A modified DSLR does not make a good multispectral camera
The first question asked by anyone new to these sensors is “Why do they cost so much?”. After a while they just get used to it, but the question still deserves an answer. We will also explore how suitable consumer cameras are for taking multispectral images.
Most (not all) multispectral imagers use the same CCD’s and CMOS sensors as consumer cameras do, and that is really what you pay for in consumer camera. And it is true that in terms of sensor quality for your buck, multispectral systems are a rip-off. An entry-level camera with APS-C -sized sensor can be had for roughly £600, and will get you ~400mm^2 of CCD sensor real estate.
By contrast a 6-channel multispectral could set you back £7,000, and typically will have very small sensors — ½ or 1/1.8 inch. That gives you roughly ~200mm^2 of sensor real estate — a 20 times difference in terms of bang for you buck!
TL;DR
The multispectral MCa cameras are still overpriced for what they are, the prices will come down but they won’t cost as little as consumer cameras do, simply because of economy of scales. Common colour cameras can be used to produce NDVI and find damaged crop cheaply, but as your requirements grow you will need capability they do not offer. It is physically possible to get ‘proper’ multispectral images out of them, but it’s quite challenging and the fact that you have to do that highlight a the fact that market for this sensors is totally immature.
A proper multispectral imager is quite different to a consumer camera in it’s design. Colour cameras are designed to create a visually pleasing image, it has no regard for accuracy and is not intended to be a measurement tool. Camera sensors respond to a broader range of light than a human eye does, starting in near UV at roughly 300 nm. and extending all the way to near-infrared. We perceive these wavelengths as colours. An image from a monochrome camera will be a combination of all this light. Colour cameras employ Bayer filter, which separates blue, red and green light, but usually does not deal with IR and UV light. A special filter, called Hot mirror, blocks UV and IR light from reaching the sensor. In many camera models it can be removed, granting partial sensitivity to these wavelengths.
A multispectral camera will respond to specific, and usually narrow bands. Instead of corresponding to colours, each band will usually correspond to wavelengths where chlorophyll, water or other substance of interest can be observed best. In their design multispectral sensors have more in common with satellite imaging than they do with consumer cameras. Below is selection of bands on the Sentinel 2 satellite, launched in June 2015 (bands 10, 11 and 12 are excluded, response is relative).
Calibration
Calibration is crucial in remote sensing, through it we relate arbitrary images to the real world, and make conclusions. Spectral response of an imaging system is a combination of that of image sensors, filters and lenses. Data usually needs to be compared to that acquired by different equipment. Even if you are using two imagers of the model, their response will differ and change as equipment ages. Calibration is suppose to deal with that, and is performed regularly through the instrument’s life.
Typically a multispectral imager will come with it’s bands balanced relative to one another. The manufacturer will provide guidance on calibration, or sometimes the sensor might calibrate on the fly and produce corrected images right away. Colour cameras have no functionality to help you find calibration coefficients or apply them to images they produce. This means that you would have to do it yourself in post-processing.
The bands in consumer cameras are inconsistent
To mimic human vision the CDD/CMOS is overlaid with a Bayer filter. As shown below, a green pixel would respond to light between 480 and 600nm. This range of wavelengths, and shape of the curve vary substantially between camera models and still produce a realistically looking image. Because the shape of the curve varies, no matter the calibration it is not possible to produce precisely equivalent image from two different models. Bayes filter cannot be removed without damaging the sensor. The only way of controlling the wavelengths you are imaging and producing a calibrated result is to place a desired filter over the sensor, ‘average’ RGB values and treat it as a monochrome camera. That way different cameras can be calibrated, and their images compared like-for-like.
Reduced light response
If a consumer camera is used with a dedicated filter, the Bayer filter will block roughly 2/3 of the light, depending on the wavelengths involved. A monochrome camera of similar sensor size and quality would have superior light performance and shutter speed. It also means that 2/3 of the CCD you paid for and 2/3 of the lens you are carrying around are going to waste.
True Resolution
A 16 megapixel camera would have 8 million green pixels, 4 million blue pixels, and 4 million red pixels. For every spot on the image, the camera only knows one out of three colours. The raw (hence the format, RAW) image looks like this.
For every pixel, the missing colours are estimated using information form the nearby pixels. This is fine for producing photographs, but it will not produce an actual measurement at that location.
Multispectral imagers work differently. Most common one, multi-camera-array (MCA), will have a camera dedicated to each channel. It will produce a green image, a red image, an infrared image, ets. These images are combined to produce one where each pixel has a real reading for every colour/channel.
If the goal is for each pixel to provide a true reading at that location, the image must be sub-sampled by a factor of four. That has serious implications for resolution, as 20MP camera suddenly becomes a 5MP camera.
Nir-enhanced sensors
There are some indexes that require a special infrared-enhanced CCD.
One of them is a water based index, which requires a measurement to be taken at 970 nm, the very limit of what common CCD sensors are capable of.
The following is a comparison of NIR-enhanced and ‘ordinary’ ccd sensor.
Productivity and Integration
A good multispectral system should produce calibrated, Geo-tagged images ready for further processing. It will trigger itself and operate reliably, images will be well — registered and aligned, and there will be no switching round of memory cards.
Unfortunately the market for small multispectral cameras is still in it’s infancy, and charges extraordinary prices for equipment that‘s only moderately impressive, and not very high resolution. All of this means that a lot of people are have to spend time tinkering with their own equipment instead of getting on with the work they want to be doing.
