Vegetation phenology, the seasonal timing of leaf growth and senescence, strongly influences land–atmosphere exchanges of carbon, water, and energy. Accurate representation of phenology in land surface models is critical for predicting ecosystem–climate interactions but is limited by threshold-based approaches in the Community Land Model (CLM), which rely on static temperature and photoperiod thresholds. These schemes struggle to capture interannual variability, nonlinear responses to climate extremes, and environmental legacy effects from prior years.

We address this by integrating a transformer-based phenology model directly into CLM5. Pretrained on 40 years of satellite-derived solar-induced fluorescence (CSIF) and climate reanalysis, the model predicts daily vegetation activity while dynamically responding to multi-scale hydroclimatic drivers. Its attention-based interpretability framework identifies the historical conditions most responsible for the onset and end of the growing season, producing global maps of phenological climate controls and their long-term trends.

This integration improves growing season interannual anomaly prediction, and enhances the model’s ability to capture biome-specific climate-vegetation sensitivities in Earth system simulations.