Hi - I’m working with several InVEST models for a very large study area (~300 miles of coastline and inland ~20 miles). Because of the resolution of my input landcover dataset (30x30m), the model takes many days to run for the entire study area.
My solution to this is to segment the overall study area into 5 different AOIs. All of the input datasets are the same across all segments (habitat boundaries, stressors, and criteria scores/weights/DQ). Does the HRA have “edge effects” that would prevent me from confidently stitching together these segments into an overall HRA output?
Phrased another way - will an HRA risk score of 1.5 in one segment be the same as a risk score of 1.5 in a different segment if all of the input habitat/stressor datasets and criteria scores are the same?
If not - could anyone suggest an alternative that gets me an HRA output across this entire study area?
I think your approach of dividing the AOI might be okay for HRA, but there is at least one reason why it might not. The stressor buffer parameter in the Habitat Stressor Info table (along with the decay equation parameter) is controlling the distance around a stressor at which a habitat will be impacted. If you have buffers greater than 0, then you might need to be extra careful about where you split your AOI so that a given habitat pixel will see the same presence of a stressor as it would if all the inputs were completely continuous and buffers were allowed to travel across the split AOI.
Assuming you are working with version 3.7.0, there may be a better way. You can enable parallel processing with multiple CPU’s from the File > Settings > Taskgraph n_workers parameter in the InVEST UI. HRA should be able to take advantage of parallel processing in many places. If your original plan was to divide your AOI and then launch several parallel instances of InVEST on your machine, I would definitely try setting the n workers parameter instead.
Thanks for the quick response, Dave. I am hopefully working around the buffer distance issue at the edges of each segment by making sure they overlap by at least the distance of the maximum stressor buffer (in this case 1km). It sounds like that should overcome this issue to some extent.
As far as the parallel processing - I haven’t tried that yet. I’ll look into that as an alternative to let me run each model (HRA and NDR) across the entire study area.
Appreciate the response - I can post a reply to this message if/when I figure out the right approach.