Machine learning based downscaling of global satellite SIF products for carbon cycle research

The recent advance in satellite measurements of solar-induced chlorophyll fluorescence (SIF) has motivated a number of studies to monitor terrestrial ecosystem productivity from space. However, none of the existing space-borne instruments with SIF capabilities so far were optimized for observing SIF, thus its full potential for operational application requires resolving multiple technical roadblocks. First, existing SIF datasets have either coarse spatial resolutions or incomplete spatial coverage, limiting their potential for operational applications at scales that can directly benefit stakeholders and inform decision making. Second, existing SIF datasets have coarse temporal resolutions, making it challenging to sufficiently capture the rapid onset of stress dynamics. This project aims to develop a SIF product at fine spatial and temporal resolutions using machine learning technique. SIF retrievals from Global Ozone Monitoring Experiment 2 (GOME-2) native observations were transformed to fine resolutions of 0.05^o and 16-day at global uniform coverage by synergistically exploiting MODIS reflectance. We used two machine learning approaches, i.e., artificial neural network (ANN) and random forest (RF), to train and predict the SIF at high resolution. We assessed the quality of the generated SIF product using various metrics, including R-square, RMSE, regression slopes, and heteroscedasticity. Comparisons with independent air-borne and ground measurements were also conducted.