Advances in Land Surface Model-based Forecasting: A comparative study of LSTM, Gradient Boosting, and Feedforward Neural Network Models as prognostic state emulators

TL;DR

The paper tackles the computational burden of land-surface state forecasting by evaluating three surrogate emulators—an LSTM encoder–decoder, XGB, and an MLP—trained on ERA5-forced ECLand simulations to emulate seven prognostic states with a physics-informed, multi-objective loss. The study finds that LSTM excels in continental long-range forecasts, XGB provides robust, consistent performance across variables, and MLP offers the best runtime-accuracy trade-off, with all models achieving substantial speedups over the full LSM. This work demonstrates the feasibility of fast offline experimentation and potential integration into data-assimilation workflows for land-surface forecasting, while highlighting region- and depth-dependent strengths and weaknesses. The results underscore the value of combining memory-enabled and memory-less surrogates to balance accuracy, interpretability, and computational efficiency in earth-system emulation.

Abstract

Most useful weather prediction for the public is near the surface. The processes that are most relevant for near-surface weather prediction are also those that are most interactive and exhibit positive feedback or have key role in energy partitioning. Land surface models (LSMs) consider these processes together with surface heterogeneity and forecast water, carbon and energy fluxes, and coupled with an atmospheric model provide boundary and initial conditions. This numerical parametrization of atmospheric boundaries being computationally expensive, statistical surrogate models are increasingly used to accelerated progress in experimental research. We evaluated the efficiency of three surrogate models in speeding up experimental research by simulating land surface processes, which are integral to forecasting water, carbon, and energy fluxes in coupled atmospheric models. Specifically, we compared the performance of a Long-Short Term Memory (LSTM) encoder-decoder network, extreme gradient boosting, and a feed-forward neural network within a physics-informed multi-objective framework. This framework emulates key states of the ECMWF's Integrated Forecasting System (IFS) land surface scheme, ECLand, across continental and global scales. Our findings indicate that while all models on average demonstrate high accuracy over the forecast period, the LSTM network excels in continental long-range predictions when carefully tuned, the XGB scores consistently high across tasks and the MLP provides an excellent implementation-time-accuracy trade-off. The runtime reduction achieved by the emulators in comparison to the full numerical models are significant, offering a faster, yet reliable alternative for conducting numerical experiments on land surfaces.

Figures (4)

• Figure 1: LSTM architecture. Blue shaded area indicates the encoder part, where the model is driven by a lookback $\lambda$ of meteorological forcing and state variables. The light-blue shaded area indicates the decoder part that is initialized from the encoding to unroll LSTM forecasts from the initial time step $t$ up to a flexibly long lead time of $\tau$.
• Figure 2: a: Total aggregated distributions of (log) scores averaged over lead times, i.e. displaying the variation among grid cells. b: The distribution of the anomaly correlation in space on the European subset (b.1: XGB, b.2: MLP, b.3: LSTM). c: Model forecasts over test year 2021 for grid cell with minimum and maximum RMSE values (LSTM).
• Figure 3: a Total average scores, representing spatial variation among grid cells. b Total average ACC in space. Note that ACC remained undefined for regions of low signal in snow cover and soil water volume, see Appendix.
• Figure 4: Emulator forecast skill horizons in two European subregions, aggregated over prognostic state variables. Scores are computed with the anomaly correlation coefficient (ACC) at 6-hourly lead times (y-axis) over approx. one year, displayed as a function of the initial forecast time (x-axis). As horizon we define the time at which the forecast has no value at all, i.e. when ACC is 0 (or below 10%). The diagonal dashed lines indicate the day of the test year 2021 as labelled on the x-axis, the arrows indicate where forecasts reach the second test year 2022.