4

I am working on digitizing the Agriculture land cover in a certain region. I did something very similar before and couldn't find a good alternative method beside manually digitizing, which is utterly time-consuming.

I have tried several methods. but none seems to work perfectly.

  1. Image Classification This is part of the original map

OGAG

What I got from the Image Classfication Result:

Classification Result

And here is the legend:

Legend

The big problem here is that the tool seems to confuse some of the personal yards with Ag Land, which is totally understandable as some of the Ag land is like this:

Green Ag

But I think this can still be a foundation to base my further work on.

  1. I tried to extract certain classes from a landcover raster and then erase tree canopy off it. The results are not quite optimal as you can see here:

Landcover Result

Too many "noises" and the boundary are too coarse and edgy.

Is there any solution?

  • Nothing can replace manual digitising. Anyway, I'd try 3 distinct classes first. White roofs, brown fields, the rest. By shrinking / expanding rasters try to eliminate noise. Erase 2 successful classes from original, try with remaining. Good luck. Also having roads polygons might help – FelixIP Feb 18 '17 at 2:59
  • Thanks for your comment. Your raised a very good point about doing it in steps. For some reason this strategy didn't quite occur to me. I always wanted to finish it in one step, which may not seem that prudent anymore. Thanks again, I will definitely give it a try – AndrewLebron Feb 21 '17 at 14:50
3

There are many things you can do to optimize your land cover classification. Most important, you need to acquire the highest quality imagery available. In your case, WorldView-2/3/4, Landsat 8 OLI or Sentinel-2 sensors would be highly suited to classifying agricultural fields--often to crop species. Higher spatial resolution data such as 4-band NAIP can be effective for agricultural applications too, especially if crop boundaries are important.

I can see you used a pixel based classifier, most likely a maximum liklihood algorithm in ArcGIS. You will certainly have to deal with mixed pixels using a pixel based approach. One way to overcome these mixed pixels is to use a single or double pass majority filter, available in ArcGIS.

You will likely have better results if you use an object oriented image analysis (OBIA) approach. ArcGIS has an image segmentation algorithm available that I would recommend. You can learn more about this approach here: Understanding Segmentation and Classification.

Finally, I would recommend incorporating data in addition to your spectral bands such as vegetation indices (NDVI) and various texture metrics. These can be easily incorporated into the classification algorithm as you would the spectral bands.

In sum:

  1. Find better imagery with higher spectral resolution
  2. Use OBIA such as image segmentation/classification
  3. Incorporate additional data as information bands such as vegetation indices (e.g. NDVI) and texture metrics into your classification
| improve this answer | |
  • Hi, thanks for your reply. So so many good points that I feel like I have so much more to learn. Several questions for your input: 1. For the data source, the aerial I'm using is actually 1*1 ft. Do you think this is good enough? If not, where can I download these high-res imagery. I didn't find any downloading option for Earth Explorer. 2. I am doing the whole thing in ArcGIS, which tool shall I use for OBIA? 3. What do you mean by incorporating NDVI in the bands? Would you mind terribly elaborating on it a little bit? Thanks – AndrewLebron Feb 21 '17 at 18:43
  • 1
    @AndrewLebron In reference to your questions: 1) when classifying agricultural areas, spectral resolution is more important then spatial resolution. The reflectance properties of the imagery are what will allow you to distinguish classes--not the spatial resolution, 2) this document should get you started: desktop.arcgis.com/en/arcmap/10.3/tools/spatial-analyst-toolbox/…, 3) NDVI is a vegetation index that you can create using the raster calculator in ArcGIS. Use the created NDVI raster as you would the red, green, blue, nIR, ... bands. – Aaron Feb 21 '17 at 19:14
  • somehow my previous comment replying to your comment is lost. No wonder I couldn't find any new response from you. Thanks for your timely reply first. And some following question: 1) TBH never heard of spectral resolution before. Is any of the source data you mentioned originally free to download? If so a link will be utterly appreciated as I can't find one. 2) Thanks, I will read up on it. 3) Can I create my own NDVI in ArcGIS? If so, how? I thought it shall be some data source I import elsewhere and could help me with my classification. – AndrewLebron Feb 21 '17 at 21:14
  • @AndrewLebron It is best to post these as new questions. 1) edc.uri.edu/nrs/classes/NRS409/RS/Lectures/…, 3) You can use the raster calculator to calculate NDVI: (nIR - Red) / (nIR + Red). – Aaron Feb 23 '17 at 13:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.