0

I have tried several methods including unsupervised classification, supervised classification, random-forest classification, and spectral indices methods to estimate the amount of impervious surface in Midwest US town from Landsat 8 Image.

The results did not converge reasonably. The core problem seems to be the case of mixed-pixels in the training shapefile which overestimate the impervious surface. I also suspect spectral confusion in which certain land-use features such as soil are wrongly mapped as impervious because of spectral similarity of these features.

I am using ArcGIS 10.3 and ENVI for processing the imagery.

How can I use a training data (polygon or point shapefile) that only contains pure-pixels, so that I don't overestimate Imperviousness?

I would prefer to work on ArcGIS 10.3.

|improve this question
  • Besides the spectral bands from Landsat 8, do you have any other datasets? For example, a soils layer or perhaps a transformed land layer? A way to try and improve the classification would be to replace one of the bands you are using in your classification with a known dataset. Spectral mixing and confusion will always be present in a coarse dataset like Landsat 8. – Keagan Allan Nov 27 '17 at 11:37
  • @KeaganAllan Unfortunately i don't have any other dataset. I was thinking may be there is a way to identify pure pixels of Impervious Surface. – Kaleab Woldemariam Nov 27 '17 at 13:26
1

Some things to try improve your estimation:

  1. Create an indices layer and add it to your Landsat 8 bands. Try the NDVI and create an Image stack. This will add the NDVI to your list of bands for the L8 image.
  2. The NDVI could slightly improve the definition of your classes by helping separate out vegetation and man-made surfaces. I am guessing you are looking at a hydrological application of this method.
  3. Have you tried using a different platform? ASTER and Sentinel-2 have a higher spatial resolution. The smaller the pixel the less chance there is for spectral mixing to occur.
  4. Try increasing the number of classes you define in the unsupervised classification, then manually reclassify the image comparing the classes with recent aerial imagery.
  5. You can try these in ENVI and change the type of algorithm used for the classification. Parallel-piped / Maximum Likelihood etc. Last time I used ENVI there was a Artificial Neural Network classification technique. It proved to be quite successful.

It is going to be tough to get a 100% result from this technique, but hopefully you can get closer to the result you are looking for. I have been playing around with the idea of using RS to define landtypes and surface characteristics in floodline mapping.

|improve this answer
  • I'm using Summer Image, probably with significant fallow land. I'm thinking i should use SAVI ( Soil Adjusted Vegetation Index ) for the 8th Indices band you suggested, What do you suggest? – Kaleab Woldemariam Nov 27 '17 at 19:07
  • I understand that land-use classification can't be 100% accurate.I am sure the NDVI image will improve the accuracy. But how can i use a good training dataset? I think ASTER has 15m resolution and would improve the spatial resolution, but how can i make the training data ( most probably point dataset, because polygon will be inadequate to select pure impervious pixels) without spectral mixing to represent impervious surface? – Kaleab Woldemariam Nov 27 '17 at 20:22
  • That is a tough question. Honestly I think using a platform such as Google Earth or any other up to date aerial photography (even GPS'ing the areas in the field" will be the best way to find the "Pure" signatures. There may be other methods, I am unfortunately not familiar with them. – Keagan Allan Nov 28 '17 at 6:44

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.