Extracting raw data from an unsupervised classification of multiple Landsat scenes

[DanielWhite5]

Hi All, I'm trying to extract the raw data values for the 5 spectral bands of a Landsat scene that I have conducted an unsupervised iso-cluster unsupervised classification on.

I am looking into the variability of sea-ice using Landsat images.  One of the fundamental steps I need to do in order to achieve this is to classify the snow/ice/water based on its reflectance.  So far, I have run an unsupervised iso-cluster unsupervised classification based on 5 rasters, being the spectral bands 1,2,3,4 and 6, and have opted for 10 classes. What I need to achieve is a classification with enough classes that are statistically distinct from one another, which I can then use to classify the medium on all my other images.

As an example, I have a composite raster with 10 classes, each a product of the 5 bands as mentioned.  Class 1 has a count of 464285, but I would like to know for that class, what is the raw data value for each band within the composite image?'

Any advice greatly welcomed thanks!

[GabrielUpchurch1]

Hi Daniel,

Could you provide some clarification on what you mean by "raw data value?"  The reason I ask is because each class is going to consist of a range of values from each band.  For example, class 1 will contain a range of values from band 1.  Also, what format (raster, vector, table) do you need the extracted values to be in?

[DanielWhite5]

Hi Gabriel,

Thanks for your response.  The process is entirely new to me so I'm afraid I'm a little unsure how to further refine the concept of the raw data value for each band, however, what I want to do is to draw the stats on each of the classes from each band of an unsupervised classification, composed of 5 bands - the key point being I want to interrogate the classes within the raster, not the entire image.

I've been advised elsewhere that Class Probability in Spatial Analyst is the way to go with this, and will post back here when I've determined how well that solution fits the problem.

Best wishes,
Daniel

[GabrielUpchurch1 · Accepted solution]

Hi Daniel,

Class Probability should be a useful tool for what you are trying to accomplish.  You might also take a look at the Zonal Statistics or Zonal Statistics as Table tools in the Spatial Analyst Zonal toolset.  You can get at a variety of descriptive stats for each class by using your classes as the input zones.  I think the two tools only work on single band rasters so you would need to run each Landsat band individually.

[DarrenWiens2]

Have you inspected the signature file? It should tell you the mean for each input band in each class, along with a covariance matrix. I'm not up on stats enough to know if you can convert that into a range of values for each band, but it's a start.

[DanielWhite5]

Hi Darren,

Yes I have the signature file, but as you've pointed out, it specifies the mean for each band within the composite image. What I'm after is the mean by class, whether that be 2,5,10,20... classes. I think I've found a way of doing this through the Class Probability function in Spatial Analyst.

Thanks for your suggestion

[DanielWhite5]

Many thanks Gabriel.

I think the Zonal Statistics function that you've suggested is going to be the most robust way of completing this process