Clustering grid cells in a land surface model

Land surface models such as JULES (Joint UK Land Environment Simulator) are usually run on a rectilinear grid, yielding gridded outputs for variables such as soil moisture and evapotranspiration. JULES also models surface and subsurface water fluxes, and these can be used as inputs to a river routing model to predict river flows. Here we investigate the effect of clustering groups of grid cells into ‘Land Response Units’ (LRUs) in JULES, using a hierarchical multivariate clustering technique to group underlying grid cells together based on characteristics including soil type, elevation and land cover. Using LRUs rather than grid cells has the potential to reduce computational expense as well as providing an alternative to tiling approaches for capturing sub-grid heterogeneity. Here, LRUs are used exclusively in the land surface part of modelling, i.e., separate from river routing.

We investigate the effect of the LRU approach on JULES soil moisture in part of the Thames catchment in the UK, and compare LRU and gridded soil moisture predictions with measurements from the UKCEH COSMOS-UK soil moisture observation network. We find that use of LRUs leads to good soil moisture prediction while reducing computational expense compared to a gridded approach, but that this is strongly dependent on the characteristics used to create the LRUs. We also consider how the LRU approach impacts predicted river flows, and compare routed JULES outputs with observed river flow from a number of NRFA gauges in the catchment. We show that less computationally expensive LRU JULES outputs give similar river flow results to standard 1 km gridded JULES outputs when routed at 1km resolution, and that the LRU approach can outperform gridded river flow predictions when routed at higher resolution.