SUMMARY AND CONCLUSIONS

The commonly applied spin-up method for land surface model (LSM) initialization was assessed, under different climate conditions for the spin-up year and using various initial soil moisture states, at a sporadic and a discontinuous permafrost site in the Liard River Basin in Canada. The single-year spin-up technique for 2000 cycles was evaluated to initialize and simulate permafrost dynamics using 1D MESH/CLASS model simulations. The study highlighted that employing a deep soil configuration in LSMs requires adequate attention to the initial states as these play a central role in the rate of spin-up stabilization and the fidelity of subsequent simulation. The initial water storage and its partitioning into liquid and frozen contents can affect the system hydraulic and thermal memories, and consequently, the capability of initializing permafrost in LSMs/ESMs. Previous studies focused on identifying appropriate initial soil saturation for model initialization, which is shown in this study to be insufficient due to the interplay between soil liquid and frozen contents on the quality of spin-up. For instance, five different partitioning conditions of a fully saturated soil column required different spin-up effort to attain stable state-variables. Our analysis shows that model spin-ups based on a back-to-back repetition of 200-1000 cycles could be appropriate for initializing soil temperature and water content profiles under different climate conditions, moisture conditions, and model configurations. Such a conclusion can be extendable to other LSMs/ESMs, given the immense computational resources needed for large-scale applications.
Further, initializing the soil column with near field capacity conditions (25% saturation: 25% liquid + 0% Ice, or as 18.75% liquid + 6.25% Ice) required minimal spin-up effort to form permafrost under different climates. Similarly, the wet climate spin-up year led to the shortest spin-up to initialize permafrost in the deep soil column. On the other hand, utilizing only the annual totals/averages while identifying the initial year’s climate condition could be insufficient and leads to non-representative transient conditions. The selection of the initial year’s climate is challenging as the interplay between the external forcing (i.e. precipitation and air temperature) dominantly control permafrost initialization behaviour in LSM. Considering additional statistical measures is advisable, especially those measuring the seasonal patterns (monthly/seasonal statistics) or using more comprehensive measures such as the coefficient of variation along with the annual totals/averages. Further, it is suggested to avoid any peculiarities around the beginning and the end of the spin-up year to ensure successful initialization of permafrost. The large variations observed in the initialization experiments necessitate assessing the associated impact of the uncertain initial conditions on the simulation.
We analyzed the effect of initialization uncertainty on various soil states at the end of spin-up. The portion of the soil column between the permeable depth (SDEP) and the organic depth (ODEP) showed a high range of variability for frozen water content and soil temperature to uncertainties of model initialization for both setups. Below SDEP, temperature profiles showed a decaying sensitivity to the initial condition perturbation, with no impact at the bottom of the soil column. The magnitude of variability for soil temperature was 4-5°C for the permeable part of the soil column, and 0.4 m3m-3 for the frozen water content down to SDEP. Layers at the ODEP and SDEP interfaces showed significant oscillations in soil liquid and frozen contents due to the abrupt change in soil properties, which requires further modelling efforts to improve the smoothness of transition, reducing numerical issues and enhancing the realism of natural systems’ representation. Further, the initial climate condition has a dominant role in the simulated soil temperature and liquid moisture content. In contrast, the initial water content (and its partitioning into liquid and frozen) had a stronger influence on the formed ice than the initial climate condition.
The assessment also incorporated different aspects that describe permafrost dynamics on annual basis, noting that previous studies on permafrost simulation in LSMs considered limited features of permafrost in their assessments. We selected two performance metrics, the bias in simulated active layer thickness (ɛALT) and the root mean square error (RMSE) of temperature envelopes, to examine the impact of different spinning conditions on the simulation quality. ɛALT showed high dependency on soil-texture, and land-cover parameterizations, with systematic errors in the range of ±1 m observed at the two sites. Also, RMSEs of maximum and minimum annual temperature envelopes (Tmax and Tmin) varied by ~1.5 °C and ~0.75°C at the two sites, noting that the two sites yielded poorer RMSE of Tmin compared to Tmax. The mean annual ground temperature at the permafrost table (MAGTp) showed a stronger response to the driving climate over initial soil storage components, ranging between 2-3 °C annually at the two sites. Examining the temporal evolution of freezing/thawing cycles highlighted the high variability of the date of maximum thaw (ALT-DOY), shifting by up to four weeks between August and October. The depth of the zero-annual temperature amplitude (DZAA) and the depth to the base of permafrost (BP) exhibited similar responses to the initialization’s uncertainty, as the results indicate considerable variability to the initial soil moisture, with a minor impact of the initial climate condition, having a magnitude of variability of three- to four-fold among all the designed experiments.
Notably, modelers employ different initialization techniques to generate self-consistent model states, which are assumed sufficient for the subsequent simulation once it attains quasi-equilibrium. The main assumption at the start of model initialization is the presence of a quasi-equilibrium with the external forcing. However, the atmospheric climate has been transient over the last millennium (Mann et al. , 1999) and is in strong disequilibrium with the ‘transient’ ground thermal regime at decadal-to-millennial scales (Zhang et al. , 2008b). In the current study, we followed the same conventional approach of assuming an equilibrium state at the end of the successful spinning. However, the study showed that there are several self-consistent states, generated under different initial conditions, which would yield divergent simulations of permafrost. This outcome raises the fundamental issue of attempting to initialize models to a steady state while the real system is transient, which yet has no simple resolution.
To conclude, our study accentuated the importance of LSM initialization for permafrost-related analysis, which could alter state-variable stabilization and, therefore, the simulation itself. The work assessed the propagation of initialization uncertainty on different aspects characterizing permafrost dynamics and underscored the huge variability in permafrost simulation. In terms of simulation quality, the two setups were able to produce Tmax envelopes and ALTs in reasonable agreement with observation, which is not the case for Tmin envelopes that were colder than observed. The relatively poor simulated cold soil envelope (Tmin) suggests inadequate surface insulation that could be attributed to the quality of snow simulation, which can be addressed through integrating a multi-layer snow scheme (e.g. JULES: Burke et al. , 2013), a complex canopies module (e.g. CLASS-CTEM: Melton et al. , 2019), and/or representing the lateral migration of heat/moisture fluxes (e.g.Noah-MP: Aas et al. , 2019). Therefore, further development is needed in MESH/CLASS to elevate the realism of permafrost simulations, and consequently, the hydrologic and climate simulations. Future work can be directed towards generalizing our analysis outcomes to other observational sites in other permafrost regions/classes, and extension to different regional and global models with varying complexity levels in large-scale applications. Lastly, assessing the influence of LSM parameters on simulated permafrost through a comprehensive sensitivity analysis is recommended in light of the large impact of initial conditions on LSM permafrost simulation.