Understanding Types of Satellite Imagery

Background

There has recently been great progress in the world of satellite imagery, this introduces vast opportunities for a whole host of industries. Few people I've spoken to recognise the potential, I usually hear some variation of "Using satellite data for x is so 2010s". I would be tempted to agree if it were not for the fact that the world of satellites is not static. Regardless of the pace of improvements we see in AI development, any model remains reliant on high quality data. Access to superior data, obtained through processing improved remote sensing data creates the possibility of superior estimations.

Monitoring and forecasting mechanisms that previously required feet on the ground, expensive drones and flux towers can increasingly be done through remote sensing.

Increasing constellations of micro-satellites that can give us frequent updates thus removing temporal resolutions limitations that existed in the past, sensors that can give us greater spectral and spatial resolution as well as greater callibration algorithms that give us data with lower noise.

The increase in spectral resolution is what I personally find most exciting, we're quickly entering a world of Hyperspectral satellites and I find myself being one of the few startup founders with sufficient understanding of these and the relevant software engineering capabilities to take advantage of them.

Below I detail some of the different types of satellite instruments and the types of images they produce.

Panchromatic Images

"A panchromatic image is a single-band grayscale image with a high spatial resolution that “combines” the information from the visible R, G, and B bands."[1] Recently these have started including Near Infrared bands as well.

The wide spectral resolution (measurement of R,G,B and perhaps Near Infrared in a single band) allows for a lot of signal and limited noise for an area. This allows for a very high spatial resolution, in some cases allowing you to obtain images at a resolution of up to 0.5m

Multispectral Images

Multispectral images are created by spectroradiometers that are able to detect and isolate specific broad bands of electromagnetic radiation. These images usually involve splitting the measurement of solar radiation reflected off a target into 3-20 bands, with each band composed of a bandwidth (fwhm) of ~15-40nm.

The most basic of these involve splitting the visible Red, Green and Blue lights as well as Near Infrared into separately identifiable bands. With more advanced satellites such as the ESA's Sentinel-5 having 7 different bands ranging from Ultraviolet to Visible and even Short Wave Infrared. All light detected within a certain band is treated the same so there's no significant distinction between specific wavelengths within a band.

The narrower bands handled by each radiometer also reduce the signal to noise ratio, this raises the impact of any atmospheric interference such as cloud cover.

Hyperspectral Images

These are images split into hundreds of contiguous spectroscopic bands with a bandwidth (fwhm size of each band) of ~5-15nm.

All materials have a sort of solar reflected signature. Absorptions of certain wavelengths of electromagnetic radiation and reflections of others in certain scattering patterns that allow us to identify the presence of a material, its characteristics and perhaps its health. This has previously been applied for use cases such as identification of some types of rocks and minerals and understanding of soil moisture. Usage of hyperspectral data has yet to become widespread however. This is due to 4 key reasons:

Hyperspectral instruments are relatively new, the first (successful) one was launched by NASA in 2000 (Hyperion). You can still test this here Of the almost 6,000 satellites orbitting Earth, only about 19 of them contain Hyperspectral imaging sensors.
Hyperspectral instruments split electromagnetic radiation into very narrow bands so are highly impacted by noise caused from atmospheric interference or even noise produced by the sensor itself. This means to acquire a good image they require a lower spatial resolution, a higher pixel size allowing more light and thus more signal. ESA's Sentinel-5 for example has a spatial resolution of 7km x 3.5km. This is useful for region level analysis but not for anything granular such as monitoring or discovering specific mineral deposits.
Hyperspectral data also suffers from the curse of dimensionality when engaging in supervised classification models due to the sheer amount of detail unless there is a lot of training data.
There is limited software available to process this data and insufficient remote sensing experts with advanced software skills available to process these complex images.

The Exciting Bit

The above limitations are changing very, very quickly! Massive investment has been made in improving sensors and callibration algorithms, this has led to advancements in spectroscopy instruments that are being trialed as we speak. These possess both high spectral resolution as well as high spatial resolution, creating the potential to revolutionise how we view and analyse our world.

References

[1] https://gisrsstudy.com/panchromatic-sensor/

[2] Lechner, Alex & Foody, Giles & Boyd, Doreen. (2020). Applications in Remote Sensing to Forest Ecology and Management. One Earth. 2. 405-412. 10.1016/j.oneear.2020.05.001.