Addressing the Uncertainty in Estimation of Ecological Indicators of the Marine Ecosystem

Remote sensing satellites provide high-resolution (~4 km) routine spatio-temporal measurements of physical and biological parameters for the global oceans. These ocean color measurements span for more than two decades now, bringing us close to monitoring the response of ocean biology to the unprecedented climate change, particularly ocean warming. However, till date it remains a challenge to study the ocean ecosystem response at finer temporal scales (less than a month) due to the presence of gaps (upto 25%, Fig. A) in satellite databoth frequent and persistentespecially during the southwest monsoon (upto 40%, June-September) owing to the persistent cloud cover (low outgoing longwave radiation, Fig. B) over the tropical Indian Ocean. The most common techniques to fill the gaps of missing pixels involve interpolation, filtering, and substitution by mean or medianlimited to single value imputation. Though these might work for data with few missing pixels, but lead to potential biases if the data contains a higher percentage of gaps such as the tropical Indian Ocean. Furthermore, the variability and uncertainty associated with single value imputation techniques are often overlookedan important aspect of gap-filling. We present a methodology of gap-filling using statistical methods (Monte Carlo) and machine learning. These advanced computational machine learning tools aim at multiple imputation which takes uncertainty into consideration by accounting for the dispersion from the true value. Most importantly, this work addresses the uncertainty in the ecological indicators derived using the gap-filled observations thus giving an estimation of the signal to uncertainty ratio. This is also needed to highlight the importance of improving the existing observations at basin to global scale and their need to assist in ecological forecasting using biophysical models. We use our methodology to produce a long-term, reliable gap-free chlorophyll dataset in the Indian Ocean. Though used for ocean color, these techniques can be extended to estimating other parameters of the climate system.