Estimation of global satellite-derived chlorophyll-a as function of sea-surface height using shallow neural networks

Herein we apply Nonlinear Autoregressive models with exogenous Inputs (NARX) to estimate the interannual variability of satellite-derived chlorophyll-a (CHL) at a global scale, as function of sea-surface height (SSH) from a satellite product provided by Copernicus. A previous analysis shows that SSH is one of the top drivers of CHL in key regions of the tropical and south Atlantic, which is herein corroborated at a global scale, showing a significant CHL-SSH correlation in most of the world ocean between 60°S and 60°N (where the most continuous data series are available). This correlation, generally low for a linear estimation, opens the possibility to CHL reconstruction using higher-performance non-linear techniques like NARX. Herein the NARX model was generated with 10 neurons in the hidden layer, trained with a Levenberg-Marquardt algorithm, and applied to the CHL and SSH monthly composites from Oct 1997 to Sep 2024. A noise level of 0.57 for the model correlations was defined as the 95^th percentile of 10,000 NARX-modeled random series. This noise level is exceeded by 97% of the CHL-anomaly series modeled for the 1997-2024 period. The NARX-model successfully reproduces the CHL interannual variability: 59% of the modeled CHL present correlations > 0.90. Then, the NARX-model can be potentially used to predict CHL beyond the training period. In this study’s next stage, the predictability of CHL will be evaluated using SSH for a post-training period, and an ultimate goal for the NARX-model will be a predictability assessment using numerical-model predictions. Thus, the proposed method opens the possibility for reconstruction and prediction not only for CHL but also for other related biogeochemical variables.