Lab 4: Supervised Classification of Landsat 7 ETM+ Imagery

Introduction

  The goal of this lab is to to perform supervised classification using Erdas Imagine to produce a land use land cover (LULC) map of Eau Claire and Chippewa counties in Wisconsin. Supervised classification consists of three main steps:

                                         1. Select the Training Samples

                                         2. Evaluate and Refine the Training Samples if Needed

                                         3. Run the Classifier 

Each section in the methods details one of these steps. The supervised classification in this lab is performed on a Landsat 7 ETM+ image captured on June 9, 2000. There are five LULC used to classify the image: water, urban/built up, forest, bare soil, and agriculture. Each pixel in the image will be classified as one of these LULC classes.

Methods

Step 1: Collecting Training Samples
  First, the Landsat 7 ETM+ NIR image was brought into Erdas. Then, the training samples were collected. The training samples are collected by creating areas of interest (AOIs) and then importing each AOI into the signature editor. To do this, one can click on the the AOI to select it in the viewer, and then click the "Create New Signature(s) from AOI" button, which has the symbol "+↳" in the signature editor window to make it a training sample.
  When collecting training samples it is important to collect quality samples. Quality training samples can be taken if the following guidelines are followed.
                                         1. Only allow samples to contain pixels which share the same LULC type.                                                                                           For example, it would be good to collect a training sample that has only a water LULC, and it                                                         would be bad to have a training sample contain water and urban LULC.
                                         2. Training samples should be at least 10 pixels in size because otherwise, a signature plot                                                               cannot be created for the sample.
                                         3. Training samples should be collected from all areas of the image being classified. Don't collect                                                     all samples from a small portion of the image.
                                         4. More Training samples should be collected where there is greater variably in the spectral                                                               profiles of the samples. For example, there should be more training samples collected for                                                               urban/built up than there should be for water.
                                         5. Training samples should include all of the spectral variability for a LULC class.
                                         6. Collect at least 5 to 10 training samples per class at the minimum.

  Figure 1 shows the water and forest training samples that were collected. In total 77 training samples were collected, 16 for water, 22 for forest, 13 for agriculture, 12 for urban/built up, and 14 for bare soil.

Figure 1: Signature Editor Window of Training Samples

Step 2: Evaluate and Refine the Training Samples
  There are a few different ways to evaluate the quality of training samples, in this lab, the histogram, image alarm, using a spectral plot, and looking at spectral separability.
  To use a histogram to evaluate the samples, one can click on the Display Window Histograms button in the signature editor window.  Then, the a histogram can be created for each spectral band included in the image, so because there are 6 reflective bands for Landsat 7 ETM+, there are 6 histograms created. One can create a histogram for a single training sample, or for multiple. It depends on the number of training samples highlighted in the signature editor window. Figure 2 shows the histgram values for all 6 bands for all of the water training samples. As one can see, most of the histograms are fairly normally distributed which is good. It is good to have the histograms normally distributed than it is for them to be bi-model. If 3 or more of the hisograms are bi-model. then, the training samples should be deleted and retaken until at least 4 bands display a normal distribution.

Fig 2: Histogram for all Water Training Samples for all Six Bands

 Image alarm allows one to see the computer take a guess at what pixels will be classified depending on which training samples are highlighted. One can access image alarm in the signature editor window by first highlighting the training samples in the signature editor window for which one wishes to have the computer estimate the classificaiton and then by navigating to View → Image Alarm → Edit Parallelepied Limits → Set → Signatures Selected → Ok. Figure 3 shows what using the image alarm looks like when using the water training samples to estimate the classification. The water appears to line up very nicely with the water in real life, so this means that the water training samples were good. If areas other than water were highlighed in turquoise, then only would should go through and recollect the training samples for that LULC class.

Fig 3: Water Image Alarm

  Next, the signature mean plots were used to see if the training samples were of good quality, To access the spectral plots for selected training samples, one can click on the Display Mean Plot Window button. Figure 4 shows the spectral mean plots for all of the signatures collected for each of the LULC classes. It is ideal to have all the spectral plots line up very similarly, much like the agriculture LULC plot shows below. If the signatures do not line up or are not uniform, then one should revisit the image alarm and histogram analysis to determine whether the signature should be kept or deleted.

Fig 4: Spectral Plots for the LULC Training Samples

   Looking at the spectral plots, there appears to be some variability in the urban/built up LULC class plot, and a bit in the forest LULC plot. The water, bare soil, and agriculture LULC plots all look uniform which is was is intended. Because the urban/built up and forest LULC plots are not uniform, the histograms and image alarm were revisited for these classes. After looking at these it was determined that the samples should be left in as they are all part of their respective LULC class.

  The last thing to evaluate the quality of the training samples is the separability between bands. This can be determined by generating a separability report. A separability report is created by navigating to Evaluate → Separability in the signature editor window, and then by chaning the number of layers to 4. This makes sure the report will determine which 4 bands provide the greatest separability. The report also generates a value which can be found under the Best Average Separability section of the report. This is the most important part of the report. This part of the report for this lab's training samples can be seen below in Figure 5. 

Fig 5: Best Average Separability Report

The numbers on the left (1, 2, 4, and 5) represent the bands which will provide the most separability between differentiating the LULC classes with the provided training samples. the number 1964 represents the quality of the training samples. If that number is above 1900 then the training samples were of good quality which in this case is true. If this number is above 2000, then the quality of the training samples is excellent. If this number is below 1700, then the quality of the training samples is poor. If one records a value below 1700, then he or she should recollect many if not all of the training samples.

Step 3: Classify the Image Using the Supervised Classifier
  This is done by first merging all of the training samples for each LULC class into their respective classes to create 5 signatures. Then, the names of these merged signatures/training samples were named to their respective LULC classes. Then, the signature file was saved.
  Then, the Supervised Classification window was opened inputting the original Landsat 7 ETM+ image into the input raster, the saved merged signature as the input signature file, and then naming the output file. The classifier chosen was the Maximum Likelihood classifier. These parameters can be seen below in Figure 6.

Fig 6: Supervised Classification Input Parameters

  Lastly a map of the output image was created in ArcMap.

Results

  The result of performing the supervised classification can be seen below in Figure 7.

Fig 7: LULC Map of Chippewa and Eau Claire Counties for the Year 2000

  This map looks very similar to the map created using unsupervised classificaiton. Overall, by just visually looking at the LULC map and knowing what the LULC is for the area, the output image is ok. The largest difference in this LULC map from the LULC created in Lab 3 is that there is a lot more area classified as bare soil in this map. This is because in this lab, fallow agricultural fields were trained and chosen to be classified as bare soil whereas in Lab 3, these fallow fields were chosen to be classified as agriculture LULC.

Sources

United States Geological Survey. (2017). Earth Resources Observation and Science Center
Wilson, C (2017) Lab 3 Unsupervised Classification Data retrieved from
     https://drive.google.com/open?id=1w2WHUATIj8EfztCA0o6iD8Zv2tdpfvm2