Sentinel-1 SAR Processing with S1ARDProcessor

Sentinel-1 C-band SAR sees through clouds and night, and therefore fills the gap left by optical sensors in cloud-prone areas (boreal, tropical, monsoon) and in rapidly changing systems (crops, wetlands, floods). But raw GRD scenes are not directly comparable in time: they are affected by speckle noise and by topographic distortions that depend on the scene’s acquisition geometry.

S1ARDProcessor turns Sentinel-1 GRD data into Analysis Ready Data (ARD) inside Google Earth Engine. It applies radiometric terrain correction and one of five speckle filters, producing time-consistent VV and VH backscatter that can be used for indices, composites, classification, or time-series analysis.

The processor follows the angular-based radiometric slope correction method of Vollrath et al. (2020).

---

What ARD gives you

Problem	Consequence	What ARD does
Speckle noise	High pixel-to-pixel variance, unreliable thresholds	Adaptive spatial filtering (Lee variants, Gamma MAP, Lee Sigma, Boxcar)
Terrain distortion	Same land cover with different backscatter on pole-facing vs sun-facing slopes	Radiometric correction from local incidence angle (slope, aspect, orbit)
Geometric inconsistency between orbits	Asc vs desc scenes not comparable	Filter by orbit keeping only one pass direction
Raw linear power is hard to interpret	Non-Gaussian dynamic range	Optional conversion to decibel scale

In short: without ARD, pixel values between two dates are not directly comparable. With ARD, they are.

---

Two ways to use it

S1ARDProcessor can be used directly on any ee.ImageCollection of Sentinel-1 GRD scenes, or indirectly through NdviSeasonality by passing use_sar_ard=True. The latter is the usual path: it reuses the same ROI, date range, and orbit logic that you already configure for optical sensors.

import ee
from ndvi2gif import NdviSeasonality, S1ARDProcessor

ee.Initialize()

---

Quick start — through NdviSeasonality

roi = ee.Geometry.Rectangle([-6.55, 36.85, -6.10, 37.20])

ns = NdviSeasonality(
    roi=roi,
    sat='S1',
    index='vh',
    start_year=2022,
    end_year=2023,
    periods=12,
    key='median',
    orbit='DESCENDING',       # pick one pass direction for temporal consistency
    use_sar_ard=True,          # enable ARD pipeline
    sar_speckle_filter='REFINED_LEE',
    sar_terrain_correction=True,
    sar_terrain_model='VOLUME',
)

composite = ns.get_year_composite()
# ee.ImageCollection: 2 images (2022, 2023), each with 12 bands (january..december)

That’s the full workflow: orbit-filtered, terrain-corrected, speckle-filtered, reduced to monthly medians of VH backscatter. Everything happens lazily in GEE — no download.

---

Direct use — any GRD collection

If you need ARD outside of NdviSeasonality (e.g. to build a custom pipeline, to feed a classifier directly, or to preprocess a single scene), instantiate the processor and map it:

processor = S1ARDProcessor(
    speckle_filter='REFINED_LEE',
    speckle_filter_kernel_size=7,
    terrain_correction=True,
    terrain_flattening_model='VOLUME',
    dem='COPERNICUS_30',
    format='LINEAR',
)

s1 = (ee.ImageCollection('COPERNICUS/S1_GRD')
        .filter(ee.Filter.listContains('transmitterReceiverPolarisation', 'VV'))
        .filter(ee.Filter.listContains('transmitterReceiverPolarisation', 'VH'))
        .filter(ee.Filter.eq('instrumentMode', 'IW'))
        .filter(ee.Filter.eq('orbitProperties_pass', 'DESCENDING'))
        .filterBounds(roi)
        .filterDate('2022-01-01', '2023-01-01')
        .select(['VV', 'VH', 'angle']))

s1_ard = s1.map(processor.process_image)

process_image runs the full chain in order: terrain correction → speckle filter → format conversion. Terrain correction expects the angle band, so make sure to include it in the select() before mapping.

