Estimating inland water quality parameters using Wyvern Dragonette-001 hyperspectral imagery, a case study from the St. Lawrence River, Canada

Abstract. This study evaluates the feasibility of using Wyvern Dragonette-001 hyperspectral imagery for retrieving inland Water Quality Parameters (WQPs), turbidity, Suspended Sediments (SS), and Dissolved Organic Carbon (DOC), in a section of the St. Lawrence River, Québec, Canada. Two field campaigns were conducted on 2 August and 6 September 2024, providing in-situ measurements for model development and validation. Water-leaving reflectance (ρ_w) was derived using ACOLITE, and a spatial blocking approach was applied to reduce spatial autocorrelation between training and testing datasets. Random Forest (RF) models were trained with hyperparameter tuning using Optuna and 5-fold cross-validation. Results show good performance for turbidity estimation (R² = 0.91), while SS and DOC exhibited lower accuracy in compared to turbidity, partly due to limited sample sizes. Feature importance analysis indicated that green bands were most relevant for DOC, near-infrared (NIR) bands for SS, and red bands for turbidity. Elevated ρ_w values in NIR bands suggest potential influences from adjacency effects, atmospheric correction limitations, or sensor characteristics. Overall, Dragonette-001 demonstrates capability for turbidity mapping, while improvements in spectral coverage, atmospheric correction, and in-situ data availability are needed for accurate retrieval of DOC and SS.