Light, spectra and the physics behind remote sensing

A satellite image looks like a photograph, so it is easy to forget that it is a measurement. Every pixel is a number that says how much light of a certain colour arrived at a sensor. Between the Sun and that sensor, the light was reflected, absorbed, scattered and filtered. Everything in remote sensing, and everything that a machine learning model learns from these images, rests on that physics.

You do not need to be a physicist to use it. A grasp of a few ideas will help you choose the right data, understand strange results and explain to others why a model works or fails.

Light as a wave

Light is an electromagnetic wave, and its colour is set by its wavelength. Visible light spans roughly 400 to 700 nanometres, from violet to red. Beyond red lies the near-infrared, then short-wave infrared, then thermal infrared. Beyond violet lies ultraviolet. Radio waves and microwaves, used by radar, have much longer wavelengths.

• Visible, about 400 to 700 nm: what we see.
• Near-infrared, about 700 to 1,300 nm: invisible to us, but strongly reflected by healthy leaves.
• Short-wave infrared, about 1,300 to 2,500 nm: sensitive to water in plants, soil and snow.
• Thermal infrared: emitted heat, which measures temperature.
• Microwaves: used by radar, which supplies its own illumination.

What happens when light hits the ground

When sunlight reaches a surface, some of it is reflected, some is absorbed and some passes through. The share reflected at each wavelength is the surface’s reflectance. A sensor mostly measures reflected light, so reflectance is the quantity that links the image to the land.

Spectral signatures

Each material reflects different amounts at different wavelengths, giving it a characteristic curve called a spectral signature.

• Healthy vegetation absorbs red and blue light for photosynthesis and reflects green, so leaves look green. Inside the leaf, cell structure scatters near-infrared strongly, so vegetation is very bright there. The sharp jump between red and near-infrared is the vegetation edge.
• Water absorbs near-infrared and short-wave infrared almost completely, so it is very dark in those bands, and this is why water is easy to find with them.
• Bare soil reflects gradually more with longer wavelengths, and its brightness changes with moisture and mineral content.
• Snow is very bright in visible light, and dark in short-wave infrared, which helps tell it from cloud.
• Built surfaces vary widely, from dark asphalt to bright concrete and roofs.

Why indices such as NDVI work

The Normalized Difference Vegetation Index uses the red and near-infrared bands, where healthy vegetation is most different from everything else. Because it is a ratio, brightness that affects both bands equally, such as overall illumination, mostly cancels out. This is why NDVI compares dates and places better than a single raw band does. It also explains its limits: very dense vegetation saturates the index, and soil can shift the values in sparse crops.

The atmosphere gets in the way

Between the surface and the sensor lies air, with gas, water vapour, dust and smoke. It affects the signal in two ways.

• Scattering. Short wavelengths scatter most, which is why the sky is blue and distant hills look hazy. Haze adds a veil of brightness to dark features.
• Absorption. Water vapour, carbon dioxide and ozone absorb at specific wavelengths, which is why sensors choose bands in the clear “windows” between them.

Atmospheric correction estimates and removes these effects to give surface reflectance, which is comparable between dates. This is why most analysts prefer corrected products over raw top-of-atmosphere data.

Sun angle, shadows and season

The same field looks different at different times. A low winter Sun casts long shadows, while a high summer Sun casts short ones. Sloping ground facing the Sun is bright, and facing away it is dark. Shadows lower brightness in every band, and a dark pixel in a shadow can resemble water. This is a classic cause of false water in mountain regions, and a reason to use terrain data when mapping.

Seasons change the land too. A crop looks like bare soil at planting, green at its peak and brown at harvest. A model that has only seen the peak will struggle at other times.

Radar is a different kind of measurement

Radar satellites send out microwaves and record what returns, which is called backscatter. Because they provide their own illumination and microwaves pass through cloud, they work at night and in bad weather. What they measure is different from light.

• Smooth surfaces such as calm water reflect the signal away from the sensor, so they look dark.
• Rough surfaces such as ploughed fields, forest and cities scatter more back, so they look brighter.
• Wavelength matters: longer waves penetrate vegetation more deeply.
• Images contain a grainy texture called speckle, which needs careful filtering.

What this means for machine learning

• A model learns the numbers, not the land. If the numbers shift because of haze or sun angle, a model can fail without the land changing at all.
• Training on one sensor and applying to another is risky, because their bands differ slightly.
• Using surface reflectance, and adding information such as terrain and season, makes models more robust.
• When a result looks odd, ask what the physics predicts. Is this a shadow? Is it haze? Is it a different season?

Frequently asked questions

Why is vegetation so bright in near-infrared?

Because the internal structure of healthy leaves scatters near-infrared light strongly, and the leaf’s pigments do not absorb it. This is not visible to the eye, but sensors measure it clearly.

What is the difference between radiance and reflectance?

Radiance is the light a sensor measures, and it depends on the Sun and atmosphere. Reflectance is the fraction of incoming light that a surface reflects, which is a property of the surface itself.

Why do I need to correct for the atmosphere?

So that images from different dates and places can be compared. Without correction, haze and sun angle can change values as much as real changes on the ground.

Can radar replace optical imagery?

Not entirely. Radar shines when cloud is a problem or when you need to measure water, roughness or movement. Optical imagery is better for colour and for vegetation health. Many projects use both.

Remote sensing is applied physics. When you keep the journey of the light in mind, from the Sun to the surface to the sensor, odd results become clues, and your choices of data and methods become much more confident.