Hi Folks, I modified a Canon Powershot SD850IS using the red filter from the Public Lab kit. I am using the camera to take pictures of melon plants, with the goal of measuring relative NDVI between plants. All images are taken at essentially the same time (within a 1-hour timespan) and all plants are in the same field.
I am analyzing the images in Fuji using Ned's plugins. However, I am not getting the results I expected. Specifically, much of the plant foliage is showing up as gray, even though it is visually green and healthy. This is worse when I use a grey card to set the white balance on the camera. Here are some pictures so you can see what I am getting.
This is an original image, taken without setting the white balance.
Color_Index.tif This is the same image after processing. I used the default settings except for setting the red channel to red, and the IR channel to blue because I have a red filter. As you can see, the foliage in full sun is registering as not photosynthetically active.
Here's the NDVI image for the same photo.
THis is the same image, analyzed with the plug-in in DVI mode, rather than NDVI mode.
Color_Index_bands_reversed.tif Here is the same image, processed using NDVI with the IR in the red, and the red in the blue (still with the red filter)
Based on these results, I am thinking that the problem is either that the NIR levels are so high in full sun that the camera sensors are over-saturating, or that I am getting a lot of NIR in the visual channel. However, I am wondering what those of you with more experience think. Also, what should I do about it?
Other questions: My plastic photo grey card shows up yellow in the NDVI images, suggesting that it is reflecting quite a bit of NIR. Suggestions for a better way to control the white balance on my modded camera?
Ultimately, my goal is to generate a single value for each image that summarizes the plant health, so I can run statistics to compare between treatments. Anyone have suggestions for the best way to do this in Fuji? I have used SigmaScan for similar analyses in the past, where I selected pixels based on color, and then counted them. Is it possible to select by color in Fuji? Even more important, can this process be scripted?
THanks in advance for your help. Rebecca
3 Comments
Hi Rebecca,
Your NDVI values are a little low, but that is to be expected when the procedure is not calibrated. To get the NDVI values closer to the proper range, you can set a custom white balance on the camera using a bright red surface as described here. It's hit or miss finding a good red surface because you need one with the proper relationship between reflectance of red and reflectance of NIR.
The red and blue channels in your photo are not over or under exposed, so that is not the problem. The blue channel (NIR) is a little bright, but still has good information in it. Exposure is an issue to remain vigilant about, because a dark photo will produce different NDVI values than a bright photo.
If your goal is to compare different melon plants, it doesn't really matter what the actual NDVI values are, only whether there is a difference among plants. So getting the NDVI values into the proper range might not be so important.
A bigger issue might be distinguishing subtle differences among melon plant health. If some plants are just a little stressed or nutrient deficient, their NDVI values might differ by only 0.1 (or less) from healthy plants. The color look up tables typically used for NDVI assign pretty much the same color to the values 0.4 and 0.5, so if want a good visualization of the study you might need a custom look up table with many distinct colors for NDVI values between 0.3 and 0.6.
Your NDVI images show a range of NDVI values within individual leaves because the leaves are not flat and sun incident angle modifies the reflectance. Darker areas are producing higher NDVI values. This is probably because darker areas have lower values for (NIR+VIS) but the difference between NIR and VIS is still large, so (NIR-VIS)/(NIR+VIS) gets bigger. You will have to decide how to sample to average all leaf surfaces or avoid some subset of darker or brighter surfaces.
Fiji has good tools for sampling and it computes the mean and variance for a sample (example here). It will be difficult to automate this process because somebody has to draw a polygon around the leaf surfaces you want to sample. Taking the photos so they include just the area you want to sample could work if each photo has the same proportion of leaf surfaces at every angle. This could allow computing mean NDVI for the entire photo, but there also has to be the same proportion of soil, stems, fruits, flowers, etc. in each photo. And no shadows of the camera, etc.
So the variability within a leaf of a healthy plant is probably going to be much greater than the difference between that plant and a stressed plant. Your sampling protocol will have to be both rigorous and extensive to demonstrate that difference.
Chris
Reply to this comment...
Log in to comment
Hi Chris, Thanks! I hadn't thought about using a red card to set the white balance, but it makes sense. I understand what you are saying about leaf-to-leaf variation within plants being large. I expected that, and am working with ~ 240 replications in order to compensate. I am planning to use the whole image - the plants are now big enough that they fill the image area. Rebecca
Reply to this comment...
Log in to comment
An update on the project. I tried using a piece of red tissue paper as a reference for setting the white balance, and it really helped. Here's the picture:
I think it will work for my needs since I can use the same tissue paper for all the images, but long-term something more reproducible would be good... too many hues of red in the world :-) Rebecca
Reply to this comment...
Log in to comment
Login to comment.