Hybrid Predictive Model: A hybrid deep learning and physically based model approach for large scale and long term estimation of water and carbon fluxes

Understanding water and carbon fluxes are of great significance for analyzing climate change's effects on watershed functionality. Gradual climate changes in weather forcing (such as temperature and precipitation) are reshaping vulnerable watersheds, leading to uncertain effects in water and ecosystem quality and quantity. To obtain a better understanding of these processes, it is crucial to improve estimation of water and carbon fluxes in watersheds throughout the world. In this study, we develop a novel data-driven approach, namely a Hybrid Predictive Model (HPM), which uses weather forcing data (including air temperature, precipitation and radiation) and remote sensing derived data (i.e., NDVI) as inputs to estimate evapotranspiration (ET) and ecosystem respiration (RECO) over large spatiotemporal scales. This approach, potentially circumvents the parametric uncertainty inherent in physically based models, including those associated with soil and vegetation parameters. We validate HPM results at various FLUXNET sites within the Rocky Mountain regions, and benchmark the method against output from the Community Land Model at three Rocky Mountain SNOTEL sites. At the East River Watershed in the Upper Colorado River Basin, we also explore how ET and RECO dynamics, estimated using HPM, vary with different dominant vegetation and changing climate. In general, we find that the use of HPM can facilitate large scale estimation of water and carbon fluxes, and advance our understanding of the linkages among hydrological dynamics, ecological functioning and changing climate.