Appendix F — Analysis-ready SAR products
The four corrections described in SAR preprocessing can be inherited rather than performed. Two products apply them over the whole globe and distribute the result free of charge, and comparing them is instructive: both arrive at terrain-corrected \(\gamma^0\), both are defensible, and they answer the four questions differently at every stage (Table F.1).
Sentinel-1 monthly mosaics
The Copernicus Data Space Ecosystem publishes global Sentinel-1 mosaics as a Level-3 product, built by calibrating GRD inputs to \(\beta^0\), removing thermal noise, flattening to \(\gamma^0_T\) by integrating the illuminated facet areas over the Copernicus DEM, and applying Range-Doppler geometric correction against the same DEM. Each month is then reduced to a weighted mean in which observations at higher local resolution carry more weