SAR

SAR, synthetic aperture radar, is imagery acquired by an instrument that emits its own microwave signal and measures what comes back. Two properties make it worth a sub toolbox of its own.

It sees through clouds and works at night, which is why it is the reference source for flood mapping, for monitoring in tropical regions, and for any emergency where you cannot wait for a clear sky.

And it measures structure rather than colour. A radar return depends on roughness, geometry and moisture, so a smooth water surface comes back almost black, a forest canopy comes back bright and diffuse, and a building facing the sensor comes back very bright through a double bounce.

Speckle, and why every workflow starts with a filter

Radar images carry speckle, a grainy salt and pepper texture that is not noise in the ordinary sense but the interference of the many scatterers inside each resolution cell. It is inherent to the technique and it defeats any classifier applied directly.

Speckle is multiplicative, meaning it scales with the signal, which is why the ordinary smoothing filters of Filters handle it badly. The filters here are built for it.

Filter	Character
Lee Filter	The classic. Adapts to the local statistics, keeps edges reasonably well.
Refined Lee Filter	Adds directional analysis, so edges and thin features survive better. The usual recommendation.
Enhanced Lee Filter	Refined variance estimation and multiple scales. The strongest of the family.
Frost Filter	Exponentially damped weighting based on local statistics. Good detail preservation.
Kuan Filter	Similar to Lee, with a different derivation of the weights.
Gamma Map Filter	Assumes a Gamma distribution explicitly, which fits radar statistics well.

Start with Refined Lee. Compare the result with the original at full zoom: a filter that removed the speckle and the field boundaries with it was too strong.

Polarimetric decompositions

Radar can transmit and receive in different polarisations. A quad polarimetric acquisition holds enough information to work out what kind of scattering happened in each pixel, which is far more informative than brightness alone.

Freeman Durden Decomposition splits the signal into three physical mechanisms: surface scattering, typical of bare soil and water, double bounce, typical of buildings and flooded forest, and volume scattering, typical of canopy. Displayed as an RGB composite, it reads almost like a land cover map.

Cloude Pottier Decomposition derives entropy, anisotropy and the alpha angle, the standard parameter space for describing scattering mechanisms.

Yamaguchi 4component Decomposition adds a fourth component for the helix scattering found in complex urban areas.

H Alpha Wisart Classification and Wisart Iterative Clustering then classify the result, the first by partitioning the entropy and alpha space into physically defined zones, the second by iterative clustering with the Wishart distance appropriate to polarimetric data.

All the tools

Tool	What it does
Lee Filter	Speckle reduction with Lee’s multiplicative model.
Refined Lee Filter	Lee filtering with directional analysis for better edge preservation.
Enhanced Lee Filter	Lee filtering with refined variance estimation and multiple scales.
Frost Filter	Speckle reduction with exponentially damped local weighting.
Kuan Filter	Speckle reduction with locally adaptive weights.
Gamma Map Filter	Speckle reduction assuming a Gamma distribution.
Freeman Durden Decomposition	Splits quad polarimetric data into surface, double bounce and volume scattering.
Cloude Pottier Decomposition	Derives entropy, anisotropy and alpha from the coherency matrix.
Yamaguchi 4component Decomposition	Four component decomposition including helix scattering.
H Alpha Wisart Classification	Unsupervised classification in the entropy and alpha space, refined by Wishart clustering.
Wisart Iterative Clustering	Iterative unsupervised clustering with the Wishart distance.

Notes

Sentinel-1 is free. The European Copernicus programme distributes SAR imagery of the whole planet at no cost, every few days. It is the practical starting point for anyone wanting to try these tools.

Decompositions need quad polarimetric data. Freeman Durden, Cloude Pottier and Yamaguchi expect a full polarimetric acquisition. A dual polarimetric product, which is what Sentinel-1 usually delivers, supports the speckle filters and some decompositions but not all of them.

Terrain distorts radar. In mountains, slopes facing the sensor are compressed and those facing away are stretched or hidden. A radar product should be terrain corrected before being compared with anything else.