A deep‐learning‐based experiment for benchmarking the performance of global terrestrial vegetation phenology models.

Aim: Vegetation phenology that characters the periodic life cycles of plants is indicative of the interactions between the biosphere and the atmosphere. Robust modelling of vegetation phenology metrics that correspond to canopy development events is essential to our understanding of how plants and ecosystems respond to a changing climate. Given considerable uncertainties associated with vegetation phenology modelling using numerical models, we explore the deep learning approach to predicting the timing of global vegetation phenology metrics. Location: Global. Time period: 2001–2015. Major taxa studied: Deciduous vegetation (DV), stressed deciduous vegetation (SDV), evergreen vegetation (EV). Methods: We developed a one‐dimensional convolutional neural network regression (1D‐CNNR) model with 10 hierarchical structures to model global vegetation phenology using meteorological variables as inputs. The developed deep learning model was evaluated using satellite‐derived phenology metrics (i.e., green‐up, maturity, senescence, and dormancy) and compared with the terrestrial ecosystem model Biome‐BGC (BioGeochemical Cycles). Results: Our experimental results show that the 1D‐CNNR model well captures both the spatial pattern and inter‐annual variation of satellite‐derived multiyear vegetation phenology metrics on a global scale. The median root‐mean‐square errors (RMSEs) and standard deviations between phenology metrics derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) data and predicted by the 1D‐CNNR model on a global scale from 2001 to 2015 are 4.1 ± 5.9, 4.2 ± 12.1, 3.0 ± 6.8, and 3.4 ± 4.3 days for green‐up, maturation, senescence, and dormancy, respectively, for the DV type; 13.3 ± 29.6, 8.4 ± 29.1, 8.1 ± 21.3, and 9.1 ± 21.6 days for green‐up, maturation, senescence, and dormancy, respectively, for the SDV type; and 13.9 ± 17.4, 17.7 ± 34.6, 18.8 ± 42.9, and 12.1 ± 17.7 days for green‐up, maturation, senescence, and dormancy, respectively, for the EV type. Main conclusions: This research demonstrates that the 1D‐CNNR model has the potential for large‐scale modelling of vegetation phenology. Results from the deep learning model suggest that there is room to improve numerical vegetation phenology models for use in land surface models. [ABSTRACT FROM AUTHOR]