ArcGIS Pro - Classification Wizard with DSM/DTM

1089
6
03-20-2019 02:04 AM
DOE_-_DavidTipping
New Contributor II

Hi - I'm looking to do small scale vegetation classification using georeferenced aerial images created from a drone. This work forms part of habitat condition monitoring on nature conservation sites. I have had some success however I think I would achieve better results using additional height information derived from DEM and DSM (DSM minus DTM). I can find nothing on the various help sites only a mention that additional datasets can be used with the ArcGIS Pro Classification Wizard. It seems to fit in the "Train the Classification" area of #the process as shown in the following link: http://proceedings.esri.com/library/userconf/proc18/tech-workshops/tw_1666-182.pdf.

 

Does anyone have experience of these routines as I can't find where to include a DSM/DTM in the process. I have been following the route of supervised object based classification, creating a segmented image, training samples, Random Trees with Active chromaticity color, mean digital number and standard deviation. 

 

Any help appreciated

 

 

Many thanks

David

0 Kudos
6 Replies
JeffLiedtke
Occasional Contributor II

Hi David,

You are on the right track. (DSM – DTM) is a valuable dataset in classification for both veg and urban landscape classification.  In the OBIA application space, the result of (DSM - DTM) should be converted to 16 bit, then use the composite bands tool to create the 2nd input to the classification training tools.

If the 1st input to the train tool is your segmented image, then the 2nd input would include your multispectral image and the 16-bit DSM - DTM layer composited into a single image file.

Color, mean DN and Std Dev are good choices for the attributes. 

Best,

Jeff

DOE_-_DavidTipping
New Contributor II

Thanks Jeff - it was your YouTube tutorial that inspired this approach https://youtu.be/v40abIwGZWA 

I have not had a chance to run through your suggestions fully yet but have snipped the various layers down to 100m squares for quicker testing and have converted the DSM-DTM down to 16 bit. Is the 16 bit step necessary as it seems to limit the level of detail?

Will report back when I have run the processh - thanks

0 Kudos
JanineWagenfeld
New Contributor

Hello,
Currently I am working on my master thesis on object-based classifications with Random Forest. Using imagery and LAS data I want to classify urban objects.
Unfortunately I could not follow the answer of Jeff Liedtke. Why do I have to convert the DSM/DTM to 16 bit? And which bands do I have to merge great with the composite band? Do I have to composited the imagery with the DSM/DTM?

Which train tool do you mean?For example the Train Random Trees Classifier?
Or do you mean the training management in the classification wizard?

Would it make more sense to use the forest based classification and regression to use multiple parameters for training?


Thanks for your help


Janine

0 Kudos
JeffLiedtke
Occasional Contributor II

Hello Janine,

All the image/raster bands to be used in your classification need to be the same datatype in order for you to composite them into one image file. If your imagery is 8-bit or 16-bit, then the DSM/DTM needs to be 8-bit or 16-bit, respectively. You can also include other layers such as SAVI, intensity, Tasseled Cap layers, etc. - depending on your imagery. Once all the data layers you want to use for classification are in the same data format, then you can composite them into a single raster layer for input into the classifier.

Classification is 3 steps; 1) collect representative training samples, 2) train the classifier using one of the train tools, 3) classify. The Esri classifier Train Random Trees is the same classifier as "Random Forest" discussed in the literature. You have 3 inputs: 1) your segmented image, 2) image composite containing all the layers you want to use in the classifier, 3) your training sample data (which includes your classification schema). The inputs into the train tool and the classify tool need to be in the same order, i.e., if your first input layer in the train tool is the segmented image and 2nd layer is the composited image, then do the same for the classify tool.

Since you are using a segmented image, you can select the attributes to be considered in the classifier, I recommend using standard deviation, along with mean color and any of the shape attributes.

Good luck and have fun with you classification,

Jeff

JanineWagenfeld
New Contributor
Hello, Jeff,
Thank you very much. Now I understand the procedure, because I found the "additional input grid" in the tools Train and Classify.
My Problem now:
When converting the pixel depth of my composited image (32 Bit to 8 Bit signed), the values change. (My first attempt was to change the pixel depth by exporting the data. My second attempt was to use the "Raster copy" tool).
My range of cell values are all between -128 to 127 (8Bit signed).
Example -
My Composited Image with the following bands:
Band 1: NDVI values with -0,53086 minimum and 0,68932 maximum
Band 2: Values of LAS data with -85,0409 minimum and 81,5664 maximum
Band 3: Values of the Green NDVI with -0,70854 minimum and 0,61344 maximum
Band 4: Values of the NDWI with -0,61344 minimum and 0,7085 maximum
When converting from 32bit to 8bit the following happens
Band 1: NDVI values with -1 minimum and 127 maximum
Band 2: Values of LAS data with -85 minimum and 82 maximum
Band 3: Values of the Green NDVI with -1 minimum and 127 maximum
Band 4: Values of the NDWI with -1 mininum and 127 maximum

How can I keep my previous values without anything changing of my values so much?
Thanks for your help,
Janine
0 Kudos
JeffLiedtke
Occasional Contributor II

Use the Stretch function with min/max. Then export the results, and composite into one multi-band layer

0 Kudos