I am currently using ArcGIS 10.3 and new to OBIA. My goal is to segment out portions of dead trees in the Lake Tahoe Basin and not sure how to go about this. I have searched extensively online and had no luck as to any tutorials or steps to follow. I am using Landsat 8 OLI (30m) and WorldView-2 (2m) spatial resolution imagery for this task. Here is my plan and please correct me if I am wrong. (For Landsat 8 OLI)
1. Create a Natural Color Image using Bands 4,3,2.
2. Establish training areas of dead trees and healthy lodgepole and red attack trees in the training sample manager.
3. Create a signature file
4. Under ArcToolbox>Spatial Analyst Tools>Segmentation and Classification>Classify Raster
Are these steps suitable to segment the landsat image?
Any help would be greatly appreciated
Norm
Solved! Go to Solution.
This may be of help in getting started. I found this with a web search. It includes a workflow.
Understanding Segmentation and Classification—Help | ArcGIS for Desktop
I'm not a Remote Sensing expert, I but have noticed several GeoNet folks are very knowledgeable. One person who comes to mind that may know an answer to your question is Curtis Price
Chris Donohue, GISP
This may be of help in getting started. I found this with a web search. It includes a workflow.
Understanding Segmentation and Classification—Help | ArcGIS for Desktop
I'm not a Remote Sensing expert, I but have noticed several GeoNet folks are very knowledgeable. One person who comes to mind that may know an answer to your question is Curtis Price
Chris Donohue, GISP
An overview of the Segmentation and Classification toolset—Help | ArcGIS for Desktop is the start of the tree in the help .
Classify Raster—Help | ArcGIS for Desktop is the tool he is referring to Chris...
Understanding Segmentation and Classification—Help | ArcGIS for Desktop is the topic you refer to. He just hadn't looked far enough down the help topic tree.
There is stuff on the arcgis blog on landsat 8 Band Combinations for Landsat 8 | ArcGIS Blog and you should do a web search to see if your band combinations etc and processing techniques meets the goals. There is no "best" approach (one example here remote sensing - Which methods are most suited to classifying urban areas from Landsat imagery? - Ge... ) and the software you use will determine what options you have available to you. So wade into this area carefully.
EDIT
tons of way-cools photos all over the web Landsat 8 Bands « Landsat Science
Hi Dan,
Cool new capabilities. Makes me wish I had 10.3.1 instead of 10.2.1.
Except for the unintended shapefile issue..... (you probably already saw this, but if not)
ArcGIS 10.3.1 does not recognize Long Integer right
Chris Donohue, GISP
Thanks Chris... Yeah, I saw that... I haven't seen any ramifications yet being posted on the Remote Sensing and image analysis place. I will have to troll around to see if anything has surfaced. Most of my comrades, use PCI, Erdas etc etc for RS stuff and use Arc* for the integration of results with vector or raster, so I know they haven't come across difficulties which may surface.
Thank you for your reply. I did read the topic tree but I forgot to mention that I am having a hard time creating the Esri Classifier Definition (.ecd). When I use the training sample manager and try to calculate the statistics, or view the scatterplots, etc..., I keep getting an "Out of Memory" warning and cannot move past this step. Is this the only way to create the .ecd? How many training samples do you think that I need? Do I need to do this classification using the Color IR image (5,4,3), or can I do this using the RGB (4,3,2)?
Also, I saw that others mentioned various programs to use for this. Unfortunately, I am working on my thesis and am at the final steps and do not have much time to learn these other programs, nor do I have access to them. I have IDRISI selva but not that proficient at it. So any help towards using ArcMaps latest edition would be helpful.
Again, I appreciate it.
Norman Nash
It's my opinion that ArcGIS Desktop at 10.1/10.3 is quite sufficient for traditional basic supervised and unsupervised per pixel classification. The out of memory errors suggest you are selecting an unreasonably large number of cells.
From what I've hard, you may want to take a gander at a basic textbook to help you understand fundamentals (not specific to ArcGIS) like band selection and get back to the thread when you have a better idea what you are trying to accomplish.
Interestingly I haven't heard anything on this thread about "object based" analysis. (In quotes because it means different things to different people.) I'm assuming you are thinking of running some kind of per pixel classifier and then doing some other analysis or workflow?
Your steps are good, because i tried the same task with a couple or aerial images to test the arcgis capabilities. And the result is not bad.
But as DAn wrote the people that work with the remote sensing are using ENVI, ERDAS, ERMAPPER, PCI geomatica or other similar program to accomplish the job, then use arcmap or Illustrator together with avenza MAPublisher to make a good presentation with vector data
Hi. (1) What I recommend is to segment the WV2 imagery at an appropriate scale (min segment size of about 20, use the raster function to try different minimum mapping sizes), then once you are happy with your segmentation settings (as a raster function) run the segment mean shift GP tool with those settings. (2) Remember, before you segment the WV2 image, to enhance it so that your features of interest are visually discernible - make those dead trees pop out of the imagery (and any other features of interest)!. (3) Then collect training sites by choosing the segmented raster file/layer in the classification toolbar, but display the L8 imagery in the ToC. So you will use the spectral info of the Landsat to choose your training, coupled with the segmented objects to get accurate training samples (i.e., the training sample polygons will be from the segmented image. Then classify the imagery using the Train SVM GP tool:P (a) segmented image, (b) Landsat imagery (all bands (except QA band etc)) maybe adding a DEM (NED 10m) composited into one image (the second image input), and (c) training file. Then (d), Classify Raster GP tool with the same inputs (in the same order) using the .ecd file from the train gp tool. Happy Classifying!
Jeff
BTW, the .ecd file does not work with the histogram and scatterplot graphs on the classification toolbar. It is generated by a new generation of classifiers that use the segmentation info as well as the spectral info. All the capability will be integrated and really sing in a later release of Pro (soon!).
Also, I forgot to mention that for the Segment mean shift tool you should set your "spectral detail" parameter high, 18-20 and the spatial detail can be left at the default or maybe lower, try out some different settings in the Raster function.