The purpose of this lab is to do a change detection analysis using landsat imagery in ERDAS Imagine. My study area is the Tri-City area of Southern Ontario Including Kitchener/Waterloo, Cambridge and Guelph. I am attempting to see the change of the urban footprint from 1990 to 2013.
The steps include:
Atmospheric correction, Radiometric calibration, Change Detection, Change Classification
Atmospheric Correction (PCI)
I do not have access to ERDAS Atmospheric correction tools and although I could use a TOA and haze removal in ERDAS, I am instead using the ATCOR module in PCI. This introduces an extra step of translating data between the two software.
The correction did not make a large difference as the scenes had little atmospheric interference. The 1990 landsat 5 imagery was not corrected as the image was clear to begin with. The 2013 imagery was corrected with haze reduction, and water and cloud masks. To then move the ATCOR 2013 image into ERDAS you must open the file tab in the content viewer (vs map tab) right click and select translate (Export). Select the .img file type and select the bands to translate. Make sure for earlier bands you select the correct landsat (TM or MSS, most data is TM)
It should also be noted that if you are using a DEM for your translation to remove shadow and other topography errors, your dem must be larger than your landsat imagery. One source for such imagery is etopo earth dem which is accurate to 1 arc second. A second option is SRTM data from EarthExplorer. SRTM data will likely need to be mosaicked to cover the entire landsat scene.
In the case of the 2013 imagery I am not convinced the the atmospheric correction is better than the original image as it darkened a lot of areas on the ground which complicated my classification in the change detection step.
Radiometric Correction (ERDAS)
Assuming that ATCOR worked on both of your images you may be able to just load the .img files (i.e ATCOR file created and translated). Otherwise you may need to import (manage data –> import data) or stack the files (spectral –> layer stack) in ERDAS. Tutorials for these steps are available under the tutorial section of the website. When you open your imagery it may appear dark. If this is the case you can try stretching the imagery using spectral options or you can try subsetting the imagery to remove the areas if possible that are causing the large contrast and darkening of the image (i.e remaining clouds or hazy areas).
For this specific scene radiometric correction was unsuccessful due to extreme variances between the images. Southern Ontario is a heavily farmed area meaning that the land cover is constantly changing from bare to vegetated, with changing vegetation. This is not idea for change detection. In retrospect a scene between harvest and snow fall would have been more appropriate for this study and may have resulted in better results.
Additional issues can include rainfall events which can preferentially affect different plots of land depending on soil type.
Change Detection (ERDAS)
If the imagery was matched well enough you can use raster subtraction to detect differences (raster functions in ERDAS). For this example the image histograms were too different to run direct change detection (subtraction). Alternately, discriminant function change detection tool can be used.
If spectral differences can’t be used you can classify the imagery and then use the Discriminant Function Change Detection tool. The images can be classified using either supervised or unsupervised classification. Supervised is appropriate for high accuracy results for a minimal number of images as this method is time consuming.
For this project I did a supervised change detection method. Including the supervised classification of both images and the application of the clump and eliminate tools.
TriCity Area, Ontario
Urban Landcover 1990 (blue)
Urban Landcover change (1990-2013 in purple)
The change detection was fairly successful using this method. The growth of many of the major cities in southern Ontario is apparent in the resulting imagery shown above.
Issues encountered were non permanent land cover change (farming) that influenced spectral change detection. Success was much better when using a supervised classification although there were still some issues detecting urban from bareland. To minimize this cropping to specific study areas could be helpful. There is some change detection error likely introduced between the two dates due to the method, but AOI’s chosen typically overlapped between both images.
The 1990 imagery achieved better classification results with a simpler process. I believe this is because the imagery was brighter possibly due to a recent rainfall in the 2013 imager (would need to be compared to weather data).
Overall results could have been improved through more rigorous selection techniques of the imagery.
I attempted the delta cue change detection method with ERDAS but was unsuccessful again due to the land cover change associated with farming. This method may be more successful between harvest and snow fall.