Smoothing the DEM

[HusamAbuOqer]

Hello everyone,

I am working on the Archydro toolset and I am facing an issue with the perona malik smoothing method, the mean and median areworking normally but the perona malik is not working and showing the error below.

(I am using the DEM I generated from the LAS file data)

 

[GinaO_Neil]

Hello @HusamAbuOqer ! Perona Malik requires the input DEM to be in the .tif format. While other formats may work with the other smoothing methods, it is highly recommended that .tif files be used as inputs anyway. Please let me know if this resolves your issue. Thanks!

-Gina

[HusamAbuOqer]

Hi Gina,

Thank you for your reply.

I think that was the issue I have been running through last, now I solved it by converting the DEM to a tif format.

Currently I am running the "Run random tree step and I am getting a raster of one value only (0), any recommendations for this issue?

Thanks.

[GinaO_Neil]

Great! 

You may be seeing a prediction value of just the 0 class if your training data is highly imbalanced towards the 0 class. If the imbalance is the issue, you can try significantly undersampling the majority class (0) when running Train Test Split: set the sampling percent for class 0 to be much lower thanthe sampling percent for class 1

[HusamAbuOqer]

Hi again Gina!

I have been trying using testing data and it worked perfectly when I used the main data I reached the train test split and I am facing this error in the screenshot below.

Any recommendations for this issue?

Thanks

[GinaO_Neil]

Hi @HusamAbuOqer ,

I believe your study area is too large. I recommend working in units of watersheds and from DEM data that is 3m resolution. I've found that 3m DEMs clipped to HUC12 watersheds works well with the tool's size limits. How large is your area? And are you confident in both the wetland coverage and nonwetland coverage in this entire area?

[HusamAbuOqer]

Hello @GinaO_Neil 

My study area is 23 sqKM with a DEM resolution of 25 centimeters. 

And I am using ground truth data for Alberta from a government data source showing the wetland and non-wetland areas.

The test data area I tried it before and it worked normally was with almost the same size (but the resolution for this DEM was 1 meter)

Is this area considered a big area for the toolset or it's because of the high resolution of the DEM I am using? 

[GinaO_Neil]

Yes, that's a very large dataset. Over several tests I found that wetlands were better predicted by topographic indices calcualted from 3m DEM than from a 1m (or higher res) DEM. I recommend resampling your data up to 3m and working in one HUC12 at a time - i.e., calculated predictor variables for a single HUC12 extent. That sounds like amazing ground truth data! You should have the opportunity in that case to train wetlands in one HIC12 and use the trained model to predict wetlands in a separate HUC 12

[HusamAbuOqer]

@GinaO_Neil  Thank you so much for your reply!

I will check on this now.

[GinaO_Neil]

No problem! Let me know if you continue to run into issues. I am happy to set up a screenshare as well. Good luck!