Public Lab Research note

NDVI two camera system calibration

by gpenzo | May 26, 2015 21:25 26 May 21:25 | #11840 | #11840

What I want to do

I have a two camera setup to capture NDVI images. Camera works fine. But I have some issues with calculating NDVI.

My attempt and results

The current way I calculate is to take pictures on both camera and keep increasing the exposure till I start having over exposure of pixels. Then I go one exposure step back. This way I make sure I have the biggest spread in both my pictures from 0 to 255. Also this way I do not need to do any streching of my channels. Then these pictures I use with the sandaard NDVI formula. The NDVI pictures look good. But I'm wondering if this is the correct way of doing it. Should there be some kind of calibration? Should the exposure time relation between NIR and RED be constant? The NDVI formula does not give any restrictions on this. If I do take pictures of the same piece of grass during diffrent time of day will I get the same result? Or will the conclusion (health) of the pice of grass be diffrent. I saw on other notes that the transmission curve of the camera is important.

wave_length_mono_response.png I do not understand yet why.

Questions and next steps

So much questions, or am I making things more difficult than it really is? Can any of the NDVI guru, lords, masters here help me with some of these questions.

Why I'm interested

I spend allot of time making a 2 camera setup. Getting RED and NIR was the easy prart. But I need to know the correct way of calculating NDVI images. Also I'm preparinge to make a 4-5 camera setup to capture other indices and I need to understand NDVI first before I start on the next camera setup.

Many thanks for any help you can give me. Regards Grayson Penzo

Note picure is of a farm land where nothing is growing except a nice green health patch of grass on the left.


If you want your DIY NDVI to be comparable to legacy (e.g., satellite) NDVI, there are a few things to consider. First, consumer cameras might not capture the same bands of visible and NIR wavelengths as the standard purpose-built orbital or high altitude sensors.
This might not be too critical because consumer cameras might be close to some official sensors, and satellite sensors have changed over the years. Some NDVI from satellite sensors might even be adjusted so it is comparable to older (or newer) values, and this could also be done with DIY NDVI.
Second, legacy NDVI is computed from radiance values, or how much energy is returned from each part of the scene in each wavelength band. Consumer cameras do not record radiance, they just record brightness, and then the camera algorithms modify that information in various ways. There is a relationship between brightness in a photo and radiance, but you have to know that relationship to convert. Ned Horning has been experimenting with this conversion. The radiance of NIR from healthy plants is typically 2 to 9 times greater than the radiance of red. So to get NDVI values in the proper range, the values of NIR and RED used to compute it must have that relationship. If you figure out how to translate brightness to radiance, your NDVI values should be good. If you just use brightness, you probably have to use brute force to push your values into the range that produces meaningful NDVI. This can be done systematically based on calibration targets, so the results could be meaningful. Without calibration targets, the results will be somewhat subjective.

Third, the Bayer filters in consumer cameras are designed to allow only one color to be captured by each pixel. But these filters are not perfect and often leak other colors (including NIR). So it is a challenge to produce an image (or channel) that has only the wavelengths you want. Investing the effort to translate into radiance values may not lead to good NDVI if your RED channel is inflated with NIR, or your NIR channel is inflated with UV. Making corrections for these leakages is possible if you can figure out how much leakage there is. Ned has been working on this too.

NDVI values do not have to be comparable to legacy NDVI to be useful. If you optimize and standardize your procedure you should be able to compare two fields photographed on different days in similar lighting. You just can't compare them very closely, so subtle differences in plant health cannot be discerned convincingly.

Reply to this comment...

@cfastie, thanks for this write-up on some of the issues that plague the cheaper approaches to VI. I'm not sure people are aware of this and this might merit it's own page of sorts (pooling the data from Ned etc.).

I was actually struggling with similar issues this weekend. I'm working on some synthetic images to model the camera response based upon hyperspectral imagery to see how different weather conditions and solar angles influence the data stream / image output. More on this will follow and hopefully will give me some insight in the sensitivity of cameras.

Reply to this comment...

Chris's post and graph are terrific. This came up again in discussion with #aren project members and I finally found this graph again! Did it ever get posted on a wiki page somewhere? Thanks!

Is this a question? Click here to post it to the Questions page.

Reply to this comment...

That figure has been in a wiki since April 2012:

Reply to this comment...

Login to comment.