In my recent research note comparing NDVI images created using blue and red filters (http://publi...
Public Lab is an open community which collaboratively develops accessible, open source, Do-It-Yourself technologies for investigating local environmental health and justice issues.
Public Lab chatroom
Reset your password
Read more: publiclab.org/n/9717
In my recent research note comparing NDVI images created using blue and red filters (http://publiclab.org/notes/nedhorning/10-30-2013/red-vs-blue-filters-for-ndvi) I noted that reflected red wavelengths of light likely have more useful information for detecting differences in plant vigor or health when compared with blue light. In addition to the images in that research note I wanted to explain the logic behind the improved NDVI performance when a red filter is used in place of a blue filter. In this note I'll use spectral reflectance curves to illustrate why there is greater contrast in NDVI when using a red filter.
A spectral reflectance curve provides information about the intensity of different colors of light reflected from a surface of an object. When plotted as a graph we can see how reflectance varies across a range of wavelengths for different materials. Scientists use this information to identify different materials. For example, geologists have libraries of spectral reflectance curves for hundreds of different minerals and by measuring spectral reflectance of a sample collected in the field they are able to identify the different minerals in the sample.
Spectral reflectance curves typically have wavelength on the x-axis and percent reflectance on the y axis. Percent reflectance is simply the percentage of energy (light) for a particular wavelength that is reflected off of a sample divided by the intensity of energy for that wavelength that illuminated the sample. In other words it's the outgoing energy divided by the incoming energy for a given wavelength.
You can see reflectance curves of different materials and how they related to different satellite images using this interactive: http://biodiversityinformatics.amnh.org/tool.php?content_id=145.
Below are a few reflectance curves for different materials. The circles represent the approximate part of the curve that is being recorded by the blue channel in a camera with a blue filter (leftmost circle ~450 nm), a red channel in a camera with a red filter (center circle ~650 nm) and the channel that records NIR (rightmost circle ~850 nm). One thing to keep in mind is that the circles represent a relatively small range of wavelengths and the wavelengths of light recorded by a single band in a photograph will be much broader. I expect the red circles are close to the center of the wavebands of a photo and it probably roughly represents what is recorded by the camera.
Reflectance curve for green lawn grass
Reflectance curve for dry grass
Reflectance curve for tar paper
Reflectance curve for pine wood
Although the reflectance curves for these materials could change from one sample to the next the general shape of the graphs are typical for these materials. For example, the maximum reflectance for different species of grass can vary significantly but the basic shape of the graph with the little bump at 550 nm and the steep curve from roughly 700 nm- 750 nm is typical of nearly all healthy green vegetation.
When you look at the spectral curve for green grass you can see that the difference in reflectance between blue (the circle furthest to the left) and NIR and the difference between red (the circle in the middle) and NIR is roughly the same. Therefore one would expect the same or similar NDVI value regardless if a blue or red filter was used. The NDVI calculated using the blue band was 0.85 and for the red band 0.82. This realization is probably why the use of blue filter cameras to create NDVI has increased in popularity.
Now look at the curve for dry grass. Since the circle for blue and red wavelengths are not at the same level or reflectance the NDVI values will be quite different. The reflectance of dry grass in the blue band hasn't changed that much but the reflectance in the red band has increased considerably. The NIR reflectance has decreased. The NDVI for dry grass when using the blue band is 0.47 and when using the red band it is 0.15. In both cases the NDVI for dry grass is less than the NDVI for green grass but when we use the red band the difference is much greater which will result in an NDVI image with greater contrast between healthy and unhealthy vegetation. When using remotely sensed data to detect or monitor healthy vegetation contrast is important and we should do what we can to maximize it.
This green grass curve highlights a common misconception that green vegetation reflects a lot of green and NIR light. As you can see only about 10% of the green light is being reflected and I certainly don't consider 10% to be "a lot". Grass looks green to us simple because it reflects more green light than blue or red and our eyes are particularly sensitive to green light. It's not because if reflects a lot of green light.
