I had time for a quick research project about using CIR(Near infrared) cameras, RGB cameras, and what those sensors can really provide for mapping vegetation. This article's purpose is to show the value of using Color Infrared (Near Infrared) sensors, and some of the misconceptions about just needing a visible light (RGB) camera.
Thanks for some great quotes from:
- Dr. Kevin Price (Agpixel, Iowa State University)
- Dr. Deon VanDerMerwe (Kansas State University)
- Gabriel Torres at Micasense
- Beau Dealy at APIS remote sensing
Here's a link to the article:
Often people are just flying RGB, and so the false-NDVI map is created simply by adding an index to the map. They only have R,G,and B to work with.
This article shows that since the dynamic ranges of a false-NDVI is so small (no matter what False-NDVI index you are using), that sometimes plants can show stress where they aren't. With a reflectance of less than 10%, a VERY accurate RGB camera is needed when doing analysis of just the RGB bands of light. If your sensor (like a typical one from a phantom) is off between pictures in the SLIGHTEST, then you will have this problem. Not true in CIR
Maybe that answers your question, maybe not ;)
Yeah...still not seeing it. I have no problem with the claim that CIR sensors give a larger dynamic range of the figure of merit (0-1 rather than 0-small) than RGB sensors, though I don't see how this article claims or proves that (apart from the quotes at the end). What I'm getting caught up on is that the image on this post shows the false NDVI image suggesting poor crop health even though the crop looks awfully green to me. If the false NDVI is constructed from both the image on the left AND a reference image, then that reference image needs to be included in the figure to have it make sense. If the false NDVI is constructed just from the image on the left (as would be suggested by only having the image on the left), then it doesn't make sense because the crop is clearly nice and green in the image on the left but the false NDVI shows poor crop health.
A plant absorbs about %90-95 of the visible light, whereas in the NIR bands it reflects most of the light. So when we look at a false-NDVI from an RGB sensor, it's hard to see the difference ls between healthy and stressed plants.
When creating a false-NDVI index, most of the time the values are around 0 because of very little light being reflected. Because of this, sometimes healthy plants show up even more stressed than a stressed plant.
Using a good CIR sensor, the NDVI has a much larger range, typically from 0 to 1. That is where we can differentiate with more accuracy a stressed plant and a healthy plant.
No matter what index/calculation you use, in the RGB with the little reflectance you will have doesn't compare to NDVI with NIR. You can't make up for the large absorbance.
Just my thoughts :)
Although I agree that the jury is still out on the utility of converted point and shoot cameras I think there is hope if appropriate filters are used, calibration routines are implemented and a host of imaging protocols are followed. I have been experimenting with methods to remove the NIR contamination from the visible channels and initial results are promising.
I also agree that in many cases RGB imagery is just fine for getting information about crops or plant health but it's hard to argue that using NDVI isn't advantageous over RGB indices for plant health/vigor simply due to the difference in dynamic range. Healthy green vegetation tends to reflect a little less than 10% of blue and red light, a little more than 10% of green, and usually over 70% of NIR. The difference might not always be significant just like 8-bit color doesn't always appear to be a big improvement over 24-bit color but for certain applications the difference is important.
I think the jury is still out as to whether these converted NIR cameras provide an improved ability to characterize and monitor vegetation. As Benjamin points out, the false NDVI image presented above could easily be inverted (or the formula changed) to generate the same scale direction. There are also other RGB-based indices that can be used such as Excess Greenness. While a clean NIR/red-based index may be theoretically preferable for vegetation, a key practical issue is how the replacement camera filter rejects visible wavelengths to generate the new RGB channels that contain IR radiation. For example some modified cameras are contaminated by NIR in all three output channels and therefore won’t provide good separation between the NIR and visible range. As an aside, the visible channels in satellite imagery (e.g. Landsat), especially the blue and green, are significantly contaminated by atmospheric scattering, while this isn’t an issue for near-surface UAV remote sensing.
We need more controlled experiments comparing the two types of cameras that are based on quantitative measurements of plant leaf area index, chlorophyll content, biomass, etc. Here are a couple of recent peer-reviewed comparisons that did not find any advantage in using converted NIR cameras…
I agree with Benjamin too, you have inverted color scheme, RGB gives important info for farmers too, depends what are you looking for, and NDVI isn't usefull in many cases; IR brings important info too, but you have to know what you are looking for.
If the NDVI & false NDVI maps are produced by some kind of difference from a baseline/reference image, then we're missing both an explanation and illustration of this. A convincing case could be:
* Here are the reference and test images in RGB and they produce this false NDVI. The reference and test images look similar, but careful pixel examination results in the differences you can see in the false NDVI.
* Here are the reference and test images from the CIR camera and they produce this NDVI. Again, the reference and test images look similar, but 1) they are actually much more different from each other than the RGB images above which gives us more dynamic range and [if applicable] 2) look: the RGB test image indicates that the crops are doing worse when actually the more accurate CIR test image shows that they're doing better.
The above case is only convincing if the two reference images were taken at the exact same time and the two test images were taken at the exact same time.
I tend to agree with Benjamin. I've looked pretty hard and haven't been able to identify super solid evidence that NDVI uncovers any information that RGB doesn't... If the field has a good reference-point for health (which is required to calibrate the NDVI image anyway), then everything else is relative to that reference point...and you can get the same thing from RGB.
Does anyone have very clear examples of where NDVI actually provides actionable information that couldn't be gleaned from an RGB image? From my perspective, the RGB image above is just as clear as the NDVI, it's just yellow instead of red on the bad parts of the field...
I previously had no opinion on this subject but now I'm very suspicious of the claim that CIR adds a lot of information. In the image shown on this post, the false NDVI has an inverted color scheme for showing crop stress -- if the algorithm to convert RGB to NDVI shows those clearly-green areas as red/bad on the false-NVDI, clearly it's your RGB-to-false-NDVI algorithm that needs to be fixed.
Likewise, in Fig 1 of the article, clearly those two small rectangular plots above the large square plot in the lower right are some of the most healthy crops -- they're bright green on the RGB, and the NVDI Fig 3 shows them as good. But somehow, they're spotty and red in Fig 4. What RGB to false NVDI algorithm are you using and how can it possibly produce these results?
The quotes at the end of the article seem to suggest that increased dynamic range in crop quality is the primary advantage of true NDVI. That seems totally plausible, but the explanation given here and in the article is not that.