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

<p>DownscalingChlorophyll-a(Chl-a)concentration de rived 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 mto30m,basedonthemultivariate  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/m3, 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/m3 and 0.866 mg/m3,  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  highnutrientcontrasts.Thesefindingscontributetothegeneration  of fine-scale Chl-a concentrations in lakes and support related  applications. </p>