Multispectral Vs. Hyperspectral Remote Sensing: A Closer Look
In this episode, the discussion delves into the realms of multispectral and hyperspectral remote sensing. Often, the common assumption is that hyperspectral remote sensing is superior. However, this is not always the case. Depending on the application, multispectral can occasionally outshine hyperspectral.
Gordon Logie, an Earth Observation Scientist at Sparkgeo. With years of experience in remote sensing and handling both hyperspectral and multispectral data, he helps us to dive into their unique characteristics and how they address different challenges.
How Are Multispectral and Hyperspectral Remote Sensing Different?
Remote sensing involves measuring the reflectance of an object’s surface, an aspect that varies based on the physical characteristics of the object. The difference between multispectral remote sensing and hyperspectral remote sensing is primarily in the number of bands (or regions of the electromagnetic spectrum) they use to gather this data.
Multispectral remote sensing utilizes around 3 to 10 bands, mostly in the visible light spectrum and often with an additional NIR (near-infrared) band. The bands are broad, meaning that they capture a wide range of wavelengths. However, this breadth also limits their ability to distinguish between closely related wavelengths. For example, different shades of green could appear identical under a multispectral sensor.
Conversely, hyperspectral remote sensing uses a vast number of narrow bands (as many as 200). With the narrow and contiguous bandwidths, it provides a higher level of detail and more accurately recreates an object’s spectral signature. This enables hyperspectral sensors to distinguish even subtle differences in closely similar shades. For instance, they can distinguish between the various shades of green on a leaf, unlike multispectral sensors.
Signal to Noise Ratio and Image Resolution
When measuring surface reflectance, external fluctuations or internal sensor noise can distort the readings. Multispectral bands often have a high signal-to-noise ratio due to their broad bands that collect light over a wide range of wavelengths. Capturing more signal minimizes the effect of noise that may be introduced by random external fluctuations or from the sensor itself.
On the other hand, hyperspectral bands, being narrower, collect light over a smaller range. This means that if noise is present, it can have a significant impact on the hyperspectral data. To compensate for this, hyperspectral sensors often use larger pixel sizes, which gather more light despite the narrow spectral bands. However, increasing pixel sizes results in a decrease in spatial resolution. Consequently, hyperspectral images often have a lower resolution than multispectral ones. Thus, in applications where spatial detail is more important than spectral detail, multispectral sensors can outperform their hyperspectral counterparts.
Flexibility in Hyperspectral Remote Sensing
Despite having a lower spatial resolution, hyperspectral remote sensing offers more flexibility in addition to providing a greater level of spectral detail. Unlike multispectral sensors that are often designed for a particular application, hyperspectral offers all-in-one access to bands that are needed for different applications. There will be no need to use data from a bunch of different sensors – which would be the case if using a multispectral sensor. With this flexibility, one pick different bands from the same hyperspectral sensor to perform different applications.
Principal Component Analysis (PCA)
PCA is a technique that is used to summarize remote sensing data by pulling only useful information. Sensors capture everything that is within their field of view. But usually, it is not the whole dataset that contains the useful information needed for analysis – but only a part of it.
PCA summarizes the dataset into components – a handful of new bands that contain the key information, which reflects the most variation within the image. With the high number of bands in hyperspectral imaging, PCA is a common technique used in hyperspectral processing in order to get the salient details without having to handle all the bands in the data.
Using Hyperspectral Bands for Data Simulation
It is possible to aggregate hyperspectral bands in different ways to create broader, synthetic multispectral data. For instance, with a hyperspectral image, you can simulate what the scene would have looked like from Landsat, Sentinel Two, or any other number of sensors. This can be used for calibrating multispectral sensors. Also, you can create a synthetic image for different sensors from your hyperspectral image, and figure out what the relationship is between them. And since it is the same data underlying it, you can tell the differences in the images that are due to the different bands between them.
The Hughes Phenomenon and Its Implications
The Hughes phenomenon or the ‘curse of dimensionality’ is an important consideration when training machine learning algorithms to classify remote sensing data. It suggests that using more bands can increase the accuracy of a classifier up to a point, beyond which adding more bands may decrease the accuracy of the classifier. In cases where gathering more training data is not feasible, reducing the number of bands can be a solution, making multispectral sensors a superior choice despite their limited detail.
Common ways to counter the Hughes phenomenon is by using fewer bands or adding more training data. But in many cases acquiring more data may be more challenging. Therefore, multispectral data, with its fewer bands, may prove advantageous in certain situations.
Applications of Multispectral and Hyperspectral Remote Sensing in Real-World Scenarios
Here are a few ways the differences between multispectral and hyperspectral remote sensing are leveraged in different applications
If the task at hand demands detailed spatial information – for example, urban planning or precision agriculture – multispectral sensors often perform better. They provide higher-resolution data, making it easier to detect smaller features and subtle changes in the landscape.
Early Disease Detection in Agriculture
Contrarily, for early detection of diseases in crops, hyperspectral sensors excel. They pick up the narrow spectral differences that indicate the onset of disease, allowing farmers to take preemptive action. A multispectral sensor may only detect the disease when the crops show significant damage.
Minerals exhibit subtle spectral differences over a narrow range of wavelengths. These are generally invisible to multispectral sensors, but hyperspectral sensors can identify these unique spectral signatures. As a result, they are invaluable in mineral exploration – significantly increasing the efficiency and success rate of identifying potential mineral-rich sites.
Distinguishing between different plant species requires detecting minute spectral differences. Here again, hyperspectral sensors provide an advantage. They can help to map invasive weeds accurately over large farming landscapes and preserve biodiversity in ecological studies.
Accessing Hyperspectral Datasets and Tools
If you’re looking to delve into hyperspectral remote sensing, there are various tools and datasets available. For example, NASA’s decommissioned Hyperion sensor has open-access data available, and so does the AVIRIS airborne sensor.
In terms of tools, there are Python packages that are specifically designed for hyperspectral data processing. Additionally, most tools originally intended for multispectral data can also handle hyperspectral data since the only difference is in the number of bands. Some desktop applications that can process hyperspectral data include ArcGIS, QGIS, and ENVI.
Hyperspectral and multispectral remote sensing each have unique strengths that make them valuable for different applications. It is essential to understand these differences when choosing between the two. While hyperspectral data provides a higher spectral resolution, multispectral data shines in spatial detail. The choice between the two should be driven by the requirements of the specific application at hand.
Sponsored by Sinergise, as part of Copernicus Data Space Ecosystem knowledge sharing.
Here are some courses that focused on hyperspectral and offer further training