Abstract. Modelling of terrestrial systems is continuously moving towards more integrated modelling approaches, where different terrestrial compartment models are combined in order to realise a more sophisticated physical description of water, energy and carbon fluxes across compartment boundaries and to provide a more integrated view on terrestrial processes. While such models can effectively reduce certain parameterisation errors of single compartment models, model predictions are still prone to uncertainties regarding model input variables. The resulting uncertainties of model predictions can be effectively tackled by data assimilation techniques, which allow one to correct model predictions with observations taking into account both the model and measurement uncertainties. The steadily increasing availability of computational resources makes it now increasingly possible to perform data assimilation also for computationally highly demanding integrated terrestrial system models. However, as the computational burden for integrated models as well as data assimilation techniques is quite large, there is an increasing need to provide computationally efficient data assimilation frameworks for integrated models that allow one to run on and to make efficient use of massively parallel computational resources. In this paper we present a data assimilation framework for the land surface–subsurface part of the Terrestrial System Modelling Platform (TerrSysMP). TerrSysMP is connected via a memory-based coupling approach with the pre-existing parallel data assimilation library PDAF (Parallel Data Assimilation Framework). This framework provides a fully parallel modular environment for performing data assimilation for the land surface and the subsurface compartment. A simple synthetic case study for a land surface–subsurface system (0.8 million unknowns) is used to demonstrate the effects of data assimilation in the integrated model TerrSysMP and to assess the scaling behaviour of the data assimilation system. Results show that data assimilation effectively corrects model states and parameters of the integrated model towards the reference values. Scaling tests provide evidence that the data assimilation system for TerrSysMP can make efficient use of parallel computational resources for  > 30 k processors. Simulations with a large problem size (20 million unknowns) for the forward model were also efficiently handled by the data assimilation system. The proposed data assimilation framework is useful in simulating and estimating uncertainties in predicted states and fluxes of the terrestrial system over large spatial scales at high resolution utilising integrated models.