Spatial Transferability and Temporal Repeatability of Water Quality Remote Sensing Inversion Models for Inland Lakes and Rivers

Previous studies have recognized that traditional empirical water quality remote sensing models are region-specific and lack spatial transferability and temporal repeatability, which prevents the mapping and monitoring inland water quality at a large regional or basin scale. This research evaluates the spatial transferability and temporal repeatability of our multi-predictor ensemble learning model in comparison with traditional empirical models. Our evaluations show that the multi-predictor ensemble model not only substantially improves the prediction accuracy in comparison with traditional individual empirical models, but also has strong spatial and temporal extensibility. The multi-predictor ensemble model calibrated at one specific place and at one certain time can be transferred and re-used in other places and in different time periods with reliable and accurate predictions. The strong spatial and temporal extensibility of the multi-predictor ensemble model is largely attributed to the selective strategy for combining component model results according to the spectral-space partition, which makes the ensemble model dynamically adaptable to a wide range of water conditions over space and time.