This post will detail an intuitive (i.e. non-scientific) color correction method for Himawari-8 web imagery. It was updated on May 12, 2017 with additional information on the revised steps used to modify full resolution data downloaded from NICT ScienceCloud (accessed from this archive page).
[Update: January 18, 2022] The JMA link below now leads to JAXA Himawari Monitor and the product in no way resembles the original example.
The above images are the result of some impractical and not easily reproducible modifications done in a very old version of Paint Shop Pro. I would like to simplify the process using readily available programs like Photoshop or GIMP. The progress so far can be seen in the next post. I will update this post with more definitive information, including sample images, once I am able to produce consistent results that I am happy with across the entire daylight range of Himawari-8 imagery. For the time being this is what I have.
THE GOOD, THE BAD, AND THE UGLY
The full resolution source data—11,000 x 11,000 pixels—as presented by NICT appears to be a simple combining of band 1 (0.47 µm blue), band 2 (0.51 µm green), and band 3 (0.64 µm red) processed to increase brightness, contrast, and saturation.
It should be readily apparent in the above image that soils are far too red. Less obvious is the cyan bias in the color of the oceans. Rayleigh scattering—the same phenomenon that gives the sky its blue color—contributes about 80% of the perceived deep ocean color when viewed from space. So they should appear similar to the deepest part of a clear blue sky, which has a relatively narrow hue range.
The issue with Himawari-8 imagery is the green band filter is less yellow-green than ideal in a RGB array, resulting essentially in a hue shift. Additionally, it is just outside the narrow 0.55 µm chlorophyll reflectance peak making it less sensitive to green vegetation and, coincidentally, mineral soils.
There are a couple of consequences to this. First, cyans will be exaggerated in the green channel while yellows will be suppressed. With cyan being unnaturally emphasized, oceans will be too green. And since most arid soils are predominantly orange, with their yellow component reduced, soils will look too red and somewhat dark. Secondly, verdant regions such as forests and grasslands will appear desaturated or brown.
The hue shift can be better illustrated using imagery from JMA that is not so saturation enhanced. The original coloration is shown in the left panel. A color sample with a hue angle of around 227° was taken from the clear ocean off the northwest coast of Australia.
For demonstration purposes a simple hue adjustment was applied to the entire image until the hue in the same region changed to about 240° as seen in the right half panel. Then the saturation and contrast were increased to get the inset image of Earth. Ocean and soil colors are now more consistent with what we should expect to see.
For the NICT imagery using the GIMP color mixing tool to add about 50% of the red band into the green band can shift the hue from the red/green axis towards the blue/yellow axis. This does not address the desaturated vegetation problem, however. In the conclusion section at the end of this post I discuss the generally accepted workaround for this.
The next issue likely is the result of saturation adjustments made to the processed image which can change the perceived lightness of colors in the HSL/HSV model. When split into their RGB components the land areas in the blue channel are darker than what they should be based on reference data from SSEC.
Notice in the left panel how much Australia stands out from the surrounding waters. In the right panel Australia is barely even discernible to the blue filter.
The contrast between the oceans and land in the red channel and the overall darkness is also much greater than expected even when differences in gamma are taken into account.
To rectify the dark land and ocean areas about 25% of the red band is mixed into the blue band and about 25% of the green band is mixed into the red band.
Finally, since the source imagery is very dark, gamma adjustment will need to be applied. This completes the basic overview of the problems with the NICT Himawari-8 imagery and their proposed solutions.
EDITING THE IMAGE
I will be using a development version of GIMP 2.9 but the procedure should be similar with Photoshop. In Step 3 the percentages may vary according to the way the color model is implemented in your editing software but this should be a general starting point.
And here is the final result:
As you see this is a relatively simple process and it should be applicable in ImageMagick as well. It probably could benefit from some minor selective color tweaking in the blue range but currently the Hue-Chroma tool in GIMP 2.9 is not fully functional. I will try this in Photoshop CS6 when I have the opportunity and update this section if necessary.
So how does it compare to an image of Earth taken from DSCOVR on the same day after adjusting to roughly the same color and contrast?
The DSCOVR image is on the left. The two are reasonably similar but Australia lacks some of the mauve color variations seen in the image from NASA. This could be due to differences in the RGB filters, particularly the green band. Because of Himawari-8's closer proximity to Earth the central Asian deserts appear nearer to the limb and are thus more affected by the atmosphere. I can only speculate on the accuracy of these images since the Apollo astronauts are the only humans to have viewed the full disk of Earth from one of these vantage points but I assume they are much too dark.
CONCLUSION
Overall this technique works well for images around solar noon but less so for images acquired earlier and later in the day. Deserts will look slightly too red and dark again. Color and lightness banding will be pronounced at the terminator and noise is visible at all times around the limb, even solar noon. Still, this is a considerable improvement over the original and the noise can be minimized by gradually clipping the darkest values, but this is not without its own problems.
