This is the first lab assigned in Geography 438, Advanced Remote Sensing, at the University of Wisconsin-Eau Claire. It focuses on learning how to extract statistical information from satellite images, developing a model to calculate image correlation analysis, and interpreting the results of the correlation analysis for image classification. The main focus of the lab is learning how to identify and eliminate data redundancy from satellite images by applying statistical techniques and analysis. This is a key part of performing image preprocessing.
Methods:
Exploring Data Quality through Feature Space Plots:
The first technique of exploring data quality used was looking at feature space plots (Figure 1). These images help show whether or not two bands may be highly correlated. If they do appear to be highly correlated, there may be redundancy present and one of the bands should be eliminated or further statistical tests should be run to detect correlation.
An image of the Eau Claire area taken in 2007 was added to the viewer in ERDAS Imagine. From here feature space plots were created using combinations of all of the available bands by using the Raster toolbar and looking under Supervised. By making feature space plots of all of the available band combinations, bands that may correlate and therefore be redundant can be identified (Figure 2). Also bands that have a high amount of variation can be located as well (Figure 3).
 |
| These are all of the feature space plots created by running the tool in ERDAS Imagine. The plots work by plotting the reflectance of one of the bands on the x-axis and the other on the y-axis. As can be seen some of the plots show a large amount of variation and little correlation between the bands, while others appear to be highly correlated with a one to one relationship. (Figure 1) |
 |
| This feature space plot shows the relationship between the reflectance in bands 2 and 3. These two bands appear to be highly correlated and one of them may need to be eliminated in order to reduce redundancy in the image for future use. (Figure 2) |
 |
| This feature space plot shows the relationship between the reflectance in bands 4 and 6. These two bands appear to be greatly varied and have a low correlation. (Figure 3) |
Assessing Image Quality through Correlation Analysis:
The next part of the lab involved creating a model to run correlation analysis on the same image that feature space plots were created for. Creating feature space plots is a good idea to explore whether or not correlation analysis should be run, while correlation analysis gives more finite information about the bands and whether or not there is redundancy present.
The first step was to open up model builder and begin constructing the necessary model (Figure 4). This model was rather simple to construct as all that was required was an input, a function to calculate correlation (Figure 5), and an output matrix table.
 |
| This is the model that was designed in order to perform correlation analysis on the image. As it can be seen, this model is rather simple and only has one input, function, and output. (Figure 4) |
 |
| This is the function that was performed to calculate the correlation of the various bands of the image. The input can be seen here along with the option to ignore the value zero, which is necessary in this case. (Figure 5) |
After the model was run an output matrix was created and cleaned up to look professional the results could be easily seen and the bands with the highest correlation could be found (Figure 6). Correlation is measure on a scale of -1 to 1. The closer the number is to 1 or -1, the higher the correlation. The closer the correlation value is to zero, the less correlation present. If two bands have a correlation value of greater than 0.95, one of them should be eliminated as to avoid redundancy.
 |
| This is the correlation matrix generated by the constructed model for the image of the area surrounding Eau Claire. The two bands with the highest correlation and the most redundancy are bands 2 and 3. This could mean that one of these bands should be eliminated before proceeding with more image processing/analysis. However, the value isn't greater than 0.95, which is the typical cutoff so it may not be necessary if it is felt by the analyzer that both of these bands will be instrumental in further analysis. (Figure 6) |
 |
| This is a high resolution image taken of an area in the Florida Keys which was analyzed using correlation analysis. (Figure 7). |
 |
| This is a high resolution image taken of an area in the Sundarbans which was analyzed using correlation analysis. (Figure 8) |
 |
| This is the final correlation matrix of the Florida Keys image. It can be seen that Band 1 and Band 2 are highly correlated and it should be highly considered to eliminate one of them to reduce the redundancy before proceeding with other analysis of the image. (Figure 9) |
 |
| This is the final correlation matrix of the Sundarbans image. It can be seen that Band 1 and Band 2 are highly correlated and it should be highly considered to eliminate one of them to reduce the redundancy before proceeding with other analysis of the image. Figure 10) |
Conclusion:
It is important when performing image preprocessing to check for redundancy in an image. This can be explored initially by creating a feature space plot which will show if correlation analysis may be necessary. If correlation analysis does appear to be necessary then it can be easily run by creating a model. Once the correlations analysis has been run and the output table created, it can be seen which bands may be redundant by looking at the correlation values and observing how near they may be to 1. From here, redundant bands should be eliminated before moving on.
