Lab 4: Pixel-Based Supervised Classification

Pixel-based Supervised Classification

Goal

The goal of this lab is to continue in learning the art of extracting biophysical and sociocultural information from remotely sensed images through pixel-based supervised classification. This is one of the most important skills in remote sensing. In the first portion of the lab, skills will be developed in selecting training samples to train a supervised classifier. Next the quality of the signatures will be evaluated. Then meaningful Land use/land cover classes will be produced in the final output. A Landsat 7 (ETM+) image captured on June 9, 2000 will be utilized with an Area of interest covering Eau Claire and Chippewa counties this image is ec_cpw2000.img.  

Methods

Part 1: Collection of training samples for supervised classification

In this first part of the lab, training samples (spectral signatures) are collected to train a maximum likelihood classifier which will be used to classify the ec_cpw2000.img image. This will be done through taking a minimum number of samples from 5 categories. The categories that will be sampled and the minimum number of samples are; Water (12), Forest (11), Agriculture (9), Urban/Built-up (11), and Bare Soil (7). This process can be done during field work or alternatively from high resolution aerial imagery. In this lab it will be done using a high resolution image. In this case Google Earth Pro is the high resolution imagery, however the closest imagery google has is from 2005. To gather these spectral signatures, the drawing tool and signature editor were used (Figure 1). The drawing tool is used to create a polygon around similar pixels that are a part of one of the 5 categories listed above. This area of interest is then brought into the signature editor and classified as Water. Then the rest of the 50 samples will be collected.

Figure 1: This image shows the process of gathering areas of interest to input into the spectral signature editor. In this image specifically, a water sample is taken.

Part 2: Evaluating the quality of training samples

The next portion of the lab is analyzing the signatures collected to make sure they are accurate enough to be used to train a supervised classifier. This will be done through analyzing Mean Plot Windows which show a specific trend for specific surface materials. Ensuring a high spectral separability will allow for the classification to be more accurate. There are 2 stages to evaluating signatures, first the signature mean plot window is used, and then an image alarm can be performed. The mean plot (Figure 2) allows for the user to individually compare samples to see if they match a normal reflectivity. The Image Alarm highlights the pixels that would be classified together, this can be overlaid over the image and can be checked for accuracy. If there are classified pixels showing up over the wrong area, then that sample that is causing the problem can be deleted and taken again to correct for error. Then a Signature Separability report is ran to provide a score out of 2000 which gives an understanding of how well the different categories are separated. This signature separability report also tells the user what 4 bands show the greatest separability, these will also be the bands that the supervised classification tool will use to create a classified image. After evaluating the samples and making sure they are accurate, then the categories were merged together to create only 5 signature, but based off the 50 that were gathered. 

Figure 2: These are the mean plot windows for each classification category. 

Part 3: Performing supervised classification

Finally a Maximum Likelihood supervised classification is performed utilizing the Merged Signatures. The maximum likelihood classifier utilizes statistical probability to determine based off user’s samples what pixel falls into each category. After this classification was ran, a Land use/Land Cover map is created in ArcMap

Results

Part 1: Collection of training samples for supervised classification

After collecting the samples, the Signature editor window will contain all the samples as shown in Figure 3.

Figure 3: This figure shows all the minimum number of signatures for each classification in the Signature Editor window.

Part 2: Evaluating the quality of training samples

When the quality is finished being analyzed, and the seperability report was ran, the report gave a score of 1972 and the most seperability was seen in bands 1,2,4,6 (Blue, Red, SWIR, and SWIR2) (Figure 4). This seperability score is good. This means that hopefully the supervised classified image output will be fairly accurate. The merged class mean plot window is shown in Figure 5.

Figure 4: This shows the output score from the average seperability report.

Figure 5: This is the merged mean plot window. This shows water (Blue), Forest (Green), Agriculture (pink), Bare Soil (brown), and Urban (red).

Part 3: Performing supervised classification

Once the supervised classification was performed, a classified image was produced and a map was made from it (Figure 6). This classification seemed to work okay but is not 100% accurate like one would hope, error is seen especially along the Eau Claire River where the banks are considered urban area, and not water.

Figure 6: This is the Land use Land Cover map that is produced through the Maximum Likelihood supervised classification method.

Sources

Landsat satellite image is from Earth Resources Observation and Science Center, United States Geological Survey.