The channel mixing step will not quite get the oceans to the proper hue however. Adding more of the red channel into the green channel will make land areas too yellow-green. This doesn't appear to be as much of a problem with the SSEC data as far as I can tell. This is very encouraging since it suggests perhaps even more accurate and aesthetically pleasing results are possible.
Ideally it would be best to work with the unretouched distribution data from JMA. NICT ScienceCloud provides full resolution imagery in lossless PNG format but it has been highly modified. In order to do color correction another level of aggressive processing must be introduced on previously processed 8-bit images—not ideal. Also the images as distributed already have significant loss of information in the darkest areas like the terminator which is the cause of the banding errors. I believe the original data is a higher bit depth and that should minimize this problem. Unfortunately that data is not easy to acquire for the casual Earth visualization enthusiast.
Another advantage of having access to the individual bands is the near infrared 0.86 µm "vegetation band" can be added to the green band to boost the chlorophyll and mineral soil signal. The amount of blending should be very subtle since the goal is to produce realistic looking vegetation colors as seen from space through Earth's atmosphere rather than the vivid greens using the Simple Hybrid Contrast Stretch (SHCS) method.
Addendum: of course a far more comprehensive and scientific method of color correction exists. I have appended this post to include the above links because it is a little embarrassing that this amateur blog, useless to scientists, ranks so highly in search results on the subject.
CREDIT
Meteorological Satellite Center (MSC) of Japan Meteorological Agency (JMA)
National Institute of Information and Communication Technology (NICT)
Space Science Engineering Center (SSEC) University of Wisconsin-Madison
NASA EPIC Team
[Update: January 18, 2022] The JMA link below now leads to JAXA Himawari Monitor and the product in no way resembles the original example.
 |
| Credit: MSC of JMA / NICT - Color correction by LoneSky. |
 |
| Credit: MSC of JMA / NICT - Color correction by LoneSky. |
The above images are the result of some impractical and not easily reproducible modifications done in a very old version of Paint Shop Pro. I would like to simplify the process using readily available programs like Photoshop or GIMP. The progress so far can be seen in the next post. I will update this post with more definitive information, including sample images, once I am able to produce consistent results that I am happy with across the entire daylight range of Himawari-8 imagery. For the time being this is what I have.
THE GOOD, THE BAD, AND THE UGLY
The full resolution source data—11,000 x 11,000 pixels—as presented by NICT appears to be a simple combining of band 1 (0.47 µm blue), band 2 (0.51 µm green), and band 3 (0.64 µm red) processed to increase brightness, contrast, and saturation.
 |
| Credit: MSC of JMA / NICT |
It should be readily apparent in the above image that soils are far too red. Less obvious is the cyan bias in the color of the oceans. Rayleigh scattering—the same phenomenon that gives the sky its blue color—contributes about 80% of the perceived deep ocean color when viewed from space. So they should appear similar to the deepest part of a clear blue sky, which has a relatively narrow hue range.
The color of the clear sky (skylight) varies substantially by geographic latitude, altitude, season, humidity, distance from the zenith, time of day and concentration of atmospheric ice, dust or smoke. The distribution of chromaticities is again roughly parallel to the blackbody locus, but the average sky chromaticity is usually above a CCT of 10,000°K, corresponding to a dominant wavelength of about 470 to 475 nm (CIELUV hue angle of about 235° to 245°).
The issue with Himawari-8 imagery is the green band filter is less yellow-green than ideal in a RGB array, resulting essentially in a hue shift. Additionally, it is just outside the narrow 0.55 µm chlorophyll reflectance peak making it less sensitive to green vegetation and, coincidentally, mineral soils.
There are a couple of consequences to this. First, cyans will be exaggerated in the green channel while yellows will be suppressed. With cyan being unnaturally emphasized, oceans will be too green. And since most arid soils are predominantly orange, with their yellow component reduced, soils will look too red and somewhat dark. Secondly, verdant regions such as forests and grasslands will appear desaturated or brown.
The hue shift can be better illustrated using imagery from JMA that is not so saturation enhanced. The original coloration is shown in the left panel. A color sample with a hue angle of around 227° was taken from the clear ocean off the northwest coast of Australia.
 |
| Credit: JMA - Modified by LoneSky. |
For demonstration purposes a simple hue adjustment was applied to the entire image until the hue in the same region changed to about 240° as seen in the right half panel. Then the saturation and contrast were increased to get the inset image of Earth. Ocean and soil colors are now more consistent with what we should expect to see.
For the NICT imagery using the GIMP color mixing tool to add about 50% of the red band into the green band can shift the hue from the red/green axis towards the blue/yellow axis. This does not address the desaturated vegetation problem, however. In the conclusion section at the end of this post I discuss the generally accepted workaround for this.
