Late last year we released our first version NDVI converted cameras using Schott BG3 filters. Using those filters we were able to generate pseudo NDVI images with a single camera, building off the work of the Public Lab project. After working with faculty at the University of Boston and the MEASA Lab at KSU, we determined that we weren’t quite getting the best results possible with the BG3 filter glass. Unfortunately, no off the shelf filter glass seemed to do exactly what we needed, block out red light and allow NIR light to pass instead. Luckily, we were able to find a manufacturer to help us build a custom filter to our specifications, and the results are looking really good! Below are two images, the first taken by a camera with BG3 filter glass and the second taken by a camera with the new custom glass, the difference is clear.
BG3 Filter Glass
Custom Filter Glass
Aside from higher overall values, there is more differentiation and detail in the leaves and a larger difference compared to the non-organic background material. This is especially important for this type of camera. Since these are uncalibrated, pseudo NDVI images, what is really needed most is high differentiation so that comparisons can be made between the majority of plants and potential problem areas. The higher detail available now makes it possible to catch unhealthy plants sooner.
Custom Filter Transmittance vs Wavelength
The new filters will be priced at $39.99 for the filter alone (8.9×7.9mm, fits SX260 and S100 at least, potentially others) and$499.90 for a ready to go converted SX260 with CHDK. It’s a little cold to be doing any crop surveys here in the Northeastern US right now, I actually ended up killing those two plants after taking them outside for just a few seconds! We’ll be out and collecting imagery as soon as anything starts growing and I’ll post updates as they become available.
Comments
@Jeff - I will definitely post my results!
@Vega - Curious, what would be a ballpark price of your 5 channel prism camera?
We have actually done some test processing using Pix4D and we can process around 300 Acres at 1" GSD in under 2 hours time. It takes under 30 minutes to collect the sensor data. The output seems very promising. As I said, we'll be doing the ground truthing during the growing season and will post some results once we have more data and find the time :)
We have been analyzing some Hyper Spectral data that was flown last year and trying to find the value... we haven't found any plastic plants in the field yet ;)
Should be interesting to see the data from the Sony. We are using a custom built 5 Channel Prism Camera. The camera uses Fairchild's Monochrome sCMOS imaging sensor. The filters are at the output face of each channel, so there is a dedicated channel for Blue, Green, Red, plus two NIR channels, covering 680-780nm & 800-1000nm.
The filters can be changed to allow for any agricultural application. By using dedicated channels and monochrome imaging sensors, all of the pixels on every sensor is used in each channel. For example if each imaging sensor is 2MP, every channel is 2MP for a combined resolution of 10mp. Speed is also part of the equation running at full frame and 100fps+ in Global Shutter mode.
All image processing is done on board the UAV with an embedded computer. Post processing using Pix4D is slow and data is inaccurate.
Since the sCMOS sensor is very low light and high dynamic range, the UAV can fly and collect data in cloudy, early morning and late evening lighting conditions, as well fly in to high light areas, such as direct in to the Sun, without saturating the sensor.
We are looking at a new Hybrid Hyper-Multi Spectral camera that will combine a 128 or 660 band Hyper Spectral imager with the above mentioned 5 Channel prism camera. This will easily create 3D images and gather data on specific plant species, even determine if they are fake or plastic plants.
A Foliage Penetration(FOPEN) project is also in the works, combining a LiDar system with the Hybrid Hyper-Multi Spectral Camera.
Systems can also image from 380 to 1700nm using a combination of Visible sensors and InGaAs sensors.
Please do post your results as the season progresses!
Jeff/Vega,
Some great dialogue going on here about sensors.
I am in no way a sensor expert like you two, but I have a Sony NEX-5N and I converted it myself with the filters from the Public Labs project that Jeff mentioned and I'm getting good results. We'll know more as the growing season progresses and we "ground truth" the sensor data.
Just wanted to let those know who were asking about the Sony NEX cameras that it is possible.
