This Weeks Lab
- Review of Classification
- Topographic Effect
The Topographic Effect
Feature extraction and classification are complicated in mountainous areas by topography, as aspect influences the DNs for similar habitats in varying slopes relative to the solar illumination at time of image capture (~10.30am local time = from the SE). In Geog 432 you looked at a few approaches to try to deal with issues of topography. You performed ratios and indexes for instance. We are going to look at another method for trying to reduce the effects of topography in this lab – but first lets review the process for classification.
You may recall Roger and Scott speaking about what we are doing when we perform operations such as classification in Remote Sensing. We are really working through multivariate statistics in hopes of recognizing patterns in the data that can help us make decisions about the type of features existing in the landscape. We are looking to see what qualities are present within a series of Digiatl Numbers, but we also have some knowledge of he landscape itself.
For this lab, we will be using a small subset from the Bowron data set we used last week. You should make an appropriate directory in your home workspace and copy /home/labs/geog457/bowron_data/bowron.pix into it. We will do a couple unsupervised classifications before we do any other work.
Once you have the file in your directory, perform an unsupervised classification by selecting bands 3,4,5 and using K-means with defaults values (16 classes). You should notice that the shadow areas are misclassed with water.
Use raster calculator to generate a 4/3 ratio … save in a 32-bit channel. Try the same classification again, but this time use channels 3,4,5 and the 4/3 ratio. How does it look now. Is there still some confusion around North West slopes.
Can you think of why some terrain is affected more that others – and why North west?
There are several approaches to trying to offset this including the use of multiple diverse training areas for each class in a supervised approach, incorporating DEM layers especially slope and incidence as input to classification, and using topographically reduced channels such as ratios and greenness. Here we will try the approach used by Kuzera, Rogan and Eastman (2005) in extracting a measure of topography from the image channels, by modelling the hill shading component from the raw imagery.
But we won’t be doing (yet) the change component.
Check through the file in focus to determine more about the bowron.pix file (sensor type, date, resolution, projection…). Do you find all the answers?
Open the bowron.meta – metadata text file in the bowron_data directory – in a text editor and look to see what date the imagery was acquired (among other things).
The date is ideal for a mid-latitude mountain area with most snow gone, and still reasonable sun angle, found in the metadata file. The azimuth should be fairly constant for BC scenes as it won’t change with time of year, but the angle will decrease further away from the summer solstice, and also the further north. The area seems to be a good test ground with ridges running mostly perpendicular to the sun direction. To estimate the effect, check some typical DNs for similar vegetation areas, e,g, conifers, alpine meadow, avalanche chutes on opposite slopes facing NW and SE in bands 3,4, and 5.
1. Save the file to you own directory, start a new project, open your file to finish the DN queries.
2. task REL, create hillshading using the sun azimuth and angle (height) values given in the metadata. There is a DEM layer (that includes a shaded relief layer to be added to your file in the bowron_dem_28.5.pix file in the bowron_data directory.
Look in the DEM sub-directory to see if you can guess how the DEM for the layer was assembled.
Perform a Google search to see if you can find the suns location and angle of incidence through a web site here are a couple examples:
3. Using layer ->scatterplot, regress the hillshading (X) against each band 1-5, 7 (Y) to generate an equation with slope and intercept unique to each band. Record this equation for each band you are interested in adding the effect to..
Note that this should be restricted to land covers well represented in all topographic aspects – this should exclude water. How would you achieve this? Ask Scott.. and suffer him singing again!
4. Create up to 12 new 8 bit channels, and from the equation under the scatterplot y = ax +b apply this to the hillshaded image DNs for each of the bands to yield 6 new ‘estimated illumination effects’ – these should be 8 bit for simplicity.
use EASI modeling e.g. if the new empty channel is 9 and the hillshading is chanel 8: %9 = .05*%8 + 53.6
.. and so on for the other new channels (10-14) and matching values of a and b
5. Use layer -> histogram to find the mean for each of the new 6 bands, record this rounding the decimal to the nearest integer.
6. Then subtract the result for each band in 4. above from each of the 6 ETM bands while at the same time adding the mean value of the band from step 5 above: input these into the same modeling equation. The result will be mostly to cancel out the large subtraction but include local differences. Again ignore or round up decimal values for the mean.
e.g. %15 = %1-%9 +51
This should now create new channels with minimal topography.
7. Run tassel first on the original bands (input = all six reflective bands) and then on these new six channels and compare for content. The first component (Brightness) should show much less topographic component.
8. With time permitting, you could now run a classification on this image.
Summary of the Kuzera et al process:
- generate hillshading with the sun azimuth and angle as given for the satellite data
- regress hillshading against the 6 Landsat bands one at a time
- Apply this regression equation to create 6 illumination effect channels
- Subtract this channel from its original band for each 6 bands, and add back in the mean for the original band
- Run Tassel cap to transform the data into components for change analysis