Towards long-term monitoring of forest phenology using Landsat time series

Monitoring phenology over multiple decades is crucial for understanding how forest productivity responds to climate change. In this regard, satellite remote sensing is indispensable to capture land surface phenology (LSP) at regional to continental scales. Sensors with high spatial or temporal resolution, such as combined Landsat/Sentinel-2 or MODIS time series, have been used to estimate annual LSP. However, these time series either cover relatively short periods (300,000) using the Framework for Radiometric Correction for Environmental monitoring (FORCE), including radiometric and topographic corrections as well as cloud and cloud shadow masking. To isolate phenological changes, we excluded forest pixels affected by disturbances including windthrow, fire, bark beetle outbreaks, or harvest. For the remaining undisturbed pixels, long-term phenological parameter distributions were first estimated from the full 40-year time series using a double-logistic Bayesian model. These parameter distributions were subsequently used as informative priors in a Bayesian hierarchical framework to estimate start (SOS), peak (POS), and end (EOS) of season from sparse annual observations, while accounting for regional tree species composition. Our two-stage modelling approach enables robust annual phenology estimation across the full Landsat era, including data-sparse early decades, and provides a basis for analyzing long-term forest phenology dynamics at continental scales.