Lab 5

Spectral Mixing Analysis

By: Jackson Radenz

Introduction:

Using 'hard' image classification techniques and object based image analysis we classify pixels and objects as one, single feature.  However, in reality we know that classifying areas in our imagery as one, single feature is false.  If we were to go to ground points throughout our image we would see that the areas not only consist of the single feature classified, but mixes of two or more features.  For example, we can classify a 30 x 30 meter pixel collected by Landsat as 'deciduous forest', but if we were to go to that 30 x 30 meter plot on the ground we would most likely find, deciduous trees, bare soil, dead leaves, grass, and possibly a small amount of water.  We call this a mixed pixel.  Therefore, this is when 'fuzzy' or mixed pixel classification becomes important.  In order to accurately classify mixed pixels, we must find 'pure pixels'.  Pure pixels are pixels that are very heavily dominated by one feature.  By finding a single feature's pixel we are able to derive its spectral characteristics that we are then able to apply to a mixed pixel's spectral characteristic to find what percentage the mixed pixel holds of a certain feature.  Pure pixel's spectral profiles are called endmembers.  It is assumed that each pixel contains a proportion of an endmember's spectral profile.  The spectral proportion of a mixed pixel's profile compared to each endmember's spectral profile is able to determine how much of a certain feature belongs to that pixel.    

Endmembers: 

Endmembers are to represent the purest pixels in the image.  Endmembers are difficult to find for several reasons.  First, endmember's features are sometimes not as large as the spatial resolution of the image.  Second, there is a large amount of variability in nature.  For example, perhaps there was a rainfall right before the image was taken that would've changed the feature's spectral profile.  Third, endmembers are not constant throughout the image.  The same features give off very similar, but slightly different spectral profiles.  Finally, shadows can be cast on endmembers, which slightly change their spectral profile.  All of these are reasons which make it difficult to find pure endmembers, however it is important to do so.   

Data:

  • Imagery Collected by: USGS
  • Imagery Sensor: Landsat 8
  • Spatial Resolution:
  • 30 m - Visual
  • 100 m - Thermal
  • Number of Bands: 11
    • Coastal Blue
    • Blue
    • Green
    • Red
    • NIR
    • SWIR 1
    • SWIR 2
    • Panchromatic
    • Cirrus
    • TIRS 1
    • TIRS 2
  • Date Acquired: 2015 - 10 - 17
  • Area: La Crosse County, WI

Methodology:

  • Open ENVI Classic and Load 'new_subset_msi_TOA_REFLECTANCE'
  • Find the deepest channel in the Mississippi River and Enhance Image (preferred is Gaussian)
  • Find the darkest pixel within the Mississippi River and digitize it by creating it into a ROI (figure 1)
Figure 1. Selecting ROI
  • 'Save ROI' and go to 'Spectral Library Builder'
  • Within 'Spectral Library Builder' import the ROI, plot it, and save the Spectral Profile of Water as an Endmember into your file (figure 2)  
Figure 2. Creating an End member
  • Repeat steps 2 - 5 for bare soil, built up, and green vegetation.
  • Once the 4 Endmembers are found, go to 'Spectral', 'Mapping Methods', and 'Linear Spectral Unmixing'
  • The 'Unmixing Input File' will be the 'new_subset_msi_TOA_REFLECTANCE' (figure 3)
Figure 3. Input File
  • Load all 4 Endmembers in and create an 'Endmember Collection Spectra' (figure 4)
Figure 4. Endmember Collection Spectra
  • Select 'Apply'
  • In 'Unmixing Parameters' select 1.00 for a 'Unit Sum Constraint'
  • Load Fractional Cover Bands in ENVI
  • Analyze Final Image

 Discussion:

In the fractional cover image we are able to create using the our selected Endmember's spectral profiles we are able to see where different types of land cover dominate.  For example, the fractional cover band of water was loaded into the red color band of ENVI.  Therefore, that is why the Mississippi appears very red.  In the cursor location values, we are able to see that the red band is very high because it is linked to the fractional cover band of water (figure 5).

Figure 5.  Linear Spectral Unmixing Map

Some of the fractional cover values are negative (as seen in figure 5 in the green and blue bands).  I believe they are negative because the error associated with mapping cover values creates negative values.  For example, I found a fractional cover value of 1.00 for water when my cursor was over the Mississippi channel.  Before creating the map we put a constraint on the values, meaning the highest a value could be is 1.  Therefore, if water is already 1.0 the values have to be zero or negative.  However, if another feature has a small, positive fractional cover value the other features must be negative so the sum constraint of 1 is not broken.     

Conclusion:

Overall, I believe Linear Spectral Unmixing is a more accurate way of classifying imagery.  Instead of giving certain areas a hard classification, it is able to give fractions of the entire cover type and paint an entire picture.  This is a much more accurate depiction of reality.  However, it is important to find pure spectra throughout the image.  With the low spatial resolution that this image contains, it is very hard to find pure spectra, if truly any at all.  This can lead to a high amount of error.  The RMS Error Band allows us to visualize this (figure 6).
Figure 6. RMS Error Gray Scale Visualization


The fractional cover type that had the lowest accuracy was bare soil and built up area. All the areas that showed up as white means they had a high amount of error relative to the rest of the image. I found built up areas and bare soil areas to be the most high and present the most contrast.
Linear Spectral Unmixing is a great technique for classifying an image, however high spatial resolution (or overlaid high spatial resolution), a basic knowledge of the land in the acquired imagery, and an ability to find pure spectra is needed in order to accurately apply the Linear Spectral Unmixing technique.



Subscribe to: Posts (Atom)