Filters

Filters is the general purpose image processing sub toolbox: 46 tools that move a window over the raster and replace each pixel by something computed from its neighbours. Smoothing, sharpening, noise removal, texture measurement and morphological operations all work that way.

They apply to any raster, not only to imagery. A slope raster, a density surface or a classified map can all be filtered.

Four families

Smoothing replaces each pixel by an average of its surroundings, which removes noise and detail together. Mean, median and Gaussian are the classics.

Edge preserving smoothing does the same while refusing to blur across boundaries. Bilateral, guided, Kuwahara, anisotropic diffusion and non local means belong here. They are what you want on imagery you intend to classify afterwards.

Sharpening and high pass do the opposite, emphasising the fine detail. Unsharp masking is the standard.

Texture and morphology measure the local structure rather than the value: how variable the neighbourhood is, or how the shapes behave under erosion and dilation.

Start here

Median Filter

The first filter to try on noisy imagery. It replaces each pixel by the median of its neighbourhood, which removes isolated aberrant pixels completely while leaving edges sharper than any averaging filter would.

The window size is the only real parameter. Larger removes more noise and more detail.

Gaussian Filter

Mathematically optimal smoothing, weighted by distance. Its sigma parameter sets how far the blur reaches. It is the reference smoother, and the first step of many other algorithms, including edge detection.

Use Fast Almost Gaussian Filter on large rasters when the exact kernel is not required, since it approximates the same result much faster.

Bilateral Filter

The edge preserving smoother to know. It weighs neighbours both by distance and by how similar their values are, so pixels across a boundary barely contribute. The result is a cleaner image with its boundaries intact, which is exactly what a classifier needs.

Guided Filter and Non Local Means Filter solve the same problem with different methods, the second being the strongest denoiser and the slowest.

Glcm Texture

Measures texture rather than brightness, from the grey level co occurrence matrix: contrast, homogeneity, entropy and related descriptors. Texture separates classes that spectra cannot, such as a forest from a meadow of the same colour, or an orchard from a plantation.

A standard input to classification, and to OBIA object features.

Unsharp Masking

The standard sharpening method: subtract a smoothed copy from the original and add the difference back. Improves the readability of an orthophoto without inventing information.

All the tools

Smoothing

Tool	What it does
Mean Filter	Moving window average. The simplest smoother.
Median Filter	Moving window median. Removes isolated outliers, keeps edges sharper than the mean.
Gaussian Filter	Distance weighted smoothing, controlled by sigma.
Fast Almost Gaussian Filter	A fast approximation of the Gaussian, for large rasters.
Total Filter	Moving window sum.
Percentile Filter	The local percentile rank of the centre pixel.
Olympic Filter	Averages after removing the extreme values, like Olympic scoring.
Conservative Smoothing Filter	Only changes a pixel when it falls outside the range of its neighbours.
Savitzky Golay 2d Filter	Smoothing by local polynomial fitting, preserving peak shapes.

Edge preserving

Tool	What it does
Bilateral Filter	Weighs neighbours by distance and by similarity, so edges survive.
Guided Filter	Edge preserving smoothing constrained by a guide image.
Non Local Means Filter	Averages similar patches from across the image. The strongest denoiser.
Kuwahara Filter	Splits the window into quadrants and takes the most uniform one.
Anisotropic Diffusion Filter	Iterative diffusion that stops at edges.
Edge Preserving Mean Filter	Averages only the neighbours close in value to the centre.
K Nearest Mean Filter	Averages the k neighbours most similar to the centre pixel.
Adaptive Filter	Varies the filtering strength with the local statistics.
Wiener Filter	Adaptive noise reduction minimising the mean squared error.

Sharpening and high pass

Tool	What it does
Unsharp Masking	Sharpens by adding back the difference with a smoothed copy.
High Pass Filter	Isolates the high frequency detail.
High Pass Median Filter	High pass built on the local median.
High Pass Bilateral Filter	High pass built on a bilateral smoothing.
Diff Of Gaussians Filter	Difference between two Gaussian blurs, isolating a scale of detail.
Emboss Filter	Directional shading that gives a relief effect.

Texture and statistics

Tool	What it does
Glcm Texture	Texture descriptors from the grey level co occurrence matrix.
Gabor Filter Bank	Directional texture response at several orientations.
Standard Deviation Filter	Local standard deviation, a simple roughness measure.
Range Filter	Local maximum minus minimum.
Diversity Filter	Count of distinct values in the window. For classified rasters.
Majority Filter	Most frequent value in the window. Cleans a classified raster.
Maximum Filter	Local maximum.
Minimum Filter	Local minimum.
Integral Image Transform	Summed area table, used internally by fast filters.

Morphology and binary rasters

Tool	What it does
Opening	Erosion followed by dilation. Removes small bright features.
Closing	Dilation followed by erosion. Fills small dark gaps.
Tophat Transform	White or black top hat, isolating features smaller than the structuring element.
Line Thinning	Reduces binary features to one pixel wide skeletons.
Remove Spurs	Prunes the short spurs a thinning leaves behind.
Thicken Raster Line	Thickens diagonal lines so flow cannot leak through them.

Detection and other

Tool	What it does
Line Detection Filter	Detects linear features at several orientations.
Frangi Filter	Multiscale vesselness, for elongated thin structures.
Scharr Filter	Gradient operator, more rotationally accurate than Sobel.
Lee Filter	Speckle reduction for radar imagery. See also SAR.
User Defined Weights Filter	Applies a convolution kernel you supply.
Flip Image	Flips the raster vertically, horizontally or both.
Lidar Ground Point Filter	Slope based separation of ground from off ground points.

Notes

Window size is the parameter. Almost every tool here takes one, and it decides the scale you are working at. It is expressed in cells, so its meaning on the ground depends on your resolution.

Filter before classifying, not after. Cleaning a classified map with a majority filter is legitimate. Trying to fix a noisy classification that came from unfiltered imagery is not.