Estimation of Water Depths and Turbidity From Hyperspectral Imagery Using Support Vector Regression

We propose and evaluate an empirical method for water depth determination from hyperspectral imagery when the benthic layer is visible using support vector regression (SVR). The implementation of the empirical method is presented, and its ability to estimate water depths is compared with a more commonly used band ratio method for two distinct fluvial environments. Our analysis shows that SVR outperforms the band ratio method by providing better root-mean-square error (RMSE) agreement and higher R² for both clear and turbid water. We also demonstrate an extension of the nonparametric properties of SVR to provide estimates of water turbidity from hyperspectral imagery and show that the approach is able to estimate turbidity with an RMSE of approximately 1.2 NTU when compared with independent turbidity measurements.