An Ensemble Approach to Retrieving Water Quality Parameters from Multispectral Satellite Imagery

This research proposes an approach based on the ensemble of multiple empirical algorithms to better estimate chlorophyll-a (Chl-a) concentration as the proxy for algal blooms in inland waters from Sentinel-2A (S2A) multispectral imagery. Recent studies on Chl-a retrieval from newly launched S2A satellite imagery mainly use a single/individual empirical algorithm. So far, no research effort has been made to synthesize the information From different empirical models for improving the estimation accuracy. Using S2A image data and coincident extensive in situ water truth data, we compared a number of empirical algorithms, among which 2BDA, 3BDA, and NDCI yield the lowest RMSE of 4.50, 4.57, and 4.71μ/L respectively. An optimally weighted combination ensemble method by pseudoinverse technique reduces the RMSE to 4.071μ/L. Our proposed spectral space partition guided ensemble method that exploits advantages of each individual algorithm can further reduce the RMSE to 3.57μ/L, 20.6% improvement over the best individual algorithm.