Landsat 8 Visible Spectrum Images

My previous articles examined the simulated Landsat 8 bands offered by the USGS and then afterwards the first "pre-production" image of Fort Collins CO USA. Since that time, large quantities of production Landsat 8 data have been made available by the USGS and are archived at GLOVIS. This article looks at two such Landsat 8 scenes and compares them to corresponding Landsat 7 scenes.

I chose a South American scene near Antofagasta, Chile for my first study. The image below is an unprocessed RGB composite image of a subset of that scene with corner coordinates:

ULX is: 339090
ULY is: -2540850
LRX is: 365850
LRY is: -2564040

An unprocessed band 2, 3, 4 RGB color composite image is shown below.

Landsat 8 bands 2, 3, 4 RGB color composite, unprocessed.

A corresponding Landsat 7 band 1, 2, 3 RGB color composite of the same coordinates is shown below.

Landsat 7 bands 1, 2, 3 RGB color composite, unprocessed.

The Landsat 8 image appears to have less contrast, more blue tones and less red tones than the Landsat 7 image. The Landsat 7 image looked a lot better "out of the box" in my opinion. Repeating this procedure using other images showed similar results between Landsat 8 images and the corresponding Landsat 7 image. The next step was to figure out what might be going on.

An examination of the band histograms indicate that for this scene at least, the Landsat 8 bands have narrower histograms than the Landsat 7 bands. Narrower histograms mean less image contrast. As a result, when the bands are turned into color RGB images by compositing or pan sharpening, the colors appear flatter than the corresponding Landsat 7 images. This may be a result of the narrower responsiveness of the Landsat 8 OLI sensors as compared with the corresponding Landsat 7 sensors. Narrower response is an advantage in multispectral analysis, which Landsat 8 was designed for. However narrowing the bands can eliminate spectral areas that the human eye is sensitive to, thereby affecting RGB image color tones.

Landsat 7 histogram on the left, and Landsat 8 histogram for visible blue bands.

The Landsat 7 band 1 histogram is on the left and the Landsat 8 band 2 histogram is on the right. The Landsat 8 band histogram is narrower and overall brighter than the Landsat 7 histogram. The other band histograms compared similarly.

Comparison Landsat 8 Multispectral Bands to Landsat 7. (Credit: U.S. Geological Survey Department of the Interior/USGS).

It is clear from the diagram above that the Landsat 8 bands are narrower than the corresponding Landsat 7 bands accross the spectrum. I attempted to improve the image for human viewing by performing a histogram stretch. This can be done on the input multispectral bands or on the color RGB or pan sharpened image. I recently added a color proportional histogram stretch feature to PANCROMA™ to complement the grayscale proportional stretch. To use it select 'Pre Process' | 'Histogram Stretch' | 'Proportional Stretch' and either 'Grayscale Image' or 'Color Image'. The Landsat 8 RGB composite image shown above has been stretched using this feature. The stretched image is shown below.

Landsat 8 image proportional stretch.

Although the image has been improved by the histogram stretch, the color tones appear somewhat 'off' compared to the Landsat 7 image. The aforementioned lower contrast in the Landsat 8 image and an apparent lack of red color tones over the land areas is probably caused by the narrowing of the sensitivity of the red band Landsat 8 sensor in particular. Surface features reflecting in the 600nm range that the human eye perceives as red is not detected by the Landsat 8 sensor. An image will appear "natural" only if the measured frequencies can be linearly transformed into the response curves of the human eye. If a section of the visible spectrum is not detected at all, this is not possible. This appears to be the case for Landsat 8 images.

I tried several remedies for this problem and determined that two corrections are necessary: increasing the contrast and gamma correction. Contrast can be improved several ways. Some methods available using PANCROMA™ utilities include simple contrast adjustment, histogram stretch, image deconvolution, and haze reduction using the Dark Area or HOT methods. I found that proportional histogram stretch (not equalization), for example as performed by the PANCROMA™ proportional stretch algorithm , haze reduction followed by gamma correction was most effective. Deconvolution also improved the images, but haze reduction worked better has the advantage of being less computationally expensive and is not RAM limited like deconvolution is. The image below has been processed using the PANCROMA™ a Dark Area haze reduction algorithm to the stretched Landsat 8 bands. This resulted in a much clearer image as shown below.

Landsat 8 stretched and Dark Area haze reduced.

The next image shows the haze reduced image gamma corrected with a gamma factor of 1.47, with the red channel boosted slightly to compensate for the aforementioned "hole" in the Landsat 8 red sensor. I used the new PANCROMA™ gamma correction feature to make the adjustment. The image can be further improved but I was satisfied in the result of only a couple of minutes of processing that yielded a pleasing image that might be useful for many purposes.

Landsat 8 stretched, Dark Area haze reduced and gamma corrected.

In order to determine if the image characteristics and remedies were applicable to other Landsat 8 images, I tried it on another image of central South Africa. The unprocessed image for the stated coordinates is shown below.

ULX is: 244440
ULY is: -3102360
LRX is: 269160
LRY is: -3125400

Landsat 8 bands 2, 3, 4 RGB color composite, unprocessed.

The corresponding unprocessed Landsat 7 image is shown below.

Landsat 7 bands 1, 2, 3 RGB color composite, unprocessed.

And finally, the color corrected Landsat 8 image.

Landsat 8 bands 2, 3 and 4 RGB color composite, Dark Area haze reduced, proportional histogram stretch and gamma corrected.

It seems clear that the new Landsat 8 data will pose image processing challenges for those wishing to use the data for visible spectrum applications. This will probably be an area of ongoing research for the immediate future. Preparing pleasing visible RGB color images from Landsat 8 data will take a bit of work, but with the right techniques and PANCROMA™ tools anyone can get acceptable results.

Clarification: I used the term "unprocessed" several times in this article to describe composite images prepared from band data downloaded directly from GLOVIS. Of course this data is processed considerably from the raw Landsat data by the USGS prior to archiving. I am referring to subsequent image processing from the L1B archived data. Also, keep in mind that when it comes to visible images, everything is subjective. Results depend on atmospheric conditions, lighting, terrain, the monitor the image is displayed on, and the particular human being that is viewing the image. This article is not meant as a criticism of any data, only a technical approach to modifying its appearance.