Estimation of Chlorophyll-a Concentration in the Nakdong River Using Sentinel-2 MSI and Algal Bloom Influence Factors Data Fusion

Abstract

Due to global warming, the frequency and extent of algal blooms are increasingly prevalent worldwide. In Korea, the Nakdong River faces severe algal bloom issues every year. Algal blooms cause various damages such as ecological, economic, and aesthetic damages, periodic monitoring is essential for preemptive management and rapid response. Chlorophyll-a (chl-a) concentration is utilized as an indicator of algal bloom occurrences, and the use of satellite enables the detection of algal blooms over extensive areas. Numerous studies recently have utilized machine learning techniques to achieve more accurate estimations of chl-a concentrations. Various factors affect algal blooms occurrence, and identifying the cause of their occurrence remains challenging. Therefore, it is essential to apply a study using diverse input data to comprehensively consider these factors. This study fused Sentinel-2 satellite data with water quality, meteorological, and hydrological factors data to estimate chlorophyll-a concentrations for eight weirs along the Nakdong River over the last five years. AutoML selected six models (CatBoost, Extra Trees, Gradient Boosting, LightGBM, Random Forest, and XGBoost), and the SHAP method identified a total of 27 fused input variables. CatBoost (R2 = 0.862, RMSE = 5.560 mg/m3, MAE = 4.120 mg/m3) demonstrated superior performance, and all six models achieved significant results with R2 above 0.8. SHAP was conducted to analyze the importance of features, and Suspended Solids (SS) emerged as the most important factor in all six models. The ranking of variable importance varied by model, water quality variables such as Dissolved Oxygen (DO), pH, Total Organic Carbon (TOC), and Biochemical Oxygen Demand (BOD), and satellite variables using the band combinations of red-edge and red band were identified as common top-ranking variables. The feasibility of chl-a estimation was assessed by exhibiting the spatial patterns of the estimated chl-a values using CatBoost. This study confirmed the applicability of fusion data for estimating chl-a concentrations, and it is expected to be utilized for nationally and globally chl-a monitoring in the future.