Fine-scale Gross Primary Productivity for Murray-Darling Basin wetlands

A seasonal basin–wide layer of terrestrial Gross Primary Productivity (GPP) of water-dependent vegetation has been developed at monthly timestep and 20-meter resolution. The layer was created using an innovative approach that combines a process-based radiative transfer model, Sentinel-2, machine learning and weather observations (Wolanin 2019). The product provides an enhanced ability to assess changes in floodplain terrestrial productivity over space and time at a Basin scale. Terrestrial GPP is highly variable across space and time in the Basin with inundation, rainfall and temperature (season) all influencing observed patterns. Generation of a Basin wide terrestrial GPP product represents the first step in creating a Basin scale remotely–sensed estimation of floodplain litter load. The layer can be used to inform both potential benefits (e.g., additional carbon for food web support) of watering actions and risks (e.g., areas that may pose hypoxic blackwater risks when inundated). \nLineage: A machine learning algorithm was employed for the prediction of gross primary productivity (GPP), alongside the utilization of a lookup table (LUT) and the Soil Canopy Observation, Photochemistry and Energy fluxes (SCOPE) model. This model orchestrates the simulation of GPP and canopy reflectance, integrating leaf biophysical parameters, plant physiological insights, vegetation structure, and meteorological records.\n\nTo establish the training dataset for the machine learning model, a comprehensive LUT was constructed, encompassing diverse scenarios achieved by varying input parameters. Eventually, the LUT was comprised of 412,000 sets of input parameters, utilized to simulate GPP and reflectance at the canopy's zenith, aligned with the bands present in Sentinel-2 surface reflectance data. Subsequently, a database encompassing GPP, surface reflectance, and meteorological data was employed to train the machine learning model.\n\nReflectance includes bands of B2 (442nm, blue), B3 (492 nm, green), B4 (664 nm, red), B5 (704 nm, vegetation red edge), B6 (740 nm, vegetation red edge), B7 (782 nm, vegetation red edge), B8 (832 nm, near-infrared), B8A (864 nm, narrow near infrared), B11 (1613 nm, short wave infrared spectral range), B12 (2202 nm, short wave infrared spectral range). Rad is the incoming shortwave radiation (W m-2), and temp is the air temperature (˚C). For the creation of the training dataset, a random selection (70%) was drawn from the simulation outcomes of the SCOPE model, while the remaining data served as a testing dataset for the evaluation of the random forest model's performance. To mitigate the risk of overfitting and enhance the predictive precision of the random model, a ten-fold cross-validation approach was employed to identify the optimally performing trained model.