The timing of vegetation phenological events is fundamental to many ecosystem and physiological processes, such as annual productivity; however, models have been found to be inadequate at correctly modeling the timing of phenological events such as leaf out, leaf senescence, and leaf abscission. Additionally, the high temporal frequency of NDVI measurements from NOAA’s newest Geostationary Operational Environmental Satellite (GOES-16 and -17) improves the monitoring of phenological change at a landscape scale, but has not been previously included in phenology forecasts.
We developed a series of Bayesian hierarchical dynamic models to forecast spring landscape vegetation phenological change (defined as percent canopy growth) at deciduous broadleaf sites based on covariate data (e.g., growing degree days). Forecasts were dynamic, iterative, and updated daily with new observations. For landscape phenology data, we used PhenoCam GCC, MODIS EVI and NDVI, and GOES NDVI data. For covariate data, the ERA5 meteorological product was used for calibration and the GEFS 16-day weather forecast data was used in the forecast. Models were calibrated at a variety of PhenoCam sites and forecasts were made at both calibration sites and out-of-sample sites within the National Ecological Observatory Network (NEON).
Results/Conclusions
We found that our models were able to forecast percent canopy growth better than non-covariate models (random walk and a basic logistic growth model) for all sites. Forecasts were better at the sites used for calibration than they were at the out-out-sample NEON sites and we were able to adequately forecast the start of leaf out approximately a week in advance at the calibration sites. We found that process uncertainty contributed the most to the overall uncertainty in our forecasts across the different sites. We also found how the inclusion of the GOES NDVI data affected the forecast depended on the spatial heterogeneity of the landscape and the mismatch between the percent canopies as measured by satellite and PhenoCam.