Spatial time series for state and transition models

[LukeMcPhan]

Hi All,

I’m currently working with a colleague on a project where we are attempting to model transitions of vegetation condition based on an annual time series of inundation. We have raster layers for inundation, vegetation type and condition at a location and, transition pathways for vegetation condition based on type and inundation history (e.g. prolonged drought, frequent flooding).

The way the package functions currently appears to be very heavily geared toward probabilistic processes and my colleague and I have had trouble finding examples using the software where a time series of disturbance events (e.g. fire/harvesting) can be prescribed for individual pixels in a deterministic way, based on their spatial location.

For example, in a simple 4 cell space (c1:c4),
c1 c2
c3 c4

over 3 years (t1:t3) the following occurs:
• Cell one receives fire in year 1 and not in 2 or 3
• Cell two receives fire in year 2 and not in 1 or 3
• Cell three receives fire in year 3 and not in 1 or 1
• Cell four receives fire in year 1 and 3, not in 2

timestep c1 c2 c3 c4
t1 1 0 0 1
t2 0 1 0 0
t3 0 0 1 1

So the question is can each cell have a time series in this way? In many ways we’re dispensing with the probabilistic aspects of disturbance and just want to supply a spatial time series of when particular transition pathways occur in a pixel based on disturbance.

Any information on this front would be very helpful, we hope to hear from you soon.

Thanks,
Luke

[leonardo-frid]

Hi Luke,

Yes. You can be prescriptive about which cell receives a transition at which timestep. Here are the steps I would follow.

1. Identify which transitions are going to be specified using prescriptive instructions for each cell with a spatial time series.
2. For the prescriptive transitions identified in #1, set the base probability in the Transition Pathways to a value of 1.0
3. Under the scenario property Advanced | Transitions – Spatial | Transition Spatial Multipliers enter rows for the Transition Types/Groups that you would like to specify a spatial raster that identifies for which cells the transitions occur. Note that you can right click on the table to expose the Timestep field and make the raster time varying. Cells with a value of 1.0 will have the transition applied to them in that timestep. Cells with a value of 0.0 will not have the transition occur. A value between 0.0 and 1.0 can be used to set a probability for a specific cell and timestep.
4. Each raster will be applied starting in the timestep that it is specified until the next timestep that another raster is specified so you must remember that if you turn a cell on in one timestep, you must turn off the cell in subsequent timesteps when you no longer want a transition to occur.
5. Transition Spatial Multiplier rasters must have the same number of rows and columns as your Spatial Initial Conditions rasters.

[leonardo-frid]

For the example you provide your rasters would look like this:
Timestep 1, Fire, Fire-Spatial-Multipliers-ts1.tif
1.0 0.0
0.0 1.0

Timestep 2, Fire, Fire-Spatial-Multipliers-ts2.tif
0.0 1.0
0.0 0.0

Timestep 3, Fire, Fire-Spatial-Multipliers-ts3.tif
0.0 0.0
1.0 1.0

[LukeMcPhan]

Leonardo, That is great news!
Thank you so much for the speedy response. I really appreciate the detail in the reply, I’ll have a good run at this and let you know how I get on, but can see how this would work.
Just another small question, because you’ve been so helpful, have you run into computing constraints with respect to raster size?
We have some rather large areas that we would like to apply this across and just wanted to get an idea if you have come across a “reasonable” size limit that you would try processing, with respect to the amount of cells in the raster?

Cheers,
Luke

[leonardo-frid]

Hi Luke,
Good question. The answer to a large part depends on what computing resources you have available to you. We do many of our simulations using cloud computing resources such as AWS EC2 virtual servers. Using these servers, it is not unreasonable to do simulations with landscapes encompassing 1-10 million cells for multiple decade runs with 100 Monte Carlo realizations. The most extreme example we are currently working on is running landscapes >100 million cells. In this case we require machines with large amounts of RAM >64GB and also make use of spatial multiprocessing to run independent landscape units. In this case it can take a few hours (3-10) to run a single timestep/realization. SyncroSim can also run in a Linux HPC environment and there are some published examples of that on the ST-Sim publications page. See for example Sleeter et al. 2019.

[Author]

Posts