Discussion Series: Education and Digitalisation

In this talk we will cover some of the technical aspects of the current ECMWF Land Data Assimilation System (LDAS). This will include both the “offline” system used for seasonal forecast initialization and reanalysis experiments and the “operational” system used in the production of forecasts. We will highlight some developments being undertaken on the CERISE (CopERnIcus climate change Service Evolution) project to move towards coupled atmosphere-land data assimilation and increase the utilization of ensemble information within the system. The current system at ECMWF is a Simplified Extended Kalman Filter (SEKF). However, the system does incorporate information from the ECMWF atmospheric Ensemble of Data Assimilations (EDA) to diagnose the Jacobian of the observations operator. We will show how use of the EDA within the specification of prior spread in the SEKF has improved forecast skill. The talk will also highlight other technical developments for the parallelization of the “offline” system using the Dask Python package, drawing inspiration from the Pangeo community.

We report on the development of the Terrestrial Carbon Community Assimilation System (TCCAS), an activity funded by the European Space Agency within its Carbon Science Cluster. TCCAS is built around the newly developed DALEC&BETHY (D&B) terrestrial biosphere model. D&B builds on the strengths of each component model in that it combines the dynamic simulation of the carbon pools and canopy phenology of DALEC with the dynamic simulation of water pools, and the canopy model of photosynthesis and energy balance of BETHY. A suite of observation operators allows the simulation of solar-induced fluorescence, fraction of absorbed photosynthetically active radiation, vegetation optical depth from passive microwave sensors, the slope of the backscatter-incidence angle relationship of an active microwave sensor, and surface layer soil moisture. The model is embedded into a variational assimilation system that adjusts a combination of initial pool sizes and process parameters to match the observational data streams. For this purpose TCCAS is provided with efficient tangent and adjoint code. TCCAS will be released as a community tool. We will present the system and show applications over study regions in Finland and Spain.

To improve agricultural soil health and carbon allocation, there has been an increased push to increase the use of cover crops in cash crop cultivation. These efforts cause challenges for agricultural system modeling, though, as many of the relevant biodiversity processes are not currently included in ecosystem models. Implemented on a test farm site in Helsinki, Finland from 2018 to 2023, the TWINWIN project examines how different combinations of cover crops grown with Barley affect both the productivity as well as the various soil conditions of the plot. The modeling component of this project has been attempting to sidestep the biodiversity model inclusion challenge by instead trying to calibrate specific Barley parameters to mirror how the plots are affected by different cover crops. While the parameter values estimated with this approach might not be the actual ones, they should allow us to better model how the mixed crops perform in different environments and climate conditions. For calibration we used the 4-Dimensional Variational Assimilation (4DEnVar) algorithm provided by Reading University research. As the crop model, we used the Stics crop model developed for both annual and perennial crops. The preliminary results have shown promise in capturing certain trends especially concerning yield, but they have also illustrated how important it is to consider both what the measurements and model variables actually represent when choosing observations to use for calibration.

Droughts form one end of the hydrologic extremes that usually evolve over months to years to reach their peak intensity. However, another category termed as flash droughts evolve and intensify very rapidly under the influence of extreme atmospheric conditions. Such events can have multiple triggering factors, including those driven by precipitation deficits or extreme heat, often leading to distinct mechanisms of progression. The northern Great Plains of the United States recently experienced two such flash droughts in 2016 and 2017. In 2016, an early heat wave during March caused warmer-than-normal conditions, leading to increased evaporative demands. The 2017 drought experienced near-record-low precipitation anomalies causing rapid depletion of soil moisture. Persistent dry soil conditions led to increased vegetation stress, severely impacting crop yield. Using the NASA Land Information System (LIS) framework, we demonstrate that data assimilation (DA) of various remote sensing observations within Noah-Multiparameterization (Noah-MP) model is essential in capturing the progression of these two contrasting flash droughts. Results suggest that during the 2016 drought, characterized by an intense heat wave, assimilation of MODIS-derived leaf area index (LAI) within Noah-MP helped the model to capture elevated transpiration at drought onset followed by declining soil moisture. LAI-assimilated soil moisture anomalies exhibit increases of 7.5% and 7.1% in similarity with Evaporative Stress Index data and U.S. Drought Monitor maps, respectively. However, for the precipitation-deficit-driven drought of 2017, assimilating SMAP soil moisture helped capture the rapid drought intensification and resulted in 6.7% and 8.8% higher similarity with respective datasets. The value in assimilating different variables in the two cases can be attributed to the distinct impacts that the precipitation deficit and heat wave had on the vegetation. The two case studies highlight the overarching need of multivariate DA using remote sensing observations to comprehensively capture the different processes controlling the progression of flash droughts.

