Deep Learning for Verification of Earth-System Parametrisation of Water Bodies

About 2/3 of all densely populated areas (i.e. at least 300 inhabitants per km$^2$) around the globe are situated within a 9 km radius of a permanent waterbody (i.e. inland water or sea/ocean coast), since inland water sustains the vast majority of human activities. Water bodies exchange mass and energy with the atmosphere and need to be accurately simulated in numerical weather prediction and climate modelling as they strongly influence the lower boundary conditions such as skin temperatures, turbulent latent and sensible heat fluxes and moisture availability near the surface. All the non-ocean water (resolved and sub-grid lakes and coastal waters) are represented in the Integrated Forecasting System (IFS) of the European Centre for Medium-Range Weather Forecasts (ECMWF) model, by the Fresh-water Lake (FLake) parametrisation, which treats $\sim$ 1/3 of the land. It is a continuous enterprise to update the surface parametrization schemes and their input fields to better represent small-scale processes. It is, however, difficult to quickly determine both the accuracy of an updated parametrisation, and the added value gained for the purposes of numerical modelling. The aim of our work is to quickly and automatically assess the benefits of an updated lake parametrisation making use of a neural network regression model trained to simulate satellite observed surface skin temperatures. We deploy this tool to determine the accuracy of recent upgrades to the FLake parametrisation, namely the improved permanent lake cover and the capacity to represent seasonally varying water bodies (i.e. ephemeral lakes). We show that for grid-cells where the lake fields have been updated, the prediction accuracy in the land surface temperature improves by 0.45 K on average, whilst for the subset of points where the lakes have been exchanged for bare ground (or vice versa) the improvement is 1.12 K. [abridged]