Data-driven modelling of canopy greenness dynamics reveals short- and long-term meteorological effects on phenology

Vegetation phenology, measured as the seasonal canopy greenness signal, is highly sensitive to present as well as past meteorological conditions. However, how these meteorological conditions affect canopy greenness on the short-term and the long-term (memory effects from previous climatic conditions) is still unclear, and modeling these effects on vegetation phenology in particular is a major challenge. In this study, we develop data-driven models to identify the influence of short- and long-term memory effects of temperature, radiation and water availability on the canopy greenness using data-adaptive approaches, such as random forest regression (RF) models and Long Short-Term Memory (LSTM) setups. We use the Green Chromatic Coordinate (GCC) from the PhenoCam network as a proxy for canopy greenness and meteorological observations from the DayMet dataset. We find that the importance of these short-term vs. long-term memory effects on canopy greenness differs across the plant functional types. For deciduous forest, roughly the last 10 days of minimum temperature and the photoperiod are identified to be the key drivers of canopy greenness, while in grasslands also the water availability and its long-term memory are important factors in controlling the seasonal course of canopy greenness. Additionally, our results show that an LSTM approach with embedded predictor memory effects outperforms a model without the memory effect (such as RF) in simulating the canopy greenness, and captured memory length varies across meteorological predictors with short temperature and radiation memory and long water memory. Our findings highlight the importance of memory effects of environmental conditions throughout the season across different time scales for canopy greenness and the fundamental role of water availability, often neglected in phenological models. Accounting for these effects in such data-driven approaches opens up new avenues for improving the representation of phenological processes in models, such as Earth system models.