Downscaling Sentinel-3 Chlorophyll-a Concentration for Inland Lakes Based on Multivariate Analysis and Gradient Boosting Decision Trees Regression.

Downscaling Chlorophyll-a (Chl-a) concentration derived from satellite image is crucial for refined applications such as water quality monitoring. However, the precision of downscaling is usually constrained by various environmental factors. In this paper, we develop a downscaling method for Chl-a concentration to improve precision, especially for inland lakes with different surrounding environment. The method downscales the Sentinel-3 Chl-a concentration from 300 m to 30 m, based on the multivariate analysis (MVA) and the gradient boosting decision tree (GBDT) model. Firstly, we analyzed 21 Chl-a concentration related indices to identify optimal factors for Chl-a concentration variability. Secondly, a GBDT model is constructed to convey the non-linear relationship between the optimal factors and Chl-a concentration at coarse resolution. Finally, fine-resolution Chl-a concentrations were produced by employing the model to refine cofactors for 12 distinct lakes. The results indicated that the proposed MVA-GBDT method effectively inferred the variability of Chl-a concentration with a mean RMSE of 4.505 mg/m(3), an improvement of 5%-39% over other methods. Furthermore, for lakes with large water quality heterogeneity, the method led to a cross validation RMSE and a difference in accuracy of 5.371 mg/m(3) and 0.866 mg/m(3), respectively. In addition, this study examined the significance of the auxiliary factors and found that the NDCI and WST were the two most important factors for MVA-GBDT to detect Chl-a concentration distributions, particularly for NDCI in lakes with high nutrient contrasts. These findings contribute to the generation of fine-scale Chl-a concentrations in lakes and support related applications.