Angle band is required for terrain correction. If terrain_correction=True, keep 'angle' in the band selection. You can drop it afterwards with .select(['VV', 'VH']).

---

Speckle filters

Five algorithms are available, covering the usual speed-vs-quality trade-off:

Filter	Strengths	When to use
REFINED_LEE	Detects edges, preserves structure, low texture loss	Recommended default — vegetation, crops, forests
LEE	Adaptive, good homogeneity/edge balance	Light, general-purpose filtering
GAMMA_MAP	Statistical, preserves texture	Areas with meaningful texture (forest canopy, urban)
LEE_SIGMA	Preserves point targets and bright features	Scenes with infrastructure, isolated bright scatterers
BOXCAR	Simple mean, very fast	Quick exploration, prototyping
None	No filter	If you intend to do your own temporal filtering

All filters assume an Equivalent Number of Looks (ENL) of 4.4 for Sentinel-1 IW GRD. The kernel size (speckle_filter_kernel_size) must be odd: 3, 5, 7 (default), 9. Larger kernels smooth more but also blur structure — start at 7 and only increase if the remaining speckle is visible in your application.

Caveat — directional kernels for Refined Lee. The current implementation uses fixed 3×3 directional kernels for edge detection regardless of the speckle_filter_kernel_size. In practice this still outperforms the plain Lee filter for S1 IW data, but if you need very large smoothing you may want to run a second pass with a larger kernel.

---

Terrain correction

Radiometric terrain correction (RTC) compensates for the backscatter anisotropy introduced by local slope and aspect relative to the sensor’s look geometry. It is essential in mountainous terrain; without it, a forest on a radar-facing slope appears brighter than the same forest on a radar-averted slope.

Scattering model

Model	Assumption	Use for
VOLUME	Scattering inside a 3D medium	Vegetation, crops, forests — default
SURFACE	Scattering from a 2D surface	Bare soil, rock, water

The VOLUME model uses the full local-incidence-angle formulation (Vollrath et al. 2020); SURFACE uses a simpler cosine-of-slope correction. For mixed landscapes, VOLUME is usually the safer choice.

DEM choice

DEM	Resolution	Notes
COPERNICUS_30	30 m	Recommended default — best global quality, covers high latitudes
COPERNICUS_90	90 m	Faster, use for very large ROIs
SRTM_30	30 m	Classic, no coverage above 60°N / below 56°S
SRTM_90	90 m	Classic, same latitude limits

The correction factor is clamped to the [0.5, 2.0] range to prevent overcorrection in very steep terrain where the angular geometry becomes unstable.

---

Orbit filtering: why it matters

Sentinel-1 sees the same area from ascending (SSW→NNE look direction) and descending (NNE→SSW look direction) orbits. Backscatter depends on look angle, so mixing orbits in a time series adds artificial variability that is not related to the ground.

Always pass orbit='ASCENDING' or orbit='DESCENDING' when building a time series. 'BOTH' is only appropriate when revisit frequency is more important than radiometric consistency (e.g. rapid flood mapping).

# Temporal consistency (time series, trend analysis, classification)
NdviSeasonality(..., sat='S1', orbit='DESCENDING', use_sar_ard=True)

# Maximum coverage (flood mapping, single-date snapshots)
NdviSeasonality(..., sat='S1', orbit='BOTH', use_sar_ard=True)

Choose the direction with better coverage over your ROI — the GEE code editor shows per-orbit footprints, and for most Mediterranean sites both directions are well covered.

