OOS 4-9
Predicting phenology: A case-study in real-time ecological forecasting
Phenological transitions have large impacts on ecosystem processes, species interactions, and climate. However, phenology is a critical source of uncertainty in projections of climate change on terrestrial ecosystems and the current generation of ecosystem models are highly variable and biased in their phenology predictions. Most phenological modeling has focused on diagnosing phenological variability and predicting long term responses to climate scenarios. Phenological predictions for the current season, on the other hand, are being made based on long-term means or expert opinion rather than real data.
To our knowledge previous research has not applied operational data assimilation approaches to produce operational, real-time forecasts of phenology. We present a phenology forecast data product that is automatically updated every day using current observations and weather forecasts. Specifically we fuse MODIS NDVI and PhenoCam based GCC with a logistic process model at five sites across eastern forests, from North Carolina to New Hampshire. Prior to application, models were calibrated (2000-2012) using a Bayesian state space model. Forecasts for fall 2013 and spring 2014 were then generated on a daily basis using a particle filter.
Results/Conclusions
The forecast system successfully tracked seasonal phenology but showed high sensitivity to tower GCC observations. Despite success in describing phenological data, both in our training data and in previous studies, the logistic model proved to be a poor choice for predicting phenological transitions. Preliminary results suggest the logistic was biased toward early transitions and could only predict phenological transitions less than 7 days in advance. Other process models are currently being explored. Work remains to be done to extend this work to a fully spatial context. In particular there is a need to determine the spatial range of influence of the tower PhenoCam data and to account for both land cover and random effects. More broadly, this work demonstrates the possibilities for the development of real-time ecological forecasting in other areas. doi:10.6084/m9.figshare.941067