These curves are made up of hundreds of points and each point is a measurement that was made from a spectrometer. When we take a photo using our blue or red filter cameras the curve we would get is just a straight line since we only have two points – on for NIR and another for either red or blue wavelengths. If we use the dual-camera NIR setup we get 4 bands (red, green, blue, and NIR) which can help in identifying features based on their spectral properties. There are image sensors called hyperspectral sensors or imaging spectrometers capable of recording hundreds of very narrow bands but those are very expensive. It was interesting to see the Adafruit website recently mention that adding a blue filter will turn your camera into a hyperspectral imager (http://www.adafruit.com/products/1567#Description).
Although it's not evident from reflectance curves another problem relying on blue light to record NDVI is that blue wavelengths of light are more easily scattered than the longer red wavelengths. Chris Fastie highlighted that fact in one of his recent research notes: http://publiclab.org/notes/cfastie/10-23-2013/ord-sfo. Images recorded using red light will be clearer than images recorded using blue light. The intensity of this effect depends on the types and amounts of particulates in the atmosphere as well as the distance between the camera and the target but it can be a problem with kite and balloon aerial photography especially if the camera is looking at an oblique angle.
One issue that has not been discussed so far is leakage of “undesirable” wavelengths of light into the photo bands. With the red filter we can be fairly confident that the blue channel contains primarily NIR light since nearly all of the blue and green light is removed by the red filter. The red band, on the other hand, records red light but also a good deal of NIR light that "leaks" in. This is because both the red Wratten 25A filter that I inserted in the camera and the red filter that is part of the Bayer filter (http://en.wikipedia.org/wiki/Bayer_filter) allow NIR light to pass. A similar problem occurs with the blue filter but it is somewhat worse when we use a red filter. This is a potential drawback to using a red filter. I'm looking into ways to effectively remove NIR light impacts from the red band but my initial thinking is that if the bands are calibrated properly the impact of NIR light leakage will be minimal in most cases. Much on the leakage impact is removed through the calibration process. I'll try to post a research note explaining light leakage and possible ways to deal with it in the not too distant future.
If you are interested in browsing reflectance spectra I have a few URLs below of libraries that allow you to view graphs and download data. It's interesting to notice the blue and red reflectance properties of different materials.
Vegetation Spectral Library: http://spectrallibrary.utep.edu/
Site has ~ 235 spectra from the field. The have photos of the cover type and ASCII data with radiance (incoming light) and irradiance (reflected light). To calculate percent reflectance divide irradiance by radiance
ASTER Spectral Library - Version 2.0: http://speclib.jpl.nasa.gov/
Only 4 vegetation typs are available [Dry grass, grass, conifer, deciduous) but it has spectral from 84 man-made materials. Data file only has percent reflectance. It includes data from includes data from three other spectral libraries: the Johns Hopkins University (JHU) Spectral Library the Jet Propulsion Laboratory (JPL) Spectral Library, and the United States Geological Survey (USGS - Reston) Spectral Library.
Spectral Library Project by the Joint Fire Science Program: http://www.frames.gov/partner-sites/assessing-burn-severity/spectral/
Spectral libraries with data for green (photosynthetic) vegetation, non-photosynthetic vegetation, soil and/or rock, and char and/or ash for three different regions (Western Montana, Interior Alaska, and Southern California). Data includes photos of cover type, ASCII file of percent reflectance.
USGS Spectral Library: http://speclab.cr.usgs.gov/spectral-lib.html
This is a comprehensive set of libraries.
This is great, Ned-- thanks for illustrating it with your examples. The stark difference between NDVI of dead grass in the red channel is a pretty convincing reason to go for this strategy. I'm looking forward to your post on calibration- I'd like to see how we can minimize leakage.
This is really good information. Another thing that might make red filters a good choice is that it is easier to make a red filter than an infrablue one. A red filter just has to pass all wavelengths longer than 600 nm and block shorter ones. An infrablue filter has to be a gap filter, blocking red light (600 to 700 nm) but passing both longer (NIR) and shorter (blue) wavelengths. Red filters with steep cutoff curves at 600 nm have been used for photography for decades and are available in inexpensive polyester versions. Gap filters which block only between 600 and 700 nm have no common use in photography and are available only in expensive glass versions. Both types of filters are approximated by inexpensive theatrical lighting filters, but the spectral properties of the red ones seem to be closer to what we need than the blue ones.
