Constraining forest dynamics in LPJ-GUESS through data assimilation of forest inventory data using LAVENDAR

The world’s forests take up ca. 30% of anthropogenic carbon emissions. However, despite the huge importance of this sink, we lack any direct method to measure it at large scales. All routes to estimate the forest carbon sink involve modelling of some kind. Dynamic vegetation models have been widely used in such estimations. These models, especially when they consider forest demography, have a clear advantage in providing a fully self-consistent digital representation of the forest, allowing all fluxes and stocks to be interrogated and drivers to be directly diagnosed. However, they also suffer from biases due to missing or simplified process representations or the use of coarse-scale or global parameters, which fail to capture the local-scale heterogeneity. These biases are particularly marked when it comes to the rates of tree growth and mortality, which are central to the forest carbon sink. Forest Inventory data offer a unique opportunity to constrain these parameters, as they provide repeated, spatially extensive observations of forest structure, growth and mortality across large regions. Here we present an approach to use forest inventory data to bias correct dynamic vegetation model simulations, to generate a hybrid product which combines the advantages of both methods.

The proposed framework uses multi-census forest inventory data to refine the performance of the latest version of the LPJ-GUESS dynamic vegetation model across temporal scales. Firstly, LPJ-GUESS employs a newly developed state initialisation method to initialise the simulated forest stands with the observed forest structures from the earliest available forest inventory census. Secondly, we integrate the Land Variational Ensemble Data Assimilation Framework (LAVENDAR) with LPJ-GUESS to assimilate observed growth rates and mortality, thereby calibrating two sets of model parameters that constrain the growth and mortality processes of the model. Specifically, we adopt a two-stage assimilation approach that not only maximises the utilisation of forest inventory data to reduce overall model bias, but also better captures temporal forest dynamics.

We apply this framework to repeated forest inventory data from Sweden and evaluate its impact on simulated forest growth and mortality. The results show that this framework improves the agreement between simulated and observed growth increments and mortality rates, while also enhancing the temporal responsiveness of the model to interannual variability. Compared to the original LPJ-GUESS configuration, the proposed framework enables a regionally and temporally adaptive parameterisation, leading to more realistic calculations of forest dynamics and carbon fluxes.

This study demonstrates the potential of combining repeated forest inventory data with advanced data assimilation techniques to provide assessments of forest carbon dynamics. The proposed two-stage framework is generic and can be extended to other regions and inventory systems, as well as integrated with complementary information from remote sensing, offering a promising pathway towards data-constrained, temporally adaptive and rapidly updatable assessments of forest carbon dynamics across large scales.