Modelling Landsurface Time-Series with Recurrent Neural Nets

Machine learning tools and semi-empirical models have been very successful in describing and predicting instantaneous climatic influences on the spatial and seasonal variability of biosphere state and function. Yet, little work has been carried to explicitly model dynamic features accounting for memory effects, where in some cases hand-designed features (e.g. temperature sum, lagged precipitation) have been employed. Here, we explore the ability of recurrent neural network variants (RNN, LSTM) to model time series of dynamic variables 1) fPAR and NDVI, and 2) Carbon dioxide uptake and evapotranspiration, with meteorological variables as the only dynamic predictors. We show that the recurrent neural net approach excellently deals with this dynamic modelling challenge and outcompetes approaches where hand-designed features are complicated to conceive.