Spatiotemporal variation reconstruction of total phosphorus in the Great Lakes since 2002 using remote sensing and deep neural network

Total phosphorus (TP) is non-optically active, thus TP concentration (CTP) estimation using remote sensing still exists grand challenge. This study developed a deep neural network model (DNN) for CTP estimation with synchronous in-situ measurements and MODIS-derived remote sensing reflectance (Rrs) (N = 3916). Using DNN, the annual and intra-annual CTP spatial distributions of the Great Lakes since 2002 were reconstructed. Then, the reconstructions were correlated to nine potential factors, e.g., Chlorophyll-a, snowmelt, and cropland, to explain seasonal and long-term CTP variations. The results showed that DNN reliably estimated CTP from MODIS Rrs, with R2, mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and root mean squared logarithmic error (RMSLE) of 0.83, 1.05 μg/L, 2.95 μg/L, 9.92%, and 0.13 on the test set. The near-surface CTP in the Great Lakes decreased significantly (p < 0.05) during 2002−2022, primarily attributed to cropland reduction, coupled with improvements in basin natural ecosystems. The sensitivity analysis verified the model robustness when confronted with input feature changes < 35%. This result along with the marginal difference between CTP derived from two sensors (R2 = 0.76, MAE = 2.12 μg/L, RMSE = 2.51 μg/L, MAPE = 11.52%, RMSLE = 0.24) suggested the model transferability from MODIS to VIIRS. This transformation facilitated optimal usage of MODIS-related archive and enhanced the continuity of CTP estimation at moderate resolution. This study presents a practical method for spatiotemporal reconstruction of CTP using remote sensing, and contributes to better understandings of driving factors behind CTP variations in the Great Lakes.