I wanted to raise a few questions around initialization of land use in FATES for global historical runs with land-use change. As we get closer to being able to do this, some of the complexities of doing so are becoming more apparent to me. This is a long post, so no need for rapid response from anyone, but hopefully it can provide some basis for discussion going forward.
Traditionally in the big-leaf model, a global transient historical run has three parts:
Spin-up under 1850 (really early 20th century) climate, 1850 CO2, and constant 1850 land-use with either accelerated decomposition or matrix spinup turned on to equilibrate vegetation and soils
Spin-up under 1850 (really early 20th century) climate, 1850 CO2, and constant 1850 land-use with normal soil dynamics turned on to re-equilibrate vegetation and soils to slightly different dynamics.
Transient dynamics of land-use, climate, and/or CO2 from 1850 onwards
In FATES with land-use change, this is going to have to work differently. There are two reasons for this. The first is that from FATES's perspective there really isn't any such thing as "constant 1850 land-use". Since harvest of any primary land creates a one-way transfer of primary land to secondary land, there can't ever be any steady state as long as there is any harvest from primary land or other nonzero net transition rates. The second reason is that the secondary land is intentionally meant to be out of equilibrium, and with an age structure that reflects its disturbance history. So if we want to start considering model results at, e.g., 1850, we need to initialize the model in such a way that the age structure is already present at that point. The GFDL group, which shares much of the underlying conceptual structure and assumptions of FATES, has been working on this for a long time and written extensively on the point, e.g.: https://doi.org/10.1175/2011EI401.1, which basically says that one needs to initialize land use in the model 100-150 years before the time point where one wants to start looking at the behavior.
So the question I want to pose is on how to set things up to do this for FATES in as simple and reproducible a way as possible. What I think a general workflow would look like might be one of the following general strategies:
Strategy 1
This would be analogous I think to the GFDL paper I linked to above:
Spin-up under 1850 (really early 20th century) climate, 1850 CO2, and no land-use (and, if nocomp configuration is active, a prescribed potential-vegetation distribution map) with either accelerated decomposition or matrix spinup turned on to equilibrate vegetation and soils.
Spin-up under 1850 (really early 20th century) climate, 1850 CO2, and no land-use (and, if nocomp configuration is active, a prescribed potential-vegetation distribution map) with normal soil dynamics turned on to re-equilibrate vegetation and soils.
Transient early land-use period under 1850 (really early 20th century) climate, and 1850 CO2. The way this could work is that:
At the very start of this phase of the initialization, all patches have the label of primary lands. On the first day of entering this mode, the LUH2 land use state vector file is read for the start year (let's say 1700 for the sake of argument), and all transitions of patch area from primary lands to secondary lands are accomplished via a single clear-cut harvest event whose rate will result in the desired area of secondary lands for 1700 at the end of the first day; and all transitions of patch area from primary lands to all other land-use types are accomplished by a set of land-use-change disturbance events whose disturbance rates on that day will result in the desired areas of all other land use types at the end of the day.
If nocomp is turned on, the disturbance events in the above procedure will lead to a change in the nocomp PFT areas on the newly disturbed patches away from potential-vegetation and to the landcover fractions associated with each land-use type. In some cases this would mean adding nocomp patch area for a PFT where none was present before, and so some strategy to seed PFTs so that they can grow will be needed.
Run with transient land-use from the period 1700-1850.
Transient dynamics of land-use, climate, and/or CO2 from 1850 onwards
Strategy 2
The idea here would be to try to create some sort of synthetic 1850-like steady-state land-use conditions and spin up the model under that land use forcing dataset. In principle, I think the way this might work would be to make a file that had a constant land use state vector for 1850 conditions, zero rates of transition between all land-use classes (and thus zero rates of harvest from primary lands), and a set of spatially-varying but temporally-constant secondary young and/or mature harvest rates that were chosen so that after some indefinitely long period of time, the secondary lands age structure for a given gridcell would be similar to what it would be if the entire land-use history up to that date on that gridcell had been applied, including all harvest and abandonment transitions. I am not sure how closely it would be possible to match the secondary age structure in this way, which might be its own science question. But if it were possible to get something that is "close enough", then one could maybe revert to a general initialization strategy that is similar to the way we currently spin up the old vegetation model. I am not sure what the model should do with the wood harvest products created in this stage; since the point is just to find an equilibrium one might want to have all the harvest product go into a fast product pool (but also maybe not). Another point is that if this strategy were pursued, soil carbon etc would reflect this 1850-like state as well, so that would seem to have both advantages and disadvantages relative to what would happen to soil carbon under strategy 1.
Note that some aspects of both strategies 1 and 2 could in principle be used together, so we might want to try to design something that would allow either or both strategies to be pursued, especially since I don't think we really know if strategy 2 would actually work at this point?
Other thoughts/questions/issues
We'll need some approach for nocomp PFT dynamics to prevent PFTs from starting out extinct and/or going extinct on managed land-use types (particularly under strategy 1). Maybe some sort of seed-rain that only occurs on managed lands and maybe also only where things are extinct?
