Improving the Estimation of Nitrogen and Phosphorus Concentrations in Lakes and Reservoirs Using a Stacked Approach

Abstract A comprehensive and accurate estimation of water quality in lakes and reservoirs is vital for the protection of the aquatic biota. Research on the spatiotemporal variations of nitrogen (N) and phosphorus (P) concentrations in lacustrine systems is typically plagued, however, by a lack of long‐term, spatially continuous monitoring data. This paper assembled a 30‐year (1989–2018) data set of water quality in 586 lakes and reservoirs in China, along with basin characteristics and climate conditions, forming the comprehensive data set available. These data were then used in a stacking model (based on random forest, support vector regression, and K‐nearest neighbor models) to identify the relationships between nutrient concentrations and their influencing factors, including net anthropogenic N/P inputs, geographical position, climate, land use pattern, and soil type. The stacking models were developed using data collected over multiple time scales (annual, seasonal, and monthly), which were then applied to reconstruct TN and TP concentrations during the periods of 1980–2018 and 2020s–2050s under the climate scenarios of RCP 4.5 and RCP 8.5. The accuracy of the stacking models was 99.1% and 98.3% for TN and TP concentrations using ensembled data, respectively. The interannual variations in TN and TP contents in the 586 lakes and reservoirs during 1980–2018 exhibited a non‐monotonic pattern with a peak of 1.12 and 0.049 mg/L in 2007, respectively. This study demonstrates that stacking machine learning models represent a new effective approach for estimating nutrient concentrations in unmonitored lakes and reservoirs across broad spatiotemporal scales.