Hyperspectral image analysis of plants
The motivation of this work is to try to determine if it is possible to detect the onset of plant
diseases before they become visible to the naked eye. To do this, we can capture very high fidelity
colour information in images, which also includes areas of the spectrum outside the range of human
vision. Given this data, can we interrogate it to identify indicative regions of the spectrum which can
identify disease onset before it becomes visible?
The colour of leaves is produced by a variety of cellular and subcellular structures absorbing and
reflecting light in differing quantities. For example, the generally green pigment of leaves is
produced by the chlorophyll inside chloroplasts, which absorbs mainly in the blue and red regions of
the spectrum (and hence reflects back mostly green light).
We are able to capture images of plants using a hyperspectral camera. This device works much like a
regular camera, but instead of capturing just the red, green and blue components of colour, it is able
to record a complete breakdown of the reflectance spectra for each pixel in the image:
UV
Blue
Green
Red
Near infrared
IR
Figure 1. Example spectra plotted for different plants imaged using the hyperspectral camera. The x-axis represents light
wavelength, from UV on the left to near infra-red on the right.
Using such a device, we are looking to identify characteristic changes in these spectra which could be
indicative of the onset of disease. Traditional ‘multispectral’ indices (meaning they do not have a
complete spectrum available, rather just select samples from the spectrum) include such metrics as
NDVI (Normalised Difference Vegetative Index), which is widely used to determine the general
‘healthiness’ of a plant:
where NIR is typically measured around 800 nm and VIS around 680nm. As you can see, this is
calculated using just two representative wavelength samples (Near infrared, and visible). For an
example see here
http://earthobservatory.nasa.gov/Features/MeasuringVegetation/measuring_vegetation_2.php
This type of index was derived from satellite imagery, where only select wavelengths, and a very low
spatial resolution was available. Our question is:
Given we have a full spectral response for a plant, and a high spatial resolution (we can see these
spectra as they originate from different leaves, or even different spectra within a leaf area), how
can we go about identifying changes in regions of these spectra over time which could indicate the
onset of disease?
Could we identify a new index from an objective data mining approach which could be used as an
indicator to identify the onset of a particular disease?
References:
A selection of existing indices and their background reasoning are available here. Each is used for a
particular purpose.
https://www.exelisvis.com/Learn/WhitepapersDetail/TabId/802/ArtMID/2627/ArticleID/13742/Veg
etation-Analysis-Using-Vegetation-Indices-in-ENVI.aspx
Peñuelas, J., Baret, F., & Filella, I. (1995). Semi-empirical indices to assess carotenoids/chlorophyll a
ratio from leaf spectral reflectance. Photosynthetica, 31, 221-230.
Peñuelas, J., & Filella, I. (1998). Visible and near-infrared reflectance techniques for diagnosing plant
physiological status. Trends in plant science, 3(4), 151-156.
http://qubitphenomics.com/hyperspectral-imaging/ (and refs therein)
Some example existing indices, drawn from the references above:
INDEX
Normalized Difference Vegetation Index
Simple Ratio Index
Greenness index
Zarco-Tejada & Miller Index
Simple Ratio Pigment Index
Normalized Phaeophytinization Index
Photochemical Reflectance Index
Normalised Pigment Chlorophyll Ratio Index
Carter Index 1
Carter Index 2
Lichtenthaler Index 1
Lichtenthaler Index 2
Structure Insensitive Pigment Index
Gitelson and Merzlyak Index 1
Gitelson and Merzlyak Index 2
Derivation
NDVI (r800 – r680)/(r800+r680)
SR
R800/r680
GR
r554/r677
ZMI 750/710
SRPI r430/r680
NPQI (r415 - r435)/(r415 + r435)
PRI
(r531 - r570) / (r531 + r570)
NPCI (r680 - r430) / (r680 + r430)
Ctr1 r695/r420
Ctr2 r695/r760
Lic1 (r790 - r680) / (r790 + r680)
Lic2 r440/r690
SIPI
(r790 - r450) / (r790 +r650)
GM1 r750/r550
GM2 r750/r700
So the question is: How can we improve on these simple ratios with our higher-resolution and
continuous (rather than discrete) spectral measures?
Data:
Data is extracted from images, so we have several colour measurements per leaf (See example
image below for an idea of resolution available)
Typical visible disease image:
Most foliar diseases are clearly visible on leaves when mature. How can we detect them earlier?
There have been many papers detecting disease in colour images after it is visible. We are interested
in whether it can be detected before it is visible (to a human observer).