Remote Sensing Labs: October 2012

An important part of remote sensing is change detection. Change detection allows people to see changes that have occurred on land over varying periods of time. Knowledge of these changes is important for many purposes such as urban planning and development, analyzing forest change, or noting changes in land cover. In this lab, I have performed a Change Detection with ENVI for the Washington D.C. area using an earlier image from Landsat TM taken September 16, 1991, and a later image from Landsat ETM+ taken July 28, 1999. These images have been georeferenced and corrected using DEM data to correct for possible distortions. A color composite of both images, composed by combining the Red, Green, and Blue bands (bands 3-2-1 respectively), is shown under Figure 1. This color composite image allows us to see the natural state of the area as we would see it in real life. Some key features to notice are the Potomac River, the Reagan International Airport, the pentagon, and overall regions of brightness.

The first step in the lab was loading the color composite images to later perform analysis and classifications on them. The color composite images were created as mentioned above by loading bands 3-2-1 in this specific order. The combination of these bands in this specific order allows the color image to appear in a spectrum we can see. After creating the color composites, we performed a Write Function Memory Insertion. The Write Function Memory Insertion takes information from the near infrared band (Band 4: wavelength 0.76- 0.90) of each image (1991 & 1999) to detect change in brightness values between both years. The near infra-red band is useful for detecting the amount of vegetation and helps emphasize the soil/land, and water/land contrasts.

First I created a blank band by using the Generate Test Band feature in ENVI which I later used as the blue band. Once I created the test band I started loading bands to create the change detection. I loaded Band 4 of the 1991 image as the Red band, then Band 4 of the 1999 image as the green band, and finally the blank band as the blue band. The resulting image was in yellow, red, and green. You can refer to Figure 2 for a visual. The yellow in this image means there was no change in brightness values between the 1991 and 1999. The red areas mean there was a higher brightness value in 1991, and the green areas mean there was a greater brightness value in 1999.

These changes in brightness values can be due to changes in landscape over the years in terms of vegetation or commercial. One area of interest is the Reagan International Airport. Here we can see a greater area of green meaning there was a great amount of change in this area between the years. This change is the renovation and expansion of the Reagan International Airport, which took place in the 90’s. You may also notice shades of green in the area between the rivers’ divide, probably signifying there was more development of commercial areas between the years. The Write Memory Function Insertion allows you to have an overall detection of change in brightness values which can be used as a reference when doing further classifications and change detection.

The next step in my overall change detection was creating a supervised classification. For the supervised classification I created 6 regions of interest (ROI’s) which consisted of: older residential, newer residential, commercial, parks & recreation, forest, and water. I chose these six different types of land-cover because it would allow for a more specified classification, especially in areas of commercial and residential change which may seem similar.

When choosing my ROI’s I used the 1991 image because we are trying to detect change from this point on. I referred to the National Land Cover Dataset classification created in 1992 as a guide because although it is not completely accurate, it more accurately divides these different types of land-cover. I also looked at satellite images on Google Earth to further ensure I was choosing correct regions by zooming in and out of the Washington D.C. area and comparing them to both my color composite image of 1991 and the National Land Cover Dataset of 1991. However, when looking at the Google Earth images, I used images from 1999 because this was the closest date to 1991 that was not in grayscale. The grayscale images made it hard for me to distinguish certain areas like residential and commercial or forest and parks, therefore I did not use the grayscale images even though they were closer in date to the 1991 image.

After creating my ROI’s I proceeded to performing a Maximum Likelihood supervised classification and a Minimum Distance supervised classification for both my 1991 and 1999 images. I was able to do my classifications through the Classification function in ENVI. When performing the classification, I spectrally subset bands 5, 4, and 2, because these are the least correlated bands and therefore the optimum bands to use for classification in terms of information. After comparing both classifications, I decided that the Maximum Likelihood classification did a better job than the Minimum distance. The Maximum Likelihood classification assumes a normal distribution of pixels, and places pixels in each class based on its probability of belonging in that class while the Minimum Distance classification uses mean vectors of each endpoint and calculates the Euclidean distance of the pixels to the mean vector of each class and places the pixel in the class closest to it. The Maximum Likelihood classification seemed to work better and this may have been because of the ROI’s I selected or the method it uses to place pixels in each category, or a combination of both. After choosing my method of classification, I edited the class colors/names to produce my final classified images of both 1991 and 1999 which can be seen in Figure 3. These images show the different regions classified using different colors which will later be used to detect changes by land cover type.

When comparing the classified image for 1991 and 1999, you may notice more red in the area between the fork of the river. This means there was more new residential in 1999 than in 1991. This may be true if the population was expanding, causing areas that were previously commercial to change to newer residential in order to accommodate the growing population. If this was not the case, more areas of newer residential may have been classified in 1999 due to the ROI’s I selected and how I selected them. The fact that I used an image from 1999 as reference in Google Earth may have created some uncertainties when I was creating my ROIs and this greater amount of newer residential may be due to this.

