Introduction
It is the investigative nature of Satellite Imagery Analytics that, I suppose, draws enthusiasts to this field - the possibility to extract meaningful insights by processing remotely-sensed data acquired by instruments hundreds of kilometers above the Earth's surface.
In this post, you will get to know about a unique technique of acquiring Earth Observation data called Imaging Spectroscopy known in common parlance as Hyperspectral Imaging and a video demonstration analyzing it in a sample workflow. At the outset, I will cover some Remote Sensing fundamentals and characteristics of the more commonly-used Multispectral and Radar techniques which will lend itself well to your understanding of where the utility of Hyperspectral Imaging lies.
HYPERLINKED SECTIONS:
Much thanks to EO College's MOOC series and the webinars by NASA ARSET which helped me to get acquainted to the world of Hyperspectral Remote Sensing.
Remote Sensing Background
Objects, materials and surfaces on earth can be imaged using two ways-
a) Passively - by only capturing reflected radiation typically originating from the Sun
b) Actively - by transmitting radiation through a sensor and capturing its reflectance
What you are seeing in Figure 1 is a natural-colored visualization of a passively-acquired Optical Imagery dataset - the dataset is formed when the imaging instrument captures reflected Solar radiation across one or more Spectral 'bands' i.e. ranges of wavelength. (Sunlight constitutes waves from the visible light, infrared & ultraviolet portions of the EM spectrum). If the imaging instrument captures reflected solar radiation across more than three Spectral bands (see the types here), then the dataset can be termed as Multispectral Imagery.
Upon interacting with an object on the surface of the earth, Solar radiation does not behave in a singular manner. Rather, some wavelengths within the radiation may be absorbed by the object, while the others may be reflected, and to varying extents.
Besides the properties of the transmitted wave, the extent of energy reflectance is influenced by-
Reflective properties of the object (whether the surface is rough or smooth),
Geometric effect of reflection (angle of illumination and angle of reflection), and
Biogeochemical characteristics of the object (moisture content, mineral properties, size etc.)
That being said, even before an energy wave interacts with an object, it has to pass through the atmosphere where it may be influenced by clouds and gases (the wavelengths that constitute Solar radiation are particularly susceptible to these). The same is applicable after interacting with the object when the reflected wave exits the atmosphere (provided the imaging instrument is spaceborne).
As can be gathered from the example in Figure 3, the rate of reflection varies considerably when soil, vegetation and water are exposed to the various wavelengths that constitute solar radiation. In the Near Infrared portion (band 4) of the spectrum for example, water has zero reflection whereas soil hovers around the 30-35% mark and vegetation has the highest reflection of them all touching 50%. This is in contrast to how the three materials respond to any of the other wavelength ranges (band 1, 2 & 3 i.e. RGB within the Visible Light portion or band 5 & 7 in the Intermediate Infrared portion of the electromagnetic spectrum). As a result, Multispectral imagery is commonly utilized to exploit such reflection variabilities in order to detect, delineate or classify distinct geologic features such as Farmlands, Lakes, Forests etc.
Because the source of illumination in Multispectral imaging is passive - Sunlight - it is an impediment at times - quite obvious that the imagery cannot be acquired during night-time. Even during the day, Solar radiation is adversely impacted by clouds, aerosols and other atmospheric features. Also, unlike what you see in Figure 3, some surfaces do have low variability in reflectance to the different energy wavelengths that constitute Solar radiation, thereby rendering the use of Multispectral imagery in detection or delineation workflows involving these to be ineffective.
Also, the Spectral resolution, which is the width of a Spectral band (Spectral band being the non-contiguous range of wavelength captured) - is an important consideration for finer classification workflows and this is where Multispectral Imagery falls short too. For example, while it is adept at distinguishing a Farmland from a Water body and is thus suitable to be used in related workflows, a researcher may not necessarily be able to distinguish coconut trees from palm trees in a plantation using Multispectral imagery alone due to its inherently low Spectral resolution i.e. data being acquired across wide ranges of wavelength which only serves to normalise the reflectance sensitivity.
Besides Spectral resolution, there are other types of resolution as well which are important considerations in their own way for any research study that incorporates Remote Sensing.
