Introduction

In order to follow a systematic method any attend of image was produce following these steps:

1. Band combination or band operation
2. Radiometric enhancement
3. Filter operation.

The band combination/band operation stage takes more of the effort because the other to steps depends upon this first step heavily. This is by far the most important task because is where you can get more variability from the satellite image.

Before performing the actual selection process of band combination or band operation an exploratory analysis was carried out by creating standard band combination images. The first image created was a ‘true’ color image where bands 3, 2, and 1 were presented on the red, green and blue channels respectively. This is a near-real color image but it was not present enough contrast and color differentiation to be really useful. The second testing combination was a color infrared image. Bands 4, 3, and 2 were placed in the RGB channels respectively. This combination was useful for distinguish vegetation from the rest but not for distinguish between different types of vegetation. Other combination like 7,5,4 or 7,5,3 were tested with similar results. One combination that appears interesting was the 5,4,3 RGB.

Now to the serious job!

First Image

Band combination/band operation

From the beginning the main goal was to produce two images that show most of the variance in the scene and subsequently, to distinguish as many different landcover types as possible.

The first step was to select the three bands less correlated (most variance) to each other. To do this, the scattergrams of each of the bands against the other were plotted. From this analysis it was concluded that the three more uncorrelated bands for this scene were the bands 3 (red), 4 (near infrared) and 5 (mid infrared). Once selected the bands, the next step was to select the channels for these bands. The widely used combination of 5 on red, 4 on green and 3 on blue was tested on the scene. The results were satisfactory especially because the vegetation appears green on this combination, which makes the image more easily interpretable.

Radiometric enhancement

Next the radiometric enhancement was performed. A “limits to actual” and “99% stretch” were executed. This provided a very good radiometric enhancement to work with. To increase the contrast in the middle zone of the histogram where normally there are grouped the majority of the values (near the peak of the histogram), a “histogram equalize” was performed. This combination of radiometric enhancements produced a good contrast into the image and was decided to be the final one for the scene. Some other contrast techniques were tested like “custom piecewise contrast stretch” or “Gaussian equalize” but in all the cases the resulted contrast was inferior to the “histogram equalize”. The “custom piecewise contrast stretch” proved to be useful to identify two pr three categories but it is not the best option when you are trying to identify a larger number of categories like in this case.

Filter operation

The last step was the filtering. Here again, a series of different filters were tested. The first important conclusion from the test was that it should be performed a small range filter, i.e. 3x3 or 5x5 maximum. Larger range filters altered significantly the image and as a result, much of the progress done with the previous steps was lost. Low pass filters, high past filters and combinations of both were tested. Finally, a high pass filter was selected to increase the borders and as a result, the contrast of the final image.

Second Image

Band combination/band operation

For this image a statistic approach was tested. A Principal Component Analysis was executed and the first three Principal Components were used to display the image, the First PC on the red channel, the Second PC in the blue channel and the Third PC on the green channel. The second and third PC were switched in order to present the final image in a more intuitive way with the ocean blue and the forest green. The first PC captures 87.94% of all the variance in the scene while the second and third PC capture 4.76% and 3.94% respectively. Together the first three PC represent 96.94% of all the variance in the scene thus this image of three bands is displaying almost all the statistical information of the six bands.

Radiometric enhancement

For this image, an approach similar to the first image was selected. A “limits to actual” and “99% stretch” were executed. Then a “histogram equalize” was carried out to increase the contrast in the middle zone of the histogram where most of the data was concentrated. Once more, some other contrast techniques were tested but in all the cases the results were inferior to the “histogram equalize”.

Filter operation

Like in the past image, a series of different filters were tested. Low pass filters, high past filters and combinations of both were performed. Finally, filter was not used for this image because any filter or combination of filters showed a significance increase in the image quality. This might result from the use of PC as layers because they should take most of the “texture” into the statistical analysis therefore non-significance improvements on image quality can be made with filtering.

Summary Table

Landcover Type	Band combination/band operation	Radiometric enhancement	Filter operation
Urban, high density	False color bands 5, 4, and 3 on channels red, green and blue respectively	· Limits to Actual · 99% Stretch · Histogram equalized	High pass filter (3x3 sharpen edges filter)
Urban, medium Density
Urban, low density
Water
Forest
Irrigated vegetation
Salt scald	Principal Component Analysis PC 1, PC 3, and PC 2 on channels red, green and blue respectively	· Limits to Actual · 99% Stretch · Histogram equalized	None
Farmland
Remnant vegetation

Bonus