Advanced Remote Sensing: Lab 1 Image Quality Assessment and Statistical Analysis

Goals and Background

This is the first of many blog posts that will be created for labs in Geography 438 Advanced Remote Sensing of the Environment. A basic introduction to software and review of topics covered in Remote Sensing were given before this lab.

The main goal of this lab is to us students with the skills of identifying and removing data redundancy in satellite images through the application of statistical analysis, an important component of image preprocessing. There are three specific objectives included in this lab: 

1) Learn how to extract basic statistical information from satellite images

2) Learn to develop models to calculate image correlation analysis 

3) Interpret the results of correlation analysis for image classification. 

Methods

The first portion of the lab dealt with the first goal of learning how to extract simple statistical information from satellite imagery. Two of the methods explored were Feature Space Plot and Correlation Analysis. ERDAS Imagine 2015 was used to conduct this analysis.

Feature Space Plot

Feature space plots are 2 dimensional graphs that allow the user to compare bands of aerial images based on pixel brightness values. Two bands are compared at a time. For this lab we are looking at a satellite image of Eau Claire, WI taken in 2007( Figure 1). This image has 6 bands in it 1,2,3,4,5, and 7 seen in the Metadata (Figure 2). The  Feature Space Plot compares each band with each other band, so for this image there are 15 comparisons that ERDAS computes. To run this analysis in ERDAS you go to the tool bar at the top of the page and click on Raster --> Supervised --> Feature Space Image. Then simply input the image you would like to analyze and select an output location for the comparisons. Figure 3 is the setup window for this tool. The results from these comparisons are seen below in Figures 6 and 7. 

Figure 1 Satellite imagery of Eau Claire WI and surrounding area collected in 2007.

Figure 2 This is the metadata file for the sattelite imagery in Figure 1. It shows information about how many bands, what data type (8 or 16 bit), and other data info.

Figure 3 This the tool setup box for the Feature Space Plots we created in the first portion of the lab.

Correlation Analysis

Feature Space Plots a preliminary comparison of data in the imagery telling the researcher whether or not correlation exists in the data they are examining. To see how much correlation exists and get a more in depth look at the data correlation analysis is done. Unlike feature space plots which only allow you to compare two bands at a time correlation analysis allows the user to compare every single band and output it into a matrix or table displaying all of the correlation values for each band comparison. This is done through the creation of a model. The model consists of an input image, definition or desired calculation, and output values which in this case are correlation values. These values range from -1 to 1. Values of .95 and above mean that there is very high correlation between the bands and there is most likely data redundancy occurring. It is then up to the analyst to choose which band, based on the type of research they are conducting, to exclude from further analysis. Correlation values of less than .95 in general mean that there is little data redundancy between the two bands and that they both can be included in further analysis. Values close to 0 or negative tell the analyst that there is very little or no correlation between two bands in which case they would be both be kept. For this lab this correlation analysis was conducted for 3 separate images. Figures 4 and 5 are the setups for that analysis. The correlation results or matrix are found below in the results section (Figures 9,11 and 13).

Figure 4 This is the model created to
conduct the correlation analysis.
It contains an input image, tool or
definition and then the final output
correlation values. This model is set up
for the Eau Claire imagery in Figure 1.

Figure 5 This is the box where the definition or analysis selection happens.
You can see that a correlation definition has been selected with Eau Claire
imagery as the input image. To run a different image simply switch out the
Eau Claire image with new image.

Results

The resulting plots from running the Feature Space Plot analysis are below in Figure 5 and 6. Each one of the those windows is a comparison of two bands. Those colorful plots are showing how high the correlation between any two bands is. A plot that is very linear and compact such as the 2nd from the right in Figure 6 shows two bands with high correlation which means similar data was collected in those two bands and one of them should possibly be removed before further analysis, but that is determined by the correlation matrix values. Again these plots are a preliminary look at the data in terms of correlation. In Figure 7 looking at the 2nd plot from the right we see that the two bands being compared are not highly correlated because the plot is very spread out and scattered.

Figure 6 These are the first 7 plots created from the Eau Claire satellite imagery collected in 2007. 

Figure 7 These are the remaining 8 plots. 

When looking at the correlation analysis results we created 3 tables for 3 different satellite images. The model is run and output a text file which is then opened and copied into Excel for formatting. Figures 8-13 are the satellite images with their corresponding correlation matrix.

Figure 8 Satellite imagery of Eau
Claire WI and surrounding areas.

Figure 9 This is the correlation matrix for the Eau Claire imagery. We see that
there are no values higher that .95 which means that data redundancy is probably
not occurring between any of the bands in this imagery. This means that all
bands can be kept moving further with image analysis. The row of red 1s are
where each band is being compared to itself. Anything compared to itself is one
so we are not interested in these values.

Figure 11 This is the correlation matrix for the Florida Keys imagery. We see
that the correlation values are fine except for the comparison of bands 1 and 2.
The correlation value of .98 means that there is most likely data redundancy
occurring between these two bands and one of them should be excluded
depending on the type of imagery analysis being conducted.

Figure 10 Satellite imagery of the
Florida Keys.

Figure 12 This is satellite imagery for
the Bengal Province in Bangladesh.
It was collected in 2004.

Figure 13  This is the correlation matrix for the Bengladesh imagery. Similar
to the Florida Keys imagery most of the values are below .95 meaning low to
no correlation or data redundancy. Again however we see that the band 1 and 2
comparison is over .95 meaning one of the bands should be thrown out.

Sources

Landsat satellite image is from Earth Resources Observation and Science Center, United States Geological Survey. Quickbird high resolution images are from Global Land Cover Facility at www.landcover.org