Implementation Strategy and Spatiotemporal Extensibility of Multipredictor Ensemble Model for Water Quality Parameter Retrieval with Multispectral Remote Sensing Data

In previous studies, empirical models have been widely used to map water quality parameters for inland waters using remotely sensed imagery. Due to the complex optical properties of inland waters, empirical models often have limited performances and they cannot be extended or reused across space and time. To overcome the limitations of traditional empirical models, this article further explores the ensemble method for water quality parameter retrieval. Based on the Sentinel-2 multispectral images and extensive coincident in situ water quality data, we examined different implementation strategies and evaluated the extensibility of a multipredictor ensemble model for retrieving chlorophyll-a (Chl-a) across space and time. Our analysis shows that the multipredictor ensemble model can improve the Chl-a prediction accuracy by 46% compared with the best traditional empirical model. Our experiments suggest that the selection of heterogeneous component models improves the performance of the ensemble model and the new iterative K -fold calibration approach results in a considerably better ensemble model than the global calibration approach. To evaluate the spatial extensibility, our ensemble model calibrated for Harsha Lake was applied to Caesar Creek Lake and Brookville Lake. This ensemble model calibrated with the data sets acquired on October 7, 2016 was also applied to Sentinel-2 images of Harsha Lake acquired on October 2, 2017 and August 23, 2018, to evaluate its extensibility for different years. Our evaluation results show that the ensemble model can be transferred and reused across space and time to provide reliable and accurate Chl-a estimates.