Here's a short summary-
a) Spectral Resolution - range of energy wavelengths which the sensor is sensitive to - for example, Cartosat-1 satellite captures reflectances in between 500 nm and 850 nm i.e. it has a single spectral band with a width of 350 nanometres
b) Spatial Resolution - measure of the smallest surface feature which the sensor can detect i.e. it is a dimension of the pixel size - for example, the dimensions of a pixel in RGB band of Sentinel-2 satellite constellation is 10 x 10 metres, hence the spatial resolution of this band is 10 metres
c) Temporal Resolution - how quickly the sensor can image the same object again i.e. the orbital cycle of the satellite - for example, Sentinel-5P satellite has an orbital cycle of 16 days which means that the same spot on the earth's surface is imaged by the satellite once every sixteen days
d) Radiometric Resolution - the quantity of energy signal information a sensor can perceive reliably i.e. with limited noise, over and over again - for example, the Radiometric resolution of OLI-2 instrument onboard Landsat 9 satellite is 14 bits sensitive which means that the sensor has 2^14 i.e. 16384 potential digital values - from 0 to 16383 - to store information in each pixel
Synthetic Aperture Radar (SAR), or simply, Radar Satellite Imagery often serves as a useful alternative in workflows where Multispectral imagery is unable to extract the desired output. It is also utilized in a complementary as well as supplementary way to validate the output derived using Multispectral Imagery.
Radar satellites acquire data through an active sensor which transmits microwaves on earth and captures its reflectance (technically known as backscatter as the captured rays have been reflected in the same direction as its source). For example, Sentinel-1 satellite transmits pulses of C-band microwaves (1 nm - 180 nm).
One major advantage of Radar satellites over Multispectral satellites (and even Hyperspectral satellites for that matter) is that, due to the presence of an active sensor which transmits illumination, radar satellites can acquire images during day or night.
[Please note, however, that while the mode of illumination for multispectral or hyperspectral earth observation satellites is passive i.e. solar radiation, acquisitions made using aerial or on-ground techniques do make use of active sensors, as do satellites purposed for outer space studies.]
Besides, microwaves are unaffected by cloud cover and aerosols by virtue of having a longer wavelength, which is also a very useful characteristic when one seeks to obtain reflectance information devoid of atmospheric influence. Moreover, materials and surfaces interact with microwaves in a different way than they do with solar radiation - a characteristic that is exploited in a standalone, supplementary or in a complementary way as I had indicated before.
A technical comparison table of the characteristics, advantages and disadvantages of multispectral and radar remote sensing can be accessed here and some of the applications involving either or both of them can be accessed here.
'So if Multispectral and Radar Remote Sensing techniques are so useful and cater to a wide variety of often-mutually-exclusive applications, then what is Hyperspectral Imaging and where does its utility lie?'.
The next section will be devoted to addressing this exact query which many of you may have.
Hyperspectral Imaging
As described in the introductory course on this topic by EO College-
"Imaging spectroscopy refers to imaging sensors measuring the spectrum of solar radiation reflected by Earth surface materials in many contiguous wavebands, on the ground as well as air or spaceborne.../...there are up to hundreds of reflectance bands that allow detection and quantification of materials based on the shape of the spectral curve. "
Contrary to what you may have perceived, Imaging Spectroscopy i.e. Hyperspectral Imaging is not a new technology. In fact, the first Imaging Spectrometer (device that acquires Hyperspectral images) became operational as far back as 1982. However, these were installed in research flights (airborne) to capture footage over relatively small areas at select locations and I suppose only a handful of researchers would have access to the data and the ability to process it. Only after spectrometers were installed onboard Satellites in the 2000s that the technology and its usage became mainstream.
As with any upcoming technology, the later versions iron out the flaws that accompany the original as well as benefit from the growth of the overall ecosystem, and so was the case with Imaging Spectroscopy - since 2019, multiple spaceborne Hyperspectral sensors have been launched and newer algorithms have been developed which promise to usher in a new era in Earth Observation to serve mankind's quest to have a deeper understanding of the geochemical, biochemical and biophysical properties of our planet's surface and atmosphere.
How is Hyperspectral imagery different from Multispectral imagery?
Have a look at the depiction below-
I had already delved into what spectral resolution is, earlier on in this post. Hyperspectral images collect information in narrow spectral bands (i.e. high spectral resolution) across a single continuous spectrum of wavelength. This is in sharp contrast to Multispectral images which capture information in wide spectral bands (i.e. comparatively much lower spectral resolution) across fewer, non-contiguous ranges of wavelength.
See the two charts in Figure 5 above which beautifully elucidate the difference - while Multispectral imagery contains information which you can draw a Bar chart with (categorical data - reflection rate per non-contiguous band), Hyperspectral imagery contains information which you can also draw a Histogram with (Continuous data - reflection rate across multiple contiguous bands). The Histogram curve, in the context of Hyperspectral imaging, is called as a Spectral Signature - it is like a fingerprint of the object / material under observation.
