An optical satellite is a camera. It records sunlight that the ground reflects, so it needs daylight and a clear sky. A radar satellite is a flashlight and a camera in one. It sends out its own pulses of microwaves, records the echo, and so it works at night and through cloud, smoke and rain. Sentinel-1 is the radar satellite of the Copernicus programme. It is free, it covers the land of the whole Earth every few days, and it is the standard choice for mapping floods in the monsoon, when clouds hide the ground for weeks.
Synthetic aperture radar
A radar image needs fine detail along the track of the satellite, and a real antenna would have to be kilometres long. Synthetic aperture radar, SAR, solves this by using the motion of the satellite. As it flies, the satellite sends many pulses and records the echoes of the same ground from many positions. The computer combines them as if they came from one very long antenna. The result is an image with cells of a few metres from an orbit 700 km up.
Sentinel-1 works in the C band, with a wavelength of about 5.5 cm. In its usual mode over land it covers a strip 250 km wide, and the image has a pixel of 10 m, although the real resolution is coarser, about 20 m. It sees the same place every 6 to 12 days, from an ascending orbit, heading north, or a descending orbit, heading south.
Backscatter
Each pixel of a radar image holds the backscatter, the share of the transmitted energy that comes back to the antenna. What comes back depends on three things about the surface.
- Roughness. A smooth surface, such as calm water or a paved runway, acts like a mirror. The energy bounces away from the satellite, and little returns. Smooth surfaces are dark. A rough surface, such as a ploughed field or a forest, scatters the energy in all directions, and some returns.
- Shape and structure. Buildings and bridges form corners with the ground, and a corner sends the energy straight back, in a strong double bounce. Cities are bright.
- Moisture. Wet soil and wet vegetation return more energy than dry, because water changes how the surface interacts with microwaves.
So in a radar image, calm water is very dark, farmland is grey, forest is a lighter grey, and built-up areas are the brightest. The radar guide shows this on a real image of Punjab, where VV over farmland is typically around -12 dB, water is below about -18 dB, and buildings are often above -5 dB.
Polarization
A radar wave vibrates in a direction, which is its polarization. A satellite can send it vertically, V, or horizontally, H, and can listen for either. Sentinel-1 over land usually gives two bands:
- VV: sent vertical, received vertical. It is strong over smooth surfaces and over water, so it is the band for finding water.
- VH: sent vertical, received horizontal. The wave changes direction when it bounces around inside vegetation, so VH responds to crops and trees. It is weaker than VV, by about 6 dB in the guide's example.
The ratio of the two is also used, as a measure of vegetation.
Decibels
Backscatter is stored as a linear number, and it covers a range of thousands, because a bright building is thousands of times stronger than dark water. On a linear scale, almost the whole image would look black. So backscatter is converted to decibels:
dB = 10 x log10(linear value)
Every step of 10 dB is a factor of 10, and 3 dB is about a factor of 2. Values are negative for most land, since the linear value is less than 1. The conversion is not a detail: the average of linear values and the average of decibel values are different numbers, so decide where to average before you start, and say so.
Speckle
A radar image looks grainy even over a flat field. The grain is speckle. Within one pixel, the echoes of many small scatterers add up, sometimes helping and sometimes cancelling each other, by chance. Two neighbouring pixels of the same field then differ, with no real difference on the ground.
Speckle is not noise to be removed by better hardware. It is part of how coherent imaging works. To reduce it, you average: over a window of pixels, as in a 5 by 5 filter, or over several dates. Both cost something. A window blurs the detail, and several dates only suit a surface that does not change between them.
Terrain and geometry
A radar looks sideways, and the distances it measures are along the line of sight. In hills this distorts the picture.
- Foreshortening. A slope facing the radar is squeezed, and looks short and bright.
- Layover. The top of a steep slope can be closer to the radar than the bottom, and the image folds over itself.
- Shadow. A slope facing away from the radar is hidden behind the ridge, and records nothing.
For flat country such as the plains of Punjab these matter little. In hills, products that are terrain corrected, with the correction for the angle of the ground to the beam, are the right choice, for example the radiometrically terrain corrected Sentinel-1 data on the Planetary Computer used in the guide. Choose scenes of the same orbit when you compare dates, because the angle changes the numbers.
Radar for floods
Calm floodwater is smooth, so it turns dark in VV. A pair of images, one before and one after, shows the new dark areas. This is the standard method, and its checks are simple: look for the dark pixels that were not dark before, remove permanent water, and remove steep slopes and radar shadow, which are also dark. Wind roughens open water and makes it brighter, and flooded vegetation or flooded towns can get brighter and not darker, because of the double bounce. The method misses these, and a radar flood map shows a minimum.
Where radar falls short
- The picture is harder to read than a photograph, and needs practice.
- A radar image does not give colour or the spectral signature of a surface, so it is poor for telling crops by their leaves.
- Rain and wind can change what the sensor sees, though far less than they affect optical data.
- The numbers depend on moisture, so a change in value can be wetter soil and not a change of cover.