Evaluating Transferability and Robustness of Process-Guided Neural Networks in Forest Carbon Flux Modelling

Abstract

Making robust and generalizable predictions within ecological systems such as forests remains challenging due to limited data availability and the slow pace of environmental change. To address this, we integrate a semi-empirical environmental process model (PRELES) to support deep learning approaches, specifically artificial neural networks (ANNs). We replicate and extend previous work on process-guided neural networks (PGNN) by introducing new model types and conducting a comprehensive hyperparameter optimisation within systematic nested cross-validation analyses in both data-thinning and extrapolative scenarios. Results show that both data-driven ANNs and PGNNs consistently outperform the stand-alone process model, while PGNNs provide additional advantages over ANNs in data-sparse settings and under transfer scenarios to unseen, changing climatic conditions. We further estimate the generalisation error for data-driven models as a function of the amount of training data, allowing for guidance on model suitability under different data availability. A variable importance analysis using accumulated local effects reveals that both PGNNs and ANNs learn simple, physically plausible relationships, whereas PRELES exhibits a strong bias toward boreal conditions and limited ability to predict unseen, climatically divergent sites. HighlightsO_LIProcess-guided, and plain neural networks outperform a calibrated process-based model (PRELES) in predicting forest ecosystem carbon fluxes. C_LIO_LIProcess-guided neural networks provide advantages over naive neural networks in sparse-data settings and show greater robustness under transferable scenarios with unseen changing climatic conditions. C_LIO_LIVariable-importance analyses using accumulated local effects show that both process-guided and naive neural networks learn simple yet physically plausible relationships between meteorological drivers and target responses, whereas the process model (PRELES) exhibits a better fit toward boreal conditions and limited ability to predict unseen, climatically divergent sites. C_LI