Exploiting Machine Learning to Develop Ocean Color Retrievals From the Tropospheric Emissions: Monitoring of Pollution Instrument

Abstract Retrievals of ocean color (OC) properties from space are important for understanding the ocean ecosystem, the carbon cycle, and monitoring events such as harmful algal blooms (HABs). The recently launched U.S. National Aeronautics and Space Administration (NASA) Earth Venture Instrument, the geostationary Tropospheric Emissions: Monitoring of Pollution (TEMPO), provides a unique opportunity to examine diurnal variability in ocean ecology across coastal waters of North America and prepare for future hyperspectral geostationary OC missions. Although TEMPO does not match the spatial resolution or spectral coverage of planned coastal ocean sensors, such as NASA's Geosynchronous Littoral Imaging and Monitoring Radiometer or the U.S. National Oceanic and Atmospheric Administration Geostationary Extended Observations Ocean Color Instrument, it provides hourly observations at approximately 5 km over U.S. coastal regions and the Great Lakes. Here, we apply a newly developed atmospheric correction approach based on principal component analysis combined with machine learning (ML) to retrieve OC properties using TEMPO's hyperspectral measurements. Principal component coefficients derived from measured reflectances are used to train a neural network to estimate OC properties, including chlorophyll concentration, informed by collocated physically‐based retrievals from MODIS, VIIRS, and Ocean and Land Color Instrument. This ML‐based approach complements traditional radiative transfer retrievals, particularly under challenging conditions such as glint and moderate cloud coverage. This approach demonstrates the value of near‐real‐time OC products, with significant potential for monitoring HABs and transient oceanic phenomena.