Scale Hyperspectral Imaging to surface level and what you get is a geospatial fingerprint - refer the Data Cube in Figure 6 which comprises stacked Hyperspectral images - these images are of the same surface albeit with visualized reflectance output across hundreds of spectral bands.
Thus, to summarize, a) higher Spectral resolution (narrow bands) & b) contiguous reflectance acquisition (spectral signature) are the standout features of Hyperspectral Imaging.
Let me elaborate the advantage of using Hyperspectral imagery using a simple analogy - imagine a transparent container stored at a cryogenic temperature (−150°C) which contains five substances in a freezing, solid state within. Upon heating the container to 150°C, you note down the exact temperature when each substance within undergoes a change in its state of matter. Upon charting your findings (Figure 7), what you are essentially depicting is the Temperature signature of the substances.
Let me ask you, 'which of the five substances is H2O (Water) and why?'
Very simple, right? Only Substance C is H2O as you know that H2O exists as ice below 0 degrees celsius, as water between 0-100 degrees and as water vapour above 100 degrees. The other four substances do not have the same temperature signature as that of water and hence, you can easily rule them out.
Now imagine that you maintain a temperature signature database of ten important substances that are of interest to you. Using the database you'll be able to detect which, if any, of the remaining four substances in the container are important to you. This is the exact utility of Hyperspectral Imaging - just that instead of temperature, we compare the energy reflectance acquired by a Hyperspectral sensor to a database of verified reflectance signatures of objects / materials that are of interest to us in order to detect / delineate / classify them.
For example, in Figure 8 below, observe the spectral signature of healthy, stressed and dry plants to solar radiation where, at lower wavelengths (400-800 nm i.e. visible light), there is distinct variability in reflectance rates. The variability can be strongly attributed to the pigments within - healthy plants appear greener due to the overshadowing presence of chlorophyll (green-colored pigment which helps the plant create food through photosynthesis) which absorbs red and blue wavelengths of visible light, reflecting only the green, hence the overall low reflectance. If due to adverse conditions the plant becomes stressed, it begins to lose its ability to develop new chlorophyll, and the proportion of other pale-colored pigments increases which gives the plant the yellowy-appearance and this is also when the rate of reflectance is comparatively the highest. However, the reflectance rate becomes very low again when the plant dries out, let's say due to drought. This is because all the pigments, be it chlorophyll or carotenoids, cease to develop and only the tannins are left behind which give the leaves its woody color and which eventually reflects only the red wavelength of visible light.
You'll appreciate how intricately visible light interacts with vegetation - the spectral signature obtained through Hyperspectral imaging gives profound insights.
Similarly, the variability in reflectance observed at higher wavelengths (1400-2400 nm i.e. intermediate infrared within solar radiation) can be strongly attributed to the water content within vegetation. However, unlike pigments, you'll observe that the variability is not intricate, rather it responds in a much more linear manner based on the health of the vegetation. This is because the amount of reflectance is directly proportional to the quantity of water - water absorbs infrared rays and it is also quintessential for healthy vegetation. Likewise, less water means more infrared reflectance and which also means stressed or dry vegetation.
Therefore, if one were to opt for Hyperspectral Imaging, one can extract the spectral signature of a pixel of agricultural land, and ascertain the health of the underlying vegetation by comparing it to research-validated spectral signatures.
The utility of Hyperspectral imaging also highlights the weakness of Multispectral imaging which is often less suitable for workflows that require finer delineation. Consider the spectral range of Sentinel-2 satellite's Multispectral Imager (MSI) instrument as depicted in Figure 9 above. While it is indeed possible to demarcate healthy vegetation from stressed or dry vegetation using the information captured within the spectral bands of Sentinel-2 imagery datasets, it would be challenging to distinguish whether the vegetation is highly-stressed or dry due to a) non-contiguous bands i.e. gaps in data and b) wider bands especially at higher wavelengths which normalizes reflectance sensitivity.
Remind yourself that how illumination interacts with an object is determined not by an individual factor but by the roughness, geometric & biogeochemical properties of the object as a whole (besides atmospheric influences and the wavelength itself). More of one and less of the other would significantly alter the spectral signature of the object under consideration - as depicted in Video 1 below which shows an iterative spectral signature of vegetation using one geometric influencer (leaf area index) and two biogeochemical influencers (chlorophyll, leaf water content). Such intricate reflectance behaviours are precisely why finer delineation using Multispectral imagery tends to be challenging and which is where the superiority of Hyperspectral imaging comes to the fore.