Currently, when land use is not enabled, but FATES is being used in a nocomp configuration, the PFT distributions are read from the present-day data rather than a potential-vegetation map. Should we allow a flag so that if land-use is not on but nocomp is, the user can choose whether to use present-day or potential-vegetation nocomp distributions? If so, I think that the present-day maps are also aligned with the satellite phenology data for SP mode, so would we only allow SP-mode under present-day nocomp data?
How to experimentally measure the effect of land use? Under, e.g., the trendy protocol, land-use effects are assessed by comparing a constant 1850 land-use with a transient land-use. Per the above, if we follow strategy 1, then there isn't an obvious way to make a constant 1850 land-use state. So would we compare a transient land-use case to a no-land-use case? Or does this mean that we would need to do something more like strategy 2 if we want to do such a comparison in a way that is consistent with the old model and/or according to MIP protocols?
What about for coupled CMIP-type cases? Obviously this is getting way ahead of where we are currently at, but could the picontrol case have no-land-use if we use strategy 1, or would we have to follow strategy 2 and make an 1850-like steady state land use dynamics (and associated land cover if run in nocomp)? And (getting even further ahead of ourselves) how would the above sequence need to be modified for emissions-driven cases? The GFDL group must have this all sorted out, but I'm not sure how they do it.
How should we approach all this from a UI perspective? E.g., in the current soil model, when accelerated decomposition (AD) is run, we set a special flag that goes into the restart file so that when a non-AD run is started from an AD run's restart conditions, the special procedure to exit AD spinup is automatically executed on the first timestep, in a way that is invisible to the user and avoids a lot of potential user errors. Should we set an analogous flag in the restart files indicating whether the land-use for the run that generated a given restart file is for a no-land-use versus a land-use configuration, and then, if a run with land-use is initialized from a run with no-land-use, the procedures sketched out in point 3.i in the Strategy 1 section above automatically happen? If so, do we also need to track in the restart file the nocomp setting (i.e., off, present-day, or potential-vegetation), and put in error-handling to disallow certain sequences that wouldn't make scientific sense (like trying to initialize a land-use-on run from a no-land-use run that used a nocomp present-day vegetation distribution)?
Anyway, just wanted to get some of this down; probably others who have been working on land use change for longer than I have have thought about lots of this as well, so please feel free to weigh in.
Hi All,
I wanted to raise a few questions around initialization of land use in FATES for global historical runs with land-use change. As we get closer to being able to do this, some of the complexities of doing so are becoming more apparent to me. This is a long post, so no need for rapid response from anyone, but hopefully it can provide some basis for discussion going forward.
Traditionally in the big-leaf model, a global transient historical run has three parts:
In FATES with land-use change, this is going to have to work differently. There are two reasons for this. The first is that from FATES's perspective there really isn't any such thing as "constant 1850 land-use". Since harvest of any primary land creates a one-way transfer of primary land to secondary land, there can't ever be any steady state as long as there is any harvest from primary land or other nonzero net transition rates. The second reason is that the secondary land is intentionally meant to be out of equilibrium, and with an age structure that reflects its disturbance history. So if we want to start considering model results at, e.g., 1850, we need to initialize the model in such a way that the age structure is already present at that point. The GFDL group, which shares much of the underlying conceptual structure and assumptions of FATES, has been working on this for a long time and written extensively on the point, e.g.: https://doi.org/10.1175/2011EI401.1, which basically says that one needs to initialize land use in the model 100-150 years before the time point where one wants to start looking at the behavior.
So the question I want to pose is on how to set things up to do this for FATES in as simple and reproducible a way as possible. What I think a general workflow would look like might be one of the following general strategies:
Strategy 1
This would be analogous I think to the GFDL paper I linked to above:
Strategy 2
The idea here would be to try to create some sort of synthetic 1850-like steady-state land-use conditions and spin up the model under that land use forcing dataset. In principle, I think the way this might work would be to make a file that had a constant land use state vector for 1850 conditions, zero rates of transition between all land-use classes (and thus zero rates of harvest from primary lands), and a set of spatially-varying but temporally-constant secondary young and/or mature harvest rates that were chosen so that after some indefinitely long period of time, the secondary lands age structure for a given gridcell would be similar to what it would be if the entire land-use history up to that date on that gridcell had been applied, including all harvest and abandonment transitions. I am not sure how closely it would be possible to match the secondary age structure in this way, which might be its own science question. But if it were possible to get something that is "close enough", then one could maybe revert to a general initialization strategy that is similar to the way we currently spin up the old vegetation model. I am not sure what the model should do with the wood harvest products created in this stage; since the point is just to find an equilibrium one might want to have all the harvest product go into a fast product pool (but also maybe not). Another point is that if this strategy were pursued, soil carbon etc would reflect this 1850-like state as well, so that would seem to have both advantages and disadvantages relative to what would happen to soil carbon under strategy 1.
Note that some aspects of both strategies 1 and 2 could in principle be used together, so we might want to try to design something that would allow either or both strategies to be pursued, especially since I don't think we really know if strategy 2 would actually work at this point?
Other thoughts/questions/issues
Anyway, just wanted to get some of this down; probably others who have been working on land use change for longer than I have have thought about lots of this as well, so please feel free to weigh in.