The NASA LIS/WRF-Hydro system is a coupled modeling framework that combines the modeling and data assimilation (DA) capabilities of the NASA Land Information System (LIS) with the multi-scale surface hydrological modeling capabilities of the WRF-Hydro model, both of which are widely used in both operations and research. This coupled modeling framework builds on the linkage between land surface models (LSMs), which simulate surface boundary conditions in atmospheric models, and distributed hydrologic models, which simulate horizontal surface and sub-surface flow, adding new land DA capabilities. In the present study, we employ this modeling framework in the Tuolumne River basin in central California. We demonstrate the added value of the assimilation of NASA Airborne Snow Observatory (ASO) snow water equivalent (SWE) estimates in the Tuolumne basin. This analysis is performed in both LIS as an LSM column model and LIS/WRF-Hydro, with hydrologic routing. Results demonstrate that ASO DA in the basin reduced snow bias by as much as 30% from an open-loop (OL) simulation compared to three independent datasets. It also reduces downstream streamflow runoff biases by as much as 40%, and improves streamflow skill scores in both wet and dry years. These results have potential applications for water management and understanding hydrologic extremes. In this presentation, we explore the technical configuration of the LIS/WRF-Hydro system and how we leverage ESMF National Unified Operational Prediction Capability (NUOPC) features to couple the LIS and WRF-Hydro systems. Ongoing work with this system, including applications to study water temperature and improve streamflow drought prediction, are also discussed.

Land-carbon dynamics are determined by not only current conditions, but also past conditions that no longer persist. Quantifying such memory effects is therefore crucial for understanding and predicting the carbon cycle. In this talk, a framework for quantifying environmental and biological memory (Ogle et al., 2015) is introduced. Then, using data from 42 eddy covariance sites across six major biomes, a set of Bayesian statistical models were implemented to quantify the strength, temporal feature, and primary contributors of memory in daily net carbon exchange (NEE; Liu et al., 2019). Memory is important for explaining the land-carbon metabolism, especially in drylands for which it explains approximately 32% of the variation in NEE. The strong environmental memory in drylands was driven by both short- and long- term moisture status. Contrary to common belief, sites of the same biome-type do not exhibit similar environmental sensitivity and memory in their NEE responses. Instead, the strength of environmental memory scales with increasing water stress both within and among major biomes (Fig. 1a), suggesting a potential adaptive response to water limitation. These findings highlight the necessity of considering ecological memory in experiment, observation, and modelling. Finally, I discuss limitations of this approach to memory quantification and introduce some future directions, including causal-state reconstruction (i.e., representing NEE dynamics as internal states of the same causal information evolving through time; Fig. 1b) and identifying anticipatory memory in fluctuating environments

Soils store the largest organic carbon in the terrestrial biosphere, yet representations of soil organic carbon (SOC) present huge uncertainty by different biogeochemical models. Model structure has been identified as a major contributor to inter-model uncertainties in simulating SOC by biogeochemical models. While biogeochemical models with different structures and ad hoc parameterizations simulate diverging spatial patterns of SOC across the globe, here we show that different models simulate converged SOC storage and its related model components after being constrained by the same observational data. We applied the PROcess-guided deep learning and DAta-driven modeling (PRODA) approach to simultaneously inform a linear model (i.e., Community Land Model version 5, CLM5) that features first-order kinetics and a non-linear microbial model that characterizes Michaelis-Menten kinetics in SOC decomposition with the same soil database containing >50,000 global distributed SOC vertical profiles. Two models after being optimized by the PRODA approach agree with each other well on simulating global SOC storage and its spatial patterns. Observational SOC data effectively constrains key model components such as carbon transfer efficiency, baseline decomposition rate, and environmental modification and thus eventually contributes to similar SOC patterns by structurally different biogeochemical models. Moreover, the Michaelis constant in the microbial model after being informed by SOC observations shows to be much larger than its corresponding substrate concentration in SOC decomposition. Thus, the nonlinear microbial model can be well approximated by a simplified first-order structure in simulating SOC at the global scale without sacrificing explanatory power. Our results highlight the importance of fusing observational data with structurally-different biogeochemical models to gain converged representations of the soil carbon cycle and identify the most probable model structure at investigated scale.

