A Deep-Learning Hybrid-Predictive-Modeling Approach for Estimating Evapotranspiration and Ecosystem Respiration

Gradual changes in meteorological forcings (such as temperature and precipitation) are reshaping vulnerable ecosystems, leading to uncertain effects on ecosystem dynamics, including water and carbon fluxes. Estimating evapotranspiration (ET) and ecosystem respiration (RECO) is essential for analyzing the effect of climate change on ecosystem behavior. To obtain a better understanding of these processes, we need to improve our estimation of water and carbon fluxes over space and time, which is difficult within ecosystems where we have only sparse data. In this study, we developed a hybrid predictive modeling approach (HPM) that integrates eddy covariance measurements, physically-based model simulation results, meteorological forcings, and remote sensing datasets to estimate evapotranspiration (ET) and ecosystem respiration (RECO) in high space-time resolution. HPM relies on a deep learning algorithm-long short term memory (LSTM) – as well as direct measurements or outputs from physically-based models. We tested and validated HPM estimation results at sites within various mountainous regions, given their importance for water resources, their vulnerability to climate change, and the recognized difficulties in estimating ET and RECO in mountainous regions. We benchmarked estimates of ET and RECO obtained from the HPM method against measurements made at FLUXNET stations and outputs from the Community Land Model (CLM) at Rocky Mountain SNOTEL stations. At the mountainous East River Watershed site in the Upper Colorado River Basin, we explored how ET and RECO dynamics estimated from the new HPM approach vary with different vegetation and meteorological forcings. The results of this study indicate that HPM is capable of identifying complicated interactions among meteorological forcings, ET, and RECO variables, as well as providing reliable estimation of ET and RECO across relevant spatiotemporal scales. With HPM estimation of ET and RECO at the East River Watershed, we found that abiotic factors of temperature and radiation predominantly explained ET spatial variability; whereas RECO variability was largely controlled by biotic factors, such as vegetation type. In general, our study demonstrated that the HPM approach can circumvent the typical lack of spatiotemporally dense data needed to estimate ET and RECO over space and time, as well as the parametric and structural uncertainty inherent in mechanistic models. While the current limitations of the HPM approach are driven by the temporal and spatial resolution of available datasets (such as NDVI), ongoing advances in remote sensing are expected to further improve accuracy and resolution of ET and RECO estimation using HPM.