Emulating ecological memory with recurrent neural networks

Ecosystem processes are driven both by contemporary and antecedent environmental and land surface conditions through ecological memory effects. This chapter provides an insight into the relevance of memory effects in the Earth system and presents an experimental case study to use an Recurrent Neural Network (RNN) model to emulate a physical model. In addition to introducing an experimental design suitable for such purposes, we demonstrate that an RNN is largely capable of learning the memory effects encoded in a physical model. A non-temporal fully connected model cannot reproduce such memory effects, especially during anomalous conditions (e.g. climate extremes).