Good day,
I want to predict a 2026 LULC map using the 2019 and 2023 random forest 10 m resolution classification maps in TerrSet 2020 with the Markov chain files and the CA-Markov chain. However, I am getting stripes of seemingly no prediction in the output image. I have 7 LULC classes.
What is the problem and what is the solution?
The cause of the striping could be a couple of things. One of the driver variables might have some striping in it. More likely, the program can’t allocate correctly based on what it has to choose from. It may be predicting that there is a certain amount of a class that should exist in the output image, but there aren't enough suitable cells to allocate to that class(es). Since it starts from the top-left and works its way across each row and down to the bottom, if there were a bunch of cells with the same suitability, their allocation to a particular class can take on that striped look. And perhaps this has happened with a few classes.
To know for sure, I'd have to look at your data. If you want to start a tech support ticket and zip up all your data used to create this, and perhaps some screenshots of the dialog box options you chose, then I could try take a look.
Tammy
Thank you - I would like to start a tech support ticket. How and where do I upload my zipped data? Here is a google drive link to the rasters and markov files: https://drive.google.com/drive/folders/1_aDuAqVUhfbjCZxJYSEAmP-gX_kH-m4b?usp=sharing
My iterations were 3 (Amount of years predicting ahead to 2026 from 2023) and I used 5 x 5 window. I used the Markov chain matrixes ran in TerrSet.
Any update on a solution? Clark Labs
Sorry. I had researched it and given a detailed answer, or so I thought. I wonder if I didn't click submit. I need to refresh my memory now and look at the data again.
I am remembering this now. The data has 2 landcover maps, only 4 years apart, but there is a very large amount of change between those 2 years. It was hard to know if this was real change or if the landcover maps weren't very accurate. The legend for the landcover categories was hard to decode:
code 1 : PF
code 2 : NF
code 3 : W
code 4 : BU
code 5 : BL
code 6 : AGRI
code 7 : S
I remember wondering why there was so much change in category PF and what that stood for. I looked at that area on Google Earth Explorer to try to see what type of landuse was there. It didn't help me that much, but I'm not a geographer.
So that brings us to the striping pattern seen in the output image. Part of the process of allocating predicted land to categories using Markov is that it predicts how much land will be in certain categories based on the previous amounts and amounts of change in the landuse categories. When it goes to allocate that land for the output image, it is going to put a certain amount of cells into the predicted categories. It starts taking them from the most suitable cells. If there are a lot of cells of the same amount of suitability, then it starts allocating them - I now see from the bottom up, left to right).
I don't have your driver variable images or your LCM project file or your transition probability files. But based on the visual result, it looks like it thinks that there are a lot of areas of equal suitability for those categories, so it is allocating them in blocks until it fills up its quota.
I think that either the landcover files are not accurate enough or the driver variables do not explain the change well enough, so it is having to do it with more randomness than would be expected.
But, I'm doing some guessing here, since I can't run your analysis without all your files.
If you want any more help on this, please open a ticket.
To open a support ticket, on the Support Page in the top right, you see small words that say "Submit a request".
The direct link to where that takes you is https://forums.clarklabs.org/hc/en-us/requests/new.
Tammy