Welcome to this weeks discussion on unsupervised classification with remotely sensed imagery. This is a multifaceted lab looking at a number of different processes culminating in the unsupervised classification and manual reclassifying of the resulting raster dataset for a permeability analysis. The main objective was to understand and perform an unsupervised classification in both ArcMAP and ERDAS Imagine. Imagery for two different areas was provided, ultimately the UWF area as seen in the map below was the final subject matter for exploration of these topics.
Unsupervised classification is a classification method such that a software suite utilizes an algorithm to determine which pixels in the raster image are most like other pixels throughout the image and groups them accordingly. After the software has grouped the various pixels together it is up to the user to define what the grouped classes represent. For this type of classification the software is given certain user defined parameters such as number of iterations to run, confidence or threshold percentage to reach, and sample sizes. These essentially tell the software how long to run, what the minimum "correctly grouped" pixel percentage is, and how many pixels to look at adjusting at a time. Lets look at how this was applied this week.
A high definition true color image of the UWF campus was used for the above analysis. This entailed performing a clustering algorithm on the true color image to group like pixels together and then export them as a slightly less defined image for storage space and processing speed concerns. The clustering algorithm created 50 classes, or shades of pixels which approximated the true color image. That was essentially the unsupervised classification. The software was told to produce 50 classes with 95% accuracy overall. Then I manually reclassified each of those 50 original classes into one of 5 labeled classes. I accomplished this by highlighting the pixel shade and reviewing it against the true color image and assigning it to the classes described. Four of the five classes are straight forward and represent what they say, with some possible error. The mixed class however is there because certain pixel shades applied to different items that represented both permeable or impermeable surfaces. For example some dead grass showing could show a tan pixel while a tan rooftop could also be showing the same value. So recoding this pixel to be grass or buildings would be wrong for at least some of that cluster of pixels. To account for this the mixed class was created, which is why you can see some rooftops as blue, grassy areas as blue or green and some blue sprinkled throughout.
Overall this is a fairly course analysis, but it does do a great job of exercising the process and creating likely results. Next week we will look at supervised classification in case that question came to mind. Thank you.
Unsupervised classification is a classification method such that a software suite utilizes an algorithm to determine which pixels in the raster image are most like other pixels throughout the image and groups them accordingly. After the software has grouped the various pixels together it is up to the user to define what the grouped classes represent. For this type of classification the software is given certain user defined parameters such as number of iterations to run, confidence or threshold percentage to reach, and sample sizes. These essentially tell the software how long to run, what the minimum "correctly grouped" pixel percentage is, and how many pixels to look at adjusting at a time. Lets look at how this was applied this week.
A high definition true color image of the UWF campus was used for the above analysis. This entailed performing a clustering algorithm on the true color image to group like pixels together and then export them as a slightly less defined image for storage space and processing speed concerns. The clustering algorithm created 50 classes, or shades of pixels which approximated the true color image. That was essentially the unsupervised classification. The software was told to produce 50 classes with 95% accuracy overall. Then I manually reclassified each of those 50 original classes into one of 5 labeled classes. I accomplished this by highlighting the pixel shade and reviewing it against the true color image and assigning it to the classes described. Four of the five classes are straight forward and represent what they say, with some possible error. The mixed class however is there because certain pixel shades applied to different items that represented both permeable or impermeable surfaces. For example some dead grass showing could show a tan pixel while a tan rooftop could also be showing the same value. So recoding this pixel to be grass or buildings would be wrong for at least some of that cluster of pixels. To account for this the mixed class was created, which is why you can see some rooftops as blue, grassy areas as blue or green and some blue sprinkled throughout.
Overall this is a fairly course analysis, but it does do a great job of exercising the process and creating likely results. Next week we will look at supervised classification in case that question came to mind. Thank you.