---

SAR indices available

Once the collection is ARD-processed, the usual NdviSeasonality index interface applies:

Key	Name	Formula	Use
vv	VV polarization	VV	Surface roughness, bare soil, water
vh	VH polarization	VH	Vegetation volume scattering — default for vegetation
rvi	Radar Vegetation Index	4·VH / (VV+VH)	Vegetation water content, robust to speckle
vv_vh_ratio	VV/VH ratio	VV / VH	Structural change, mowing, ploughing
dpsvi	Dual-pol SAR Vegetation Index	(VV−VH) / (VV+VH)	Dense canopies, crop growth
rfdi	Radar Forest Degradation Index	(VV−VH) / VV	Forest disturbance
vsdi	Vegetation Scattering Diversity Index	sqrt((VV−VH)² + (VV+VH)²)	Scattering diversity, biomass

All of them accept normalize_sar=True at the NdviSeasonality level to rescale output to [0, 1] — useful for visualisation and for stacking with optical indices in a classifier.

---

Full example — crop phenology with VH + RVI

from ndvi2gif import NdviSeasonality

# Monthly VH and RVI over an agricultural area
ns_vh = NdviSeasonality(
    roi=roi,
    sat='S1',
    index='vh',
    start_year=2020, end_year=2023,
    periods=12,
    key='median',
    orbit='DESCENDING',
    use_sar_ard=True,
    sar_speckle_filter='REFINED_LEE',
    sar_terrain_correction=True,
    sar_terrain_model='VOLUME',
)
composite_vh = ns_vh.get_year_composite()

ns_rvi = NdviSeasonality(
    roi=roi,
    sat='S1',
    index='rvi',
    start_year=2020, end_year=2023,
    periods=12,
    key='median',
    orbit='DESCENDING',
    use_sar_ard=True,
    normalize_sar=True,
)
composite_rvi = ns_rvi.get_year_composite()

Add to a map to compare summer (dry, low VH) and spring (active vegetation, high VH):

import geemap

Map = geemap.Map()
Map.centerObject(roi, zoom=11)

vh_viz  = {'bands': ['july'], 'min': 0.0005, 'max': 0.1,
           'palette': ['#000000', '#7f7f7f', '#ffffff']}
rvi_viz = {'bands': ['july'], 'min': 0.0, 'max': 1.0,
           'palette': ['#c7e9b4', '#41b6c4', '#253494']}

Map.addLayer(composite_vh.select('july').median(),  vh_viz,  'VH July (median)')
Map.addLayer(composite_rvi.select('july').median(), rvi_viz, 'RVI July (median)')
Map

---

Format: linear vs decibel

Keep the internal format in 'LINEAR' during pipeline processing — SAR indices (RVI, DPSVI, VV_VH_ratio, etc.) must be computed on linear power, not on logarithms. NdviSeasonality enforces this automatically.

Convert to decibel at the end if you want to visualise VV or VH directly:

processor = S1ARDProcessor(format='DB')  # applies 10·log10 at the end

format='DB' compresses dynamic range for display but is not interchangeable with linear for downstream arithmetic.

---

Tips and caveats

Don’t skip terrain correction in mountainous ROIs

Even in gently sloping terrain, RTC improves temporal stability of backscatter. Skip it only for flat coastal/agricultural areas where the computation cost matters and the DEM adds no information.

Match the kernel to your target size

A speckle_filter_kernel_size=7 (default) is a 70 m window at S1 native resolution. For very fine-grained work (e.g. small reservoirs, individual field boundaries), drop to 5 or 3 to preserve detail. For coarse monitoring (biomass over forest stands), 9 is reasonable.

ENL assumption

All Lee-family and Gamma MAP filters assume ENL = 4.4, which is the standard value for S1 IW GRD. If you apply ARD to other SAR modes (EW, SM) you should recompute ENL — but S1ARDProcessor is targeted at IW, which is what 99% of ndvi2gif users will use.

Cloud filter has no effect

cloud_filter and max_cloud_cover are ignored for sat='S1' — SAR is not affected by clouds. They are kept in the API only for interface consistency with optical sensors.

Computational cost

use_sar_ard=True adds meaningful server-side work — terrain correction involves a DEM join plus per-pixel trigonometry, and Refined Lee runs four directional convolutions. For large ROIs or long time spans, expect getInfo() and map rendering to be noticeably slower. Export to Drive/Asset instead of pulling via getInfo().

