Application of machine-learning techniques toward the creation of a consistent and calibrated global chlorophyll concentration baseline dataset using remotely sensed ocean color data

This paper introduces a machine-learning approach to satellite ocean color sensor cross calibration. The crosscalibration objective is to eliminate incompatibilities among sensor data from different missions and produce merged daily global ocean color coverage. The approach is designed and investigated using data from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard of the Terra satellite and Sea-viewing Wide Field-of-view Sensor (SeaWiFS). Data from these two sensors show apparent discrepancies originating from differences in sensor design, calibration, processing algorithms, and from the rate of change in the atmosphere and ocean within 1(1/2) h between sensor imaging of the same regions on the ground. The discrepancies have complex, noisy, and often contradictory time and space variabilities. Support vector machines are used to bring MODIS data to the SeaWiFS representation where SeaWiFS data are considered to exemplify a consistent ocean color baseline. Support vector machines are effective in learning and resolving convoluted data relationships between the two sensors given a variety of bio-optical, atmospheric, viewing geometry, and ancillary information. The method works accurately in low chlorophyll waters and shows a potential to eliminate sensor problems, such as scan angle dependencies and seasonal and spatial trends in data. The results illustrate that MODIS and SeaWiFS differences are noisy and highly variable, which makes it difficult to extrapolate the cross-calibration knowledge onto new time and space domains and to define representative global ocean color datasets for support vector machine training.