Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
5754660 | Remote Sensing of Environment | 2017 | 18 Pages |
â¢A model is developed to estimate sea surface salinity from MODIS measurementsâ¢The model is selected among many possible choices of solutionsâ¢The model is evaluated extensively using data from the northern Gulf of Mexicoâ¢Model uncertainties and sensitivity to input parameter errors are quantifiedâ¢Model applicability and limitations are discussed, and executable is made available
Sea surface salinity (SSS) is an important parameter to characterize physical and biogeochemical processes, yet its remote estimation in coastal waters has been difficult because satellite sensors designed to “measure” SSS lack sufficient resolution and coverage, and higher-resolution ocean color measurements suffer from optical and biogeochemical complexity when used to estimate SSS. In the northern Gulf of Mexico (GOM), this challenge is addressed through modeling, validation, and extensive tests in contrasting environments. Specifically, using extensive SSS datasets collected by many groups spanning >Â 10Â years and MODIS (Moderate Resolution Imaging Spectroradiometer) and SeaWiFS (Sea-Viewing Wide Field-of-View Sensor) estimated remote sensing reflectance (Rrs) at 412, 443, 488 (490), 555, and 667 (670) nm and sea surface temperature (SST), a multilayer perceptron neural network-based (MPNN) SSS model has been developed and validated with a spatial resolution of ~Â 1Â km. The MPNN was selected over many other empirical approaches such as principle component analysis (PCA), multi-nonlinear regression (MNR), decision tree, random forest, and supporting vector machines (SVMs) after extensive evaluations. The MPNN was trained by a back-propagation learning technique with Levenberg-Marquardt optimization and Bayesian regularization. The model showed an overall performance of root mean square error (RMSE)Â =Â 1.2, with coefficient of determination (R2)Â =Â 0.86, mean bias (MB)Â =Â 0.0, and mean ratio (MR)Â =Â 1.0 for SSS ranging between ~Â 1 and ~Â 37 (NÂ =Â 3640). Validation using an independent dataset showed a RMSE of 1.1, MB of 0.0, and MR of 1.0 for SSS ranging between ~Â 27 and ~Â 37 (NÂ =Â 412). The model with its original parameterization has been tested in the Mississippi-Atchafalaya coastal region, Florida's Big Bend region, and in the offshore Mississippi River plume, with satisfactory performance obtained in each case. Comparison with concurrent Aquarius-derived SSS maps (110-km resolution) showed similar agreement in offshore waters as indicated above, but the new 1-km resolution SSS maps revealed more finer-scale features as well as salinity gradients in coastal waters. The sensitivity of the model to realistic model input errors in satellite-derived SST and Rrs was also thoroughly examined, with uncertainties in the model-derived SSS being always <Â 1 for SSSÂ >Â 30. The extensive validation, evaluation, and sensitivity test all indicated the robustness of the MPNN model in estimating SSS in most, if not all, coastal waters and offshore plumes in the northern GOM. Thus, the model provided a basis for generating near real-time 1-km resolution SSS maps from satellite measurements. However, the model showed limitations when applied to regions with known algal blooms or upwelling as they both led to low Rrs in the blue bands that may be falsely recognized as caused by low SSS.