SAR Speckle Noise: What It Is and How to Filter It Without Losing Information

The first time I showed a Sentinel-1 image to a colleague trained in optical remote sensing, she asked why the image looked so grainy. "Is the sensor broken?" No — that granular, salt-and-pepper texture is speckle, and it's a fundamental property of coherent imaging, not a defect.

Understanding speckle is essential for anyone working with SAR data, because how you handle it determines whether your analysis preserves or destroys the information you're after.

Why SAR Images Look Grainy

SAR transmits coherent microwave radiation — the waves have a precise, known phase relationship. When this radiation hits the ground, it's scattered by everything within each resolution cell: individual rocks, soil grains, plant stems, building facets. Each scatterer returns a wave with slightly different amplitude and phase.

At the sensor, these returned waves combine through interference:

• Where wave crests align (constructive interference), the signal is amplified — the pixel appears bright
• Where crests meet troughs (destructive interference), the signal is reduced — the pixel appears dark

This interference pattern is deterministic — for the same surface viewed from the same geometry, you'd get the same speckle pattern. But it looks random because the sub-pixel scatterer arrangement varies from pixel to pixel.

The result: adjacent pixels over the same uniform surface (a flat agricultural field, for instance) can differ in brightness by a factor of 3 or more. This makes visual interpretation difficult and complicates quantitative analysis.

Speckle Is Multiplicative

This is the critical technical detail that determines which filters work and which don't.

In optical imagery, noise is additive: Observed = True signal + Noise. The noise is independent of signal strength.

In SAR imagery, speckle is multiplicative: Observed = True signal × Speckle. Brighter areas have proportionally more noise. A forest (high backscatter) has larger absolute speckle fluctuations than a calm lake (low backscatter), even though the relative variation is the same.

This multiplicative nature means standard image processing filters designed for additive Gaussian noise perform poorly on SAR data. A Gaussian blur or simple mean filter will:

• Over-smooth dark areas (where speckle is relatively mild)
• Under-smooth bright areas (where speckle is proportionally stronger)
• Blur edges between different surface types
• Destroy point targets (ships, buildings) that appear as bright spots

Adaptive Filters for SAR

SAR-specific filters adapt their behavior based on local image statistics — applying stronger smoothing in homogeneous areas (where speckle dominates) and weaker smoothing near edges and point targets (where spatial structure is important).

Lee Filter

The workhorse of SAR speckle filtering. The Lee filter estimates the "true" backscatter for each pixel using a weighted combination of the pixel's own value and the local mean:

Filtered = Mean + Weight × (Original − Mean)

The weight depends on the ratio of local variance to expected speckle variance:

• In uniform areas (variance ≈ speckle variance): Weight → 0, output ≈ local mean (strong smoothing)
• Near edges (variance >> speckle variance): Weight → 1, output ≈ original value (preserves edges)

Window sizes of 5×5 to 7×7 are typical. Larger windows smooth more but risk blurring real features.

Refined Lee Filter

An improvement that considers edge direction. Instead of using a square window, the refined Lee filter selects one of several directional sub-windows (horizontal, vertical, diagonal) that best aligns with any local edge. This better preserves linear features like roads, coastlines, and field boundaries.

Frost Filter

Similar philosophy to Lee but uses an exponentially weighted kernel that decreases with distance from the center pixel. The decay rate adapts to local statistics — faster decay (less smoothing) in heterogeneous areas, slower decay in uniform areas.

Gamma-MAP Filter

A Bayesian approach that assumes speckle follows a Gamma distribution (which it does for multi-look SAR data). The Gamma Maximum A Posteriori (MAP) estimator produces a statistically optimal estimate of the true backscatter given the observed speckled value.

In practice, Gamma-MAP produces slightly sharper results than Lee but is more computationally expensive.

Multi-Temporal Averaging

If multiple SAR acquisitions over the same area are available (common with Sentinel-1's 6-12 day revisit), temporal averaging provides speckle reduction without spatial resolution loss:

1. Co-register multiple images to sub-pixel accuracy
2. Average the intensity values pixel by pixel

Speckle from different dates is statistically independent (the sub-pixel scatterer arrangement looks different from each satellite pass). Averaging N images reduces speckle by a factor of √N:

Images Averaged	Speckle Reduction	Equivalent Number of Looks
1	None	Original ENL
4	2×	4 × original ENL
9	3×	9 × original ENL
16	4×	16 × original ENL

The catch: The surface must be stable between acquisitions. If crops are growing, water levels changing, or construction progressing, the temporal average will blur these real changes. Multi-temporal averaging is best for stable surfaces — bare soil, rock, urban areas, permanent water bodies.

When NOT to Filter

Speckle isn't always noise — sometimes it carries information:

Change detection: Comparing speckle patterns between two dates can reveal surface changes. A building demolished between two passes changes the scatterer arrangement, which changes the speckle. Aggressive filtering removes this signal.

InSAR: Interferometry requires preserving the phase information, which is destroyed by most intensity filters. InSAR has its own coherence-based noise reduction methods (multi-looking, Goldstein filtering) that operate in the complex domain.

Ship detection: Vessels appear as bright point targets against dark ocean. Speckle filters that smooth these bright pixels reduce detection sensitivity.

Texture analysis: Forest structure, urban density, and agricultural field patterns produce characteristic speckle textures. Filtering homogenizes the texture, removing the very signal you want to measure.

My Practical Recommendations

For visual interpretation and mapping: Apply a 5×5 Refined Lee filter. It preserves edges while making the image readable.

For quantitative backscatter analysis (crop monitoring, soil moisture): Use multi-temporal averaging when possible (3-5 dates from the same orbit), supplemented by a gentle 3×3 Lee filter.

For change detection: Minimize filtering. Apply the same minimal filter to both dates, or work with unfiltered data and apply statistical change detection methods designed for multiplicative noise.

For classification: Filter before classifying. Speckle degrades classifier performance significantly — the noise introduces within-class variance that confuses the algorithm.

The key principle: filter as little as necessary for your application. Every filter makes a trade-off between noise reduction and information loss. Understanding that trade-off — and knowing when speckle is noise versus signal — is what separates competent SAR analysis from processing by default.