Debugging

If output backscatter looks off, disable one stage at a time:

# Stage 1: raw + basic filter
ns = NdviSeasonality(..., sat='S1', use_sar_ard=False)

# Stage 2: ARD without terrain correction
ns = NdviSeasonality(..., sat='S1', use_sar_ard=True,
                     sar_terrain_correction=False)

# Stage 3: full ARD
ns = NdviSeasonality(..., sat='S1', use_sar_ard=True,
                     sar_terrain_correction=True)

This isolates whether a problem comes from the DEM/angle geometry or from the speckle filter.

---

API reference

S1ARDProcessor(...)

Parameter	Type	Default	Description
speckle_filter	str or None	'REFINED_LEE'	'BOXCAR', 'LEE', 'REFINED_LEE', 'GAMMA_MAP', 'LEE_SIGMA', or None
speckle_filter_kernel_size	int (odd)	7	Filter kernel size in pixels
terrain_correction	bool	True	Apply radiometric terrain correction
terrain_flattening_model	str	'VOLUME'	'VOLUME' (vegetation) or 'SURFACE' (bare soil/water)
dem	str	'COPERNICUS_30'	'COPERNICUS_30', 'COPERNICUS_90', 'SRTM_30', 'SRTM_90'
format	str	'LINEAR'	'LINEAR' (for indices) or 'DB' (for visualisation)

Methods

Method	Returns	Description
process_image(image)	ee.Image	Full pipeline: terrain correction → speckle filter → format conversion
apply_terrain_correction(image)	ee.Image	Only the RTC step
apply_speckle_filter(image)	ee.Image	Only the speckle step (routes to the selected filter)
to_db(image)	ee.Image	Convert VV and VH from linear to decibel

NdviSeasonality SAR-specific parameters

Parameter	Default	Notes
sat='S1'	—	Select Sentinel-1 as source
orbit	'BOTH'	'ASCENDING', 'DESCENDING', or 'BOTH'
use_sar_ard	True	Enable the S1ARDProcessor pipeline
sar_speckle_filter	'REFINED_LEE'	Passed to the processor
sar_terrain_correction	True	Passed to the processor
sar_terrain_model	'VOLUME'	Passed to the processor
normalize_sar	False	Rescale SAR indices to [0, 1]

---

References

Vollrath, A., Mullissa, A., Reiche, J. (2020). Angular-based radiometric slope correction for Sentinel-1 on Google Earth Engine. Remote Sensing, 12(11), 1867.

Lee, J.S. (1980). Digital image enhancement and noise filtering by use of local statistics. IEEE Transactions on Pattern Analysis and Machine Intelligence, (2), 165–168.

Lee, J.S. (1983). Digital image smoothing and the sigma filter. Computer Vision, Graphics, and Image Processing, 24(2), 255–269.

Lopes, A., Nezry, E., Touzi, R., Laur, H. (1993). Structure detection and statistical adaptive speckle filtering in SAR images. International Journal of Remote Sensing, 14(9), 1735–1758.

Kim, Y., Jackson, T., Bindlish, R., Lee, H., Hong, S. (2012). Radar vegetation index for estimating the vegetation water content of rice and soybean. IEEE Geoscience and Remote Sensing Letters, 9(4), 564–568.

Mandal, D., Kumar, V., Ratha, D., Dey, S., Bhattacharya, A., Lopez-Sanchez, J.M., McNairn, H., Rao, Y.S. (2020). Dual polarimetric radar vegetation index for crop growth monitoring using Sentinel-1 SAR data. Remote Sensing of Environment, 247, 111954.

Periasamy, S. (2018). Significance of dual polarimetric synthetic aperture radar in biomass retrieval: An attempt on Sentinel-1. Remote Sensing of Environment, 217, 537–549.