Fig. 1: Comparison of the dPL framework to the traditional calibration paradigm.
a A deep learning model is trained to mimic the outputs of a process-based model (PBM). This step is optional since one may also directly implement the model in a DL platform. b Workflow of the first dPL option, network gA: parameters are inferred by a network (in our case, a separate LSTM network) based on auxiliary attributes. These parameters are then sent into the PBM, whose outputs are compared to the observations to calculate the loss (the difference between objective function and observation). c Workflow of the second dPL option, network gZ: historical observations (meteorological forcings and observed responses) are additional inputs to the parameter estimation network. d Traditional site-by-site parameter calibration framework.