Another area that has changed between 1991 and 1999 is in the Northwest corner of the image. Here you may see a greater amount of older residential in areas that were previously forest in 1991. Some of the forested areas may have been cut down and turned into residential areas, but somehow these areas were classified as older residential than newer residential. If this was the case, these areas may have been classified as older as opposed to newer residential because they are located in areas that were once forest and therefore there are lower brightness values due to shading of nearby trees or dirt as opposed to concrete, etc.

There is also a change in the area where the Reagan International Airport is located. When comparing the 1991 and the 1999 image you can see the airport is more defined in the later image. The airports’ landing strips are more defined in the later image partially because the second image picked up the grass located between the river and the airport and classified it as parks & recreation. This may be because in 1991 the airport was still under renovation and expansion and therefore the grass was not tended to or watered as often as later in 1999 when the airport was either in the final stages of renovation or open to the public. The healthier grass displayed brighter pixels and was therefore classified under the parks and recreation. This can be seen by a greater amount of green in this area along with the definition of the landing strips. However, the amount of yellow, in this case commercial, is reduced a bit and more red, newer residential, is seen on the left side of the airport. This may be once again, because I used an image of 1999 from Google Earth when creating my ROIs, therefore creating uncertainties between what is newer residential and commercial.

The next step for my final change detection was completing a Change Detection Matrix. I created my Change Detection Matrix using the Post Classification function of ENVI by selecting Confusion Matrix, followed by Using Ground Truth Image. I used my 1991 classified image as the input, and the 1999 classified image as the ground truth. By inputting the 1991 classified image first and the 1999 image as the ground truth, I am comparing the overall change created from 1991 to 1999 which can be seen as “error” bands for each type of land cover based off the classified maps.

The Change Detection Matrix allows us to see statistics about how the pixels in each type of land cover changed between 1991 and 1999 in terms of classification. For instance, looking at the matrix in Figure 4, we can see that 88.47% of the water pixels do not change from 1991 to 1999. We can also see which pixels changed from one category to another. For example, 24.15% of what was classified as commercial in 1991 is now classified as newer residential in 1999. This means that commercial pixels now account for about 1/4th of newer residential pixels in 1999, which makes sense when looking at the classified images, especially in the regions between where the river divides, and the side left of the airport. The classified image of 1999 shows more regions with red than yellow, most likely because according to the change detection matrix 24.15% of the yellow pixels in the 1991 image have now changed to red pixels in the 1999 image. The same can be said for the change between forest in 1991 and older residential in 1999. About 14.08% of the pixels classified as forest in 1991 are now classified as older residential in 1999 which explains the higher amount of maroon in the Northwest portion of the image which was sea-green in 1991.

As mentioned earlier, creating the Change Detection Matrix also created bands which display the change from 1991 to 1999. I selected the band showing the commercial change and the band showing the change for parks & recreation as the categories I will use in the final change detection because these are the areas more significant in change detection for this area. These change images can be seen in Figure 5. Looking at these images we can see that there is a greater amount of commercial change than change in vegetation. Areas of commercial change can be seen throughout the image, but certain land features such as streets, the airport landing strips, and the edge of the river can be seen more easily. For the vegetation change image more change can be seen in areas where there was forest like in the Northwest corner and Southeast corner of the images. The area near the center of the rivers’ fork does not hold much change in vegetation, but it does show a lot of commercial change which may have occurred due to development between the years. These areas of change can be compared with the Write Memory Function Insertion and the supervised classification images and a general trend of change can be seen.

The final step in creating a Change Detection is creating ROI’s from the change images of parks & recreation and commercial areas in Figure 5. In order to do this, I used the Region of Interest feature and selected Band Threshold to ROI. I then selected the “error” band in this case the commercial and parks & recreation change images, and set the min and max threshold values to 1. By setting the thresholds to 1 I was able to create an ROI that I could overlay on any image. For the sake of the change detection, I decided to overlay my new change ROI’s onto the 4th band in grayscale of the 1999 image because this band was the lightest, allowing to new ROIs to stand out more. The final change detection map can be seen in Figure 6, where the red color represents all the pixels that changed to become commercial in 1999, and the green color represents all the pixels that changed to become vegetation in 1999. Once again, you can see a greater concentration of commercial change in the area between the rivers’ fork, but you can also see it along the streets as strips; and at the Reagan International Airport as it defines the landing strip. The vegetation lies closer to the river but is pretty well disbursed around the area. These changes may be a result of new development in the city creating more commercial areas along with healthier well irrigated vegetation. Some uncertainties such as the fact that I used images from Google Earth in 1999 to make my initial ROI’s may have affected my overall change detection because I may have classified areas that were actually commercial as newer residential and therefore it was harder to distinguish the difference between the two in the 1999 image when performing a Maximum Likelihood classification. The type of supervised classification I chose may have also affected how pixels are placed in which category with respect to brightness values and shadows. Overall, the major areas of change are located at the Reagan International Airport, along major streets, areas near the forests, and along the edges of the river.

References

Jensen, John R. Introductory Digital Image Processing: A Remote Sensing Perspective. 2^nd

Ed. Upper Saddle River, NJ: Prentice Hall, 1996. Print.