Previous studies demonstrated ASCAT backscatter and slope contain valuable information about plant water dynamics (Steele-Dunne et al., 2019). In this study, ASCAT normalized backscatter and slope are jointly assimilating into the ISBA-A-gs land surface model (LSM) to constrain plant water dynamics processes. A Deep Neural Network (DNN) is trained following approaches in Shan et al., 2022, as an observation operator embedded in an Extended Kalman Filter (EKF). Data Assimilation (DA) and model open loop (OL) experiments are ran on ASCAT grid points (GPIs) containing ISMN stations in Western Europe and validated using data from 2017 to 2019. Performances of DA and OL are evaluated against ISMN in-situ soil moisture observations at different soil depths and satellite-based LAI from the 1 km v2 Copernicus Global Land Service project (CGLS).

Overall DA shows neutral improvements in domain median values when considering unbiased Root Mean Square Error compared to OL against the observations. However, improvement is observed at specific times of year. For example, analysis of the monthly performances in Agricultural GPIs shows that DA corrects deeper soil moisture in spring. This echoes our previous studies which demonstrated an indirect link between deeper soil water availability and vegetation water status revealed by ASCAT slope. Potential reasons for worse performances at other times are also analyzed. Validation performances of DNN might have an effect on DA performances. Other reasons include that OL reaches already satisfactory ubRMSE (<0.05 m3m-3) comparing to ISMN observations. In addition, it is important to note that DA is performed at the spatial resolution of ASCAT (25 km), while the ISMN provides point-scale information. Analysis of DA performance statistics future efforts can be done to improve the robustness of DNN.

One of the challenges in land data assimilation is accounting for the carbon stock trajectories. To initialise land surface models, a two-step spin-up is often performed. The first step helps to put the prognostic variables, including vegetation station, soil carbon pools, and soil moisture at equilibrium under pre-industrial CO2 conditions. The second step is a transient run which allows for simulations to start with the correct atmospheric conditions. However, since these runs are computationally costly, they are often neglected when performing model calibrations. Instead, it is common to rely on a steady-state relaxation parameter to correct the magnitude of the initial soil carbon content during the optimisations. This presentation will discuss the challenges encountered when using such a parameter based on our experiences with the ORCHIDEE land surface model. One challenge concerns the interpretability of such a parameter when used at a site or regional scale. Another is how adding the nitrogen cycle to ORCHIDEE impacts the effectiveness of this parameter. Finally, we will discuss future avenues and whether we are in a position to remove this parameter from our carbon optimisations with the rise of emulators.

The maximum carboxylation rate (Vcmax) is an important parameter for the coupled simulation of gross primary production (GPP) and evapotranspiration (ET) in terrestrial biosphere models such as the Canadian Land Surface Scheme Including Biogeochemical Cycles (CLASSIC). Vcmax is often parameterized for plant functional types (PFTs). Assuming that Vcmax is constant in time and space introduces large uncertainties in simulated GPP and ET. Using eddy covariance observations made at eight mature boreal forest stands of North America, we optimized Vcmax25 (Vcmax at 25 °C) for six tree, shrub, and herb PFTs in CLASSIC with a Bayesian algorithm. Results showed substantial reduction in root mean square deviation for GPP and ET at almost all and two permafrost-free sites, respectively, when comparing the optimized simulations with several corresponding gridded estimates. The optimized PFT-Vcmax25 compared well with reported estimates from field observations. Remarkable Vcmax25 variations were identified among sites for shrub and herb PFTs. Variations in PFT-Vcmax25 closely associated with site conditions in latitude, air temperature and start and end of growing seasons.

Land surface model parameter calibration efforts typically utilize a surrogate model or emulator for parameter optimization due to the computational expense of simulations. Emulators generally are trained to represent the relationship between the high dimensional parameter inputs and some measure of model error. Defining the error that parameter calibration exercises aim to minimize is a critical task in the optimization process. Often, the error is a combination of multiple variables merged over space and time (referred to hereafter as ‘total error’). Building a robust emulator that accurately represents the relationship between input parameters and the total error can be a challenging task and often results in discrepancies between the emulated and modeled total error. This is partially due to the degradation of the physical relationships between parameter settings and model processes when emulating the total error. Here we experiment with an alternative approach designed to simplify the emulation task and preserve physical relationships between input parameters and modeled processes. We train a suite of emulators to predict specific model output variables for particular biomes. We then calculate the total error using the emulated values and identify model parameterizations that minimize error. We performed this experiment in the Community Land Model and compared simulations using the resulting ‘optimal’ model parameterizations identified using the two approaches. This comparison demonstrates the importance of experimental design choices made throughout the parameter calibration process.