Fig. 7
From: MetaFlux: Meta-learning global carbon fluxes from sparse spatiotemporal observations
Site-level validation errors for (a) GPP and (b) Reco across differentiable models (MLP, LSTM, and BiLSTM) that are meta-trained (orange lines) or not (blue lines). The shaded regions represent the standard deviation of RMSE across 5 model runs and 100 epochs (where a single epoch refers to a complete pass over the entire training dataset). In general, the meta-trained models perform better, and the choice of internal learner matters as demonstrated by the overall lower RMSE of the either LSTM or BiLSTM time models that account for temporal dependency.