I will try and post some stuff in the next month or so as we collect and analyze the data. Busy time of the year!
Jeff is correct. The Bayer pattern on almost all color sensors(except Foveon, www.foveon.com), make it difficult to interpolate very specific wave lengths. Remember a Bayer pattern is a group of 2-Green, 1-Red, & 1-Blue pixels.
In cheaper imaging sensors, such as those used in consumer cameras like the Canon camera, the dye sublimation process used is not nearly as accurate nor of great spectral response.
Doing this type of specific wavelength data collection, it is always best to use a monochrome imaging sensor. This is one without a Bayer pattern.
Monochrome cameras are widely available in machine & robotics vision. Visit AIA(www.visiononline.org), EMVA(www.emva.org), and SPIE(www.spie.org) for more in depth information.
Hi Gonzalo, I haven't worked with a yellow filter yet. There are no Bayer sensitivity graphs available for consumer cameras like the S110 but if it's like most others, the red pixels will read a lot of infrared light as well. This will make it hard to separate the infrared intensity from the red intensity. That's why I haven't experimented with it. The filters we use give a pretty good approximation, to improve on that I think you really need to move to a higher end solution like Micasense or Tetracam Mini MCA (not ADC or ADC Lite, these are just cameras using yellow filters as far as I can tell).
Hi Jeff/ Vega, thank you for your explanation. Jeff, another question, have you worked with some yellow filter in converted Canon S110 so that could filter bellow 460-470 nm and be able to allow red and infrared waves to make ndvi graphics?. Thank u again. Gonzalo
Jeff, the cost is cheap, but the results are poor. Are the transmission results of this filter with or without a lens? If this is just the filter, a lens designed for 380-680nm will drop off quickly outside this range. Vignetting, warping in the image, will cause distortion, as well poor MTF will result in unstable and inaccurate data collection. High F# lens should be used, below F#2, preferably close to F#1 as possible for lower light operation, on those cloudy days.
Field Curvature and Distortion as also important in any camera lens.
Hi Gonzalo, thanks for your question. You're exactly right, what we're doing with our filters is making an approximation of NDVI where instead of using the red band as the visible light channel to compare against, we use the blue band. It is an approximation of NDVI but this is basically what we're after, and it's what allows us to get this information at such a low cost and with low complexity.
To see a post-processed sample from this filter and camera, check out the last map on this sample maps page.
Gonzalo, for some attributes, the dual bandpass filter shown in the graph cuts out the red, shifts transmittance to 680nm and above, dropping out around 900nm. Unfortunately, transmittance at 900nm using the lenses that come with the camera is poor. Spectral response is horrible in this imaging sensor, with QE less than 10% at 850nm.
Here are some of the Spectral Areas for Agriculture:
NDVI Normalized Difference - Vegetation Index - (RNIR-Rred)/(RNIR+Rred)
RVI Ratio - Vegetation Index - RNIR/Rred
WDVI Weighted Difference - Vegetation Index - RNIR-C*Rred C = 2 (soil factor)
REP-LI Red edge position: linear interpolation method 700+40(Rre-R700)/(R740-R700) Rre: (R670+R780)/2
MTCI MERIS - Terrestrial Chlorophyll Index - (R754-R708)/(R708-R680)
TCARI Transformed chlorophyll absorption in reflectance index 3((R700-R670)-0.2(R700-R550)(R700/R670))
TCARI/OSAVI - Combined Index: TCARI with Optimized Soil-Adjusted Vegetation Index
TCARI/OSAVI OSAVI: 1.16x(R800-R670)/(R800+R670+0.16)
MCARI Modified Chlorophyll Absorption index [(R700-R670)-0.2x(R700-R550)x(R700/R670)]
DCNI Double-peak canopy nitrogen index (R720-R700)/(R700-R670)/(R720-R670+0.03)
NDRE Normalized Difference Red Edge Index (R780-R720)/(R780+R720)
-
1
-
2
-
3
-
4
-
5
of 5 Next