What I want to do
I want to know which filter performs better to get NDVI images with a Canon A490 (ccd)
My attempt and results
I am comparing here the results obtained with this two Rosco filters when getting NDVI images from NIR photos. The camera used was a Canon A490 with the IR filter previously removed. Camera was white balanced with a blue origami paper before taking photos with each filter. The software to process NIR photos was Fiji and Photomonitoring plugin.
The ideal criteria would be to compare NDVI values with real biological plant data, but as this is not possible for me I would compare photo histograms, that is objective data, and then interpret NDVI image, a more subjective analaysis.
Firstly it is interesting to compare filter spectral response. #74 lets less green wavelengths pass through and a bit less energy from blue channel, also less NIR wavelengths (> 700nm), so we can expect greater red-blue channel differences with #2007. The total area of passing wavelengths seems greater for #2007 so it will generate more luminous photos.
This photo contains some plants and also some other objects that are not photosynthetic, I took this photo to evaluate filter ability to distinguish photosynthetic materia in an image.
The deduction about amount of light seems true as camera used fastest shutter time for #2007. One curious thing is that leaf shadows appears with higher ndvi level than leafs, specially with #74. And also with #74 the shadow that crosses all the circle is noticeable in the #74 NDVI image, but not in the #2007. So one conclusion from this is that #74 is more affected by shadows and could lead to errors when distinguising plants in one image.
The histogram of each photo is the following. The difference between the average values of red a blue channel is 28,4 for #74 and 26,2 for #2007, what is not what I expected, but this full image histogram is not very meaningful as more than 75% of the photo is non photosynthetic material and the amount of plant reflecting more NIR is small. Anyhow I am comparing average values for each channel, which is a very rough analysis.
This other image is from just some leafs, all the objects in the image are photosynthetic so here we can evaluate better the ability of each filter to differentiate ndvi levels.
The histogram of the image reveals now a R-B everage difference of 62,1 with #74 and 56,2 for #2007. Here the comparison is for a image with 95% of photosynthetic material. And this small difference in channel levels is noticeable in the NDVI image, as the different levels in the main leaf are more clear in the #73 than in the #2007.
My first conclusion is that #74 performs a bit better for my Canon A490 CCD, I will soon try the red filters.
Questions and next steps
Is comparing histogram channel average value effective to deduce NIR image quality for getting NDVI values?, what would be better than simple average?
Why shadows have more NDVI level than plant??
Why I'm interested
Getting aerial crop NDVI images
2 Comments
This is a great analysis. I like your interpretation of the filter spectral graphs. Rosco confirms that 2007 passes more light (10% transmission) compared to 74 (4% transmission). The difference we experience is not that great, maybe because the transmission in the NIR is more similar than it seems (their graphs stop at 750 nm but the camera is sensitive to 900 or 1000 nm). Also the extra green transmission of the 2007 might not influence the blue or red channels that we need for NDVI. But it's hard to know exactly what is happening because the camera's Bayer filters allow non-target colors to get into each channel, and it is hard to find good data about that for each camera.
Your results confirm that a common artifact of this process is high NDVI values for areas of dark shade. For some reason, shady areas have higher than expected values for NIR (or lower than expected values for visible light), so NDVI values are artificially high. It would be good to learn more about this, but until then avoiding shadows and avoiding interpreting NDVI results from shadows is called for.
The white balance you used seems to produce infrablue photos that are not as orangey (for foliage) as others that have produced good NDVI results. In this note I describe differences between performing the custom white balance in the shade versus direct sunlight. I subsequently settled on doing white balance for blue filters (with blue paper) on sunny days in the shade, and white balance for red filters (with red paper) in direct sunlight. Oranger infrablue photos have higher values in the red channel which is used for NIR, so NDVI values will be higher, and closer to the expected range (0.3 to 0.7). I think oranger infrablue photos also produce a greater range of NDVI values for foliage, so more subtle difference in plant health can be discerned. Similarly, bright bluish foliage in super-red (red filter) Infragram photos produces the best NDVI results.
Comparing the difference between the red and blue histogram averages is a good way to predict the nature of the NDVI results. The greater the difference, the higher the NDVI values will be. In Fiji, if you can highlight a very small area of typical foliage with only one tone, the histograms will be narrow distributions. That will make it easier to choose the value for comparing (choose the peak) and will make the average value a better estimate of the peak.
I look forward to your results with red filters.
Chris
Reply to this comment...
Log in to comment
Hi Chris, your supposition is right, my white balance was done at direct sunlinght, I tried today balancing in the shadow and the new images appears almost the same to me (the photo was taken 2h later, so different light) , I expected a more orange hue after reading your post "balancing act", but seeing the ndvi image it has slightly more higher ndvi values, the values over a part of the leaf without any shadow now are 0,35 where in the previous photos was about 0,25, so yes, the balancing in the shadow results in higher ndvi values.
( down photo is white balanced in the shadow, and the one above with direct sunlight)
Regarding the shadows I compared histograms for a shadow part of a leaf and there is a reduction of 43% in the red channel and 65% on the blue, that agrees with your explanation of having less light than NIR at the shadows and that results in greater difference and therefore ndvi value.
Thanks, I really appreciate the great labour at this site sharing knowledge.
Reply to this comment...
Log in to comment
Login to comment.