The next issue likely is the result of saturation adjustments made to the processed image which can change the perceived lightness of colors in the HSL/HSV model. When split into their RGB components the land areas in the blue channel are darker than what they should be based on reference data from SSEC.
 |
| NICT blue channel (left) and SSEC blue channel (right). April 21, 2017 (00:20:00 UTC). |
Notice in the left panel how much Australia stands out from the surrounding waters. In the right panel Australia is barely even discernible to the blue filter.
The contrast between the oceans and land in the red channel and the overall darkness is also much greater than expected even when differences in gamma are taken into account.
 |
| NICT red channel (left) and SSEC red channel (right). April 21, 2017 (00:20:00 UTC). |
To rectify the dark land and ocean areas about 25% of the red band is mixed into the blue band and about 25% of the green band is mixed into the red band.
 |
| Land areas are now much less visible in the blue channel (left). |
Finally, since the source imagery is very dark, gamma adjustment will need to be applied. This completes the basic overview of the problems with the NICT Himawari-8 imagery and their proposed solutions.
EDITING THE IMAGE
I will be using a development version of GIMP 2.9 but the procedure should be similar with Photoshop. In Step 3 the percentages may vary according to the way the color model is implemented in your editing software but this should be a general starting point.
STEP 1: Increase levels to 1.30.
STEP 2: Increase saturation by 15%.
STEP 3: Use the color mixing tool with preserve luminosity turned on and make the following changes:
RED Channel: add 25% GREEN
GREEN Channel: add 50% RED
BLUE Channel: add 25% RED
STEP 4: Increase levels to 1.40.
STEP 5: Use levels to reduced the red channel to 0.95% and increase the blue channel to 1.05%. This step is optional.
And here is the final result:
 |
| Credit: MSC of JMA / NICT - Color correction by LoneSky. |
As you see this is a relatively simple process and it should be applicable in ImageMagick as well. It probably could benefit from some minor selective color tweaking in the blue range but currently the Hue-Chroma tool in GIMP 2.9 is not fully functional. I will try this in Photoshop CS6 when I have the opportunity and update this section if necessary.
So how does it compare to an image of Earth taken from DSCOVR on the same day after adjusting to roughly the same color and contrast?
 |
| Credit: NASA / MSC of JMA / NICT - Color correction by LoneSky. |
The DSCOVR image is on the left. The two are reasonably similar but Australia lacks some of the mauve color variations seen in the image from NASA. This could be due to differences in the RGB filters, particularly the green band. Because of Himawari-8's closer proximity to Earth the central Asian deserts appear nearer to the limb and are thus more affected by the atmosphere. I can only speculate on the accuracy of these images since the Apollo astronauts are the only humans to have viewed the full disk of Earth from one of these vantage points but I assume they are much too dark.
CONCLUSION
Overall this technique works well for images around solar noon but less so for images acquired earlier and later in the day. Deserts will look slightly too red and dark again. Color and lightness banding will be pronounced at the terminator and noise is visible at all times around the limb, even solar noon. Still, this is a considerable improvement over the original and the noise can be minimized by gradually clipping the darkest values, but this is not without its own problems.
The channel mixing step will not quite get the oceans to the proper hue however. Adding more of the red channel into the green channel will make land areas too yellow-green. This doesn't appear to be as much of a problem with the SSEC data as far as I can tell. This is very encouraging since it suggests perhaps even more accurate and aesthetically pleasing results are possible.
Ideally it would be best to work with the unretouched distribution data from JMA. NICT ScienceCloud provides full resolution imagery in lossless PNG format but it has been highly modified. In order to do color correction another level of aggressive processing must be introduced on previously processed 8-bit images—not ideal. Also the images as distributed already have significant loss of information in the darkest areas like the terminator which is the cause of the banding errors. I believe the original data is a higher bit depth and that should minimize this problem. Unfortunately that data is not easy to acquire for the casual Earth visualization enthusiast.
Another advantage of having access to the individual bands is the near infrared 0.86 µm "vegetation band" can be added to the green band to boost the chlorophyll and mineral soil signal. The amount of blending should be very subtle since the goal is to produce realistic looking vegetation colors as seen from space through Earth's atmosphere rather than the vivid greens using the Simple Hybrid Contrast Stretch (SHCS) method.
Addendum: of course a far more comprehensive and scientific method of color correction exists. I have appended this post to include the above links because it is a little embarrassing that this amateur blog, useless to scientists, ranks so highly in search results on the subject.
CREDIT
Meteorological Satellite Center (MSC) of Japan Meteorological Agency (JMA)
National Institute of Information and Communication Technology (NICT)
Space Science Engineering Center (SSEC) University of Wisconsin-Madison
NASA EPIC Team
Comments
Post a Comment