LAUSR

In this article, we propose a universal data-driven model to acquire FLUXNET-consistent annual forest gross primary productivity and net ecosystem exchange globally. The model is developed based on a deep-learning network with a time series of seven ecological and climatic parameters as inputs. To avoid tedious data downloading for large-area studies, we build this model on the Google earth engine platform with all of the input data available online. A multidimensional convolutional block is adopted to detect meaningful variation patterns between forest growth and the environment. The patterns are then encoded and adjusted with forest attributes to obtain the final estimation through a multilayer perceptron. This special working mechanism enables the model to understand and adapt to different modes of forest carbon dynamics. The new model behaves more like a human than conventional machine-learning models, which directly retrieve estimations from raw input variables. Multistep transfer learning is implemented to make the model robust to a large portion of data gaps and to obtain a balanced performance for different climates and ecological zones. The experimental results demonstrate that the model can achieve estimations that are highly consistent with the FLUXNET records. By visualizing the activation outputs of intermediate layers in the convolutional block, we show that the model can reasonably reflect key influence factors in different periods of a year for various forest types. This result means that we may be able to better understand forest carbon absorption by learning how this model works to obtain correct estimations.