Ned, this is good evidence that some materials, like grass clippings and pine boards, will be more easily distinguished from live green plants in NDVI images using the red channel instead of the blue channel to represent visible light. Cardboard will also be easier to distinguish. Other examples are tree leaves that have turned red or orange in the fall, and green vs ripe tomatoes. There is an unstated assumption that this pattern is widespread and that using the red channel for NDVI will usually allow better discrimination between live plants and other things. Is this your conclusion or do we have to test lots of other materials that commonly show up in the NDVI images we take? Aerial images might include dry and wet soil, various rock types, wet and dry sand, and various roofing materials (wood shingles, tiles, composition shingles, painted and rusted metal, tar, slate). Is a red filter always going to give better NDVI results with the surfaces common in aerial or other photos? How often will a blue filter give similar or better results?
Is this a question? Click here to post it to the Questions page.
Is this your conclusion or do we have to test lots of other materials that commonly show up in the NDVI images we take? I have looked at a lot of spectra and didn't see many cases where a blue filter would be a better choice than a red filter. It would be good to hear from other folks about exceptions. I'm sure there are exceptions but for my interests the ability of a red filter to more easily distinguish between healthy and less-healthy vegetation is important. My results indicate you are getting higher precision with a red filter than with a blue filter in detecting plant stress.
Is a red filter always going to give better NDVI results with the surfaces common in aerial or other photos? Probably not but for my interests it's unlikely that I care. If a roof produces NDVI similar to healthy vegetation I'm not going to be too worried. NDVI is a tool that is reasonably good at detecting plant stress and a few other things but it's not an all purpose tool.
How often will a blue filter give similar or better results? Twice
A blue filter will give better results if red or yellow flowers are present because the flower reflectance mixes with the leaf reflectance and deflates "red" ndvi.
Thanks very much for posting this research note. I found it to be helpful background for a related topic (http://publiclab.org/notes/WhiteRabbit/07-25-2014/can-infragrams-provide-useful-measurements-of-thirsty-my-lawn-is).
I have a infrared-visible camera, how i get the NDVI image?
Here are a couple links that should help you get started:
You said ...With the red filter we can be fairly confident that the blue channel contains primarily NIR light...
But how this happens ? As I know, the Bayer RGB act as bandpass filter therefore the blue channel receive just blue wavelength which is far from NIR. Could you explain ?
Thank you so much
All three Bayer filters (R,G,B) transmit some near IR light.
Above: Spectral response of a Canon PowerShot A2200 before (top) and after (bottom) the IR cut filter was removed.
With the IR cut filter removed and a red filter blocking all blue light, the blue channel will capture NIR and maybe some green or UV depending on the filters.
thanks to @cfastie for the interesting charts... How did you take it ?
Looking at chart, using just the blue channel as NIR the NIR/VIS sensitivity is about 0 at 750nm this will cut NDVI in this range due to camera sensitivity.
But using a Wratten 25A filter and the blue+green channels as NIR the green band might compensate the lack of signal in the range 730-800nm and reduce the effect of NIR light that "leaks" in the red band. Is there any test for this ?
Finally I think it depends a lot from camera sensitivity without the IR filter.
The spectral response figure came from an article linked in this thread: https://groups.google.com/forum/#!topic/plots-infrared/aJhM30D6bUM
That's a good point that the range of NIR wavelengths captured differs between the blue and green channels. But so does the range of visible light differ between them. With a Wratten 25A in place, the blue channel is going to capture very little visible light, but the green channel is going to capture quite a bit of red light. Eliminating all of this cross contamination is very hard with a single camera system.
The most direct way to check if your procedure for computing NDVI is good is to compare it to NDVI from a known system. This is what @claytonb did here: https://publiclab.org/notes/Claytonb/08-13-2016/plant-health-ndvi-white-balance
You must be logged in to comment.