with metrics showing R = 0.56, RMSE = 39.03 μg Chl-a/L, RPD = 1.50. Thus, Sentinel data proved viable for seasonal cyanobacteria monitoring across different fishponds. Overall, this study presents a novel approach for enhancing the precision of cyanobacteria predictions and long-term ecological monitoring in fishponds, contributing significantly to the water quality management strategies in the Třeboň region.