Good morning everyone
We have recently purchased a NGB converted camera (NIR, G and B channels) with the goal of observing different vegetation indices (DVI, NDVI, SAVI, etc) from UAV-captured images. We were hoping some of you may have some tips on calibrating the camera settings to optimise these images (NDVI in particular).
We have found that the NDVI (and DVI, as expected) images vary immensely in different light conditions and for different camera settings. This is far from the ideal normalised index, where the range of values should be environment independent. To put it simply, based on the literature we would expect the NDVI values to fall between 0.2 and 0.8 for vegetation, where the range represents stressed to healthy vegetation. However we are finding that the range of our values vary with different light conditions. While we could obviously manipulate a single image (e.g. it's colormap) to display a nice looking NDVI plot, the day-to-day variance does not allow for comparisons over time.
Below are two NDVI images of the same scene, one day apart with the same colormap, which illustrates the paragraph above.
We currently use a blue card, in the same light conditions as the image, for the white-balance, a shutter speed of 1/1000 and an ISO of 400.
We have analysed the individual NGB channels and noticed a few surprising results. While the NIR and G channels show expected values, the blue channel is producing quite different results to that in a RGB camera. The vegetation is showing up particularly bright in the NGB blue channel, opposed to the respective RGB channel. This would explain why there are problems with the NDVI values as we are using the blue channel to calculate the index. Perhaps calibrating the white balance differently would overcome this variance?
Thanks!
Benji
Replies
I will be experimenting with temporal resolution of ndvi's created from a NGB camera this summer. I believe that it is possible to compare images over time by being particular about the lighting conditions the images are taken in and I will experiment with collection at different altitudes to hopefully compensate for the mii scattering effect. I believe that the closer you are to the target (Vegetation in this case) the less variance you will see in the blue spectrum.
We are currently attempting this method. As Deon suggested it is obviously more computationally intensive however we do expect the results to drastically improve our NDVI images.
We are looking to use image-registration algorithms to produce spatially overlapping images for this purpose. The practicality of the NDVI image will be fully determined by the accuracy of the image-registration.
It is possible to use two cameras. We started with that method. However, it is a challenge to obtain precise synchronization between the cameras, it makes data processing more difficult, and it is not a complete solution in terms of band selectivity because you still deal with a wide red band. The disadvantages outweighed the advantages in our applications, but it is not a bad idea if you can overcome the technical issues.
Benji,
You are seeing some of the complexities associated with using converted "NDVI" cameras. It does not make the data useless, but you have to take the limitations into consideration. One of the important factors to realize is that the data obtained by a converted camera are not the same as data from a traditional NIR/red narrow band system designed for NDVI. You have a camera with wide bands that overlap substantially. The "green" channel responds to light in the blue and red bands, but with peak sensitivity in the green region. What is actually more problematic, however, is that there may be significant response in the blue sensor to NIR light, particularly if the conversion does not include a filter to remove the higher wavelengths of NIR light above 850 nm or thereabouts. The blue sensor response in the NIR region produces noise in the NDVI that reduces the signal to noise ratio and the sensitivity to small changes in vegetation biomass and vigor.
Calibration based on an invariant target is typically poor because it is incredibly hard to produce a true Lambertian surface. The ratio between bands of incoming light is also not constant between days or time of day. A true calibration therefore requires data about the individual bands sensed by the camera. Just point your camera upwards (use a diffuser) at different times, and calculate the NDVI of the incoming light, to see what I mean.
Bidirectional reflectance is another issue that you need to deal with using appropriate data modeling. And then you need to consider the influences of shadows.
Finally, but not least, are issues related to the use of light by the plants. Photosynthesis is not a 100% efficient, and is not a first order process across widely changing light intensities. It leads to an apparently higher NDVI at low light conditions because plants are able to utilize a greater fraction of the available light compared to bright conditions.
For these reasons and others, it is not possible to obtain absolute NDVI values using a converted consumer camera, but again, it does not make the data useless. It just means that you have to be very careful in data processing and interpretation, and understand the possible causes of variations in data. Practically, it often means that you should use a relative index to interpret the data, and you should be extremely wary when trying to compare data obtained under different atmospheric conditions, particularly when cloudy vs clear conditions are compared.
Thanks for your detailed response, Deon