Chlorophyll-a and Turbidity estimation using UAV imagery and machine learning in a tropical eutrophic reservoir

Abstract. Unmanned Aerial Vehicles (UAVs) equipped with multispectral sensors offer a promising approach for monitoring water quality in optically complex inland waters, particularly in tropical eutrophic systems. This study investigates the performance of two UAV-mounted sensors, DJI Phantom 4 Multispectral and MicaSense RedEdge-P Dual, for estimating chlorophyll-a and turbidity concentrations in the Ibirité Reservoir, Brazil. Four UAV campaigns were conducted in 2024, alongside in situ measurements, and six regression models were evaluated. Results show that chlorophyll-a was robustly predicted using ensemble algorithms, with the MicaSense sensor achieving the best performance (R² = 0.867, RMSE = 6.72 μg/L). Turbidity estimation was more variable, with linear regression outperforming complex models when using MicaSense data (R² = 0.712). The DJI sensor consistently underperformed, mainly due to limited spectral resolution. Findings highlight the critical roles of sensor configuration, spectral sensitivity, and model selection in UAV-based water quality assessment.