Lab 4: Unsupervised Classification

Goals

The goal of this lab is to help the image analyst to extract information of both biophysical and sociocultural features in remotely sensed images. In order to do so, the use of unsupervised classification algorithms will help to gain this information. The two main objectives of this laboratory exercise are to execute unsupervised classification by inputting the correct requirements for such an algorithm as well as the ability to recode spectral clusters into useful information on land use and land cover.

Methods

Experimenting with Unsupervised ISODATA Classification Algorithm
This process uses ISODATA, iterative self-organizing data analysis technique algorithm to produce an image covering both the Eau Claire and Chippewa Counties.

The first step in this process is to set up an unsupervised classification algorithm. Once the original image is opened in ERDAS Imagine 2013 and select the unsupervised classification tool under the raster toolbar. We will use our original image as the input file and we will change the number of classes to 10-to-10. This means that the algorithm will classify the brightness values within the image into a total of 10 different categories. Next, the maximum number of iterations to 250. This means that the algorithm will run up to a total of 250 times in order to make sure it does not group unlike features together (Fig. 1).Once this is complete run the model and compare the input and the output images (Fig. 2).

(Fig. 1) This image shows the input image as well as the Unsupervised Classification window that is being used to classify the data in the original image. The adjustments to the formula have been seen as described above.

(Fig. 2) The input image can be seen on the left while the output image is on the right. The output image has undergone an unsupervised classification algorithm. 

The next step of the process is to recode the clusters produced by the unsupervised classification algorithm into meaningful classes which will show land use/land cover of Eau Claire county. To do so, we will be opening the image attributes. Next we will recode the classes. The best way to do this is to select each cluster individually and change the color to a bright yellow so it stands out (Fig. 3). Then we will decide which category it belongs to and re-color it. The class names and colors we will be using area as follows: Water-Blue, Forest-Dark Green, Agriculture-Pink, Urban/built-up-Red and Bare soil-Sienna.

(Fig.3) This image displays the image which had previously undergone the unsupervised classification and has been recoded according to the class names and colors as described above.

Improving the Accuracy of Unsupervised Classification

One of the main problems that arose from the first classification process is that it was difficult to identify the differences between the forest and agriculture land use/land cover classes. Since this is the case, it is not clear which classes are correct and which are incorrectly classified. For this reason, we will be running the same process as we did above, however, we will be changing the minimum and maximum number of classes to 20-to-20 as opposed to the 10-to-10 we did in the above section. This should help to make a more accurate map when it comes to classifying the unsupervised clusters. The same process is repeated, the algorithm is run and the analyst recodes the image in order to select the appropriate classes for the clusters. Once this is complete we can compare the output image from the previous section and the new output image to see if changing the number of classes had an effect on the map and classifications (Fig. 4).

(Fig. 4) The first output image can be seen on the right which used 10-to-10 classes to perform the classification process and on the left is the image produced after using the 20-to-20 classification.

As can be seen in the comparison of the two images in Fig. 4, by increasing the number of classes in the unsupervised classification algorithm the classification of various LULC (land use/land cover) classes can be more accurately determined. The original output image was challenging to classify because the 10 clusters which the algorithm produced grouped much of the agriculture and forest land together. This lead to the image being dominated by agriculture (represented by the color pink) which is not the correct distribution of land in Eau Claire County. After more classes were added it was easier to accurately identify LULC classes because the clusters were more accurate. 

Another way that we can further organize the classes is to change the column properties in the raster attributes. This will allow us to chose which order we want the columns to be displayed in when we view the attributes of the data. 

Then we will recode the LULC classes in order to make it easier to generate maps using this data. To do so we will select the thematic tab under raster and select the recode tool. The next step is to change new values so that all the like LULC classes are grouped together. For example, the map has a number of classes which were classified as agriculture but rather than having multiple classes for the same LULC we will change the values for all the agriculture to be 3. (Water will be 1, forest will be 2, urban/built-up will be 4 and bare soil will be 5). This way when we go to create a professional looking map using this data it will be much easier. In the last step we will use ArcMap to produce a map which presents the classification data in a more understandable manner for viewers (Fig. 5).

(Fig. 5) The image above shows how the classification work done in ERDAS Imagine can be used to create a visually pleasing map which is easier to interpret by the viewer. 

Results

As a result of completing this laboratory exercise the image analyst learned a number of important techniques for classifying remotely sensed data, particularly for land use/land cover. Unsupervised classification was used to produce all the above images, however to better understand how the algorithms used in this process work we made changes to see how it effects the images and therefore the classification. In the first portion of the lab, we used 10-to-10 classes in the algorithm and tried to classify the LULC (land use/land cover) classes from there, however, it was difficult. After completing the second portion of the lab exercise where we used 20-to-20 classes in the algorithm we gained more insight into how the number of classes can affect the overall accuracy of the classification of the image. In the end we learned that the greater the number of classes used in the unsupervised classification algorithm the easier to identify LULC classess. 

Sources

Data for this lab was collected from: United States Geological Survey and Earth Research Observation Science Center.