Similarly, see Figure 10 which shows the spectral signature of Open Water & Coastal Water. Sure, one could use Multispectral Imagery (the narrow B1 & B3 bands in Sentinel 2 MSI for example) to distinguish whether a pixel over an ocean is open or coastal, however it would be complicated as the difference in reflectance is tiny. With Hyperspectral imaging, the task would become much simpler through the derivation of spectral signature.
'Everything sounds so rosy about Hyperspectral Imaging. Do we really need Multispectral Imaging?'
Yes we do. Excess information is not always good information - it can be a double-edged sword. Because the spectral bands in Hyperspectral images are narrow and contiguous, researchers often encounter interferences from neighbouring bands while processing selected bands for a particular workflow. This is akin to cross-talk in telecommunications and is very problematic as it hampers the accuracy of delineation. Filtering the unwanted data signals / noise away isn't simple either - it necessitates the use of specialized and computationally-intensive corrective techniques. In contrast, the non-contiguous nature of spectral bands within Multispectral images ensures that data leakage doesn't happen. Moreover, the cost of acquiring Hyperspectral data and the expertise required to analyze it is very high compared to Multispectral data as of today.
[Some of the drawbacks of Multispectral imaging are applicable to Hyperspectral imaging as well - spaceborne acquisitions for earth observation can't be captured at night and are also susceptible to atmospheric influences during daytime]
Hyperspectral data can be acquired using different techniques, the choice of which depends on aspects such as the biogeochemical characteristics of the object or surface, sensitivity to external influences, desired resolution, cost constraints, and coverage requirements.
Some acquisitions are best captured using Ground-based sensors in a field / laboratory environment - especially when the object or surface under consideration is highly sensitive to external influences or where there is a need for high spatial resolution. Have a look at the two videos below - you'll realize that acquiring hyperspectral data using ground-based techniques requires considerable expertise.
Another laboratory-based Hyperspectral imagery acquisition video can be viewed here.
On the other hand, Spaceborne sensors are preferred when one desires wider and faster coverage, has cost considerations, has less need for high spatial resolution, and to study surfaces which have less reflectance sensitivity to external influences.
Airborne sensors - where the sensor system is installed on a research flight or a drone - are used extensively too - and are particularly effective for workflows which require more and faster coverage than a field acquisition and where there is a need for higher spatial resolution than what a spaceborne sensor provides.
Besides the mode of acquisition, there are different scanning technologies to choose from as well-
Fair to imagine that some research studies may require Hyperspectral data acquired using different modes and techniques in order to derive or validate the output.
The scope of utilizing Hyperspectral Imaging in scientific research is vast - it is already being used extensively in Agricultural Applications, Soil Applications, Mining Applications, Coastal Applications, Hazard Applications, Archaeological Applications and Military Applications, among several others (explore some Hyperspectral missions and applications here).
I was pleasantly surprised to know that what was originally designed for Remote Sensing has even found applications in crime scene detection, forensic medicine and biomedical sector, the latter where Imaging Spectroscopy is deployed as a non-invasive technique to distinguish cancerous tissues from healthy ones through the use of wavelengths which are able to penetrate outer skin.
Now that you've become familiar with the concept of Hyperspectral Imaging and its utility, I'll leave you with a video demonstration which depicts its practical use for Land Cover classification purpose-
Timestamps:
00:04Â - Video Details
00:19Â - Getting familiar with the Datasets
01:09Â - Exploring Airborne & Spaceborne Hyperspectral Imagery
03:23Â - Visualizing the Spectral Signature
04:50Â - Visualizing the Spectral Library
06:05Â - Using Regression Analysis to generate Land-Cover Map
Thanks for Watching!
ABOUT US
Intelloc Mapping Services | Mapmyops.com is based in Kolkata, India and engages in providing Mapping solutions that can be integrated with Operations Planning, Design and Audit workflows. These include but are not limited to - Drone Services, Subsurface Mapping Services, Location Analytics & App Development, Supply Chain Services, Remote Sensing Services and Wastewater Treatment. The services can be rendered pan-India, some even globally, and will aid an organization to meet its stated objectives especially pertaining to Operational Excellence, Cost Reduction, Sustainability and Growth.
Broadly, our area of expertise can be split into two categories - Geographic Mapping and Operations Mapping. The infographic below highlights our capabilities.
Our 'Mapping for Operations'-themed workflow demonstrations can be accessed from the firm's Website / YouTube Channel and an overview can be obtained from this flyer. Happy to address queries and respond to documented requirements. Custom Demonstration, Training & Trials are facilitated only on a paid-basis. Looking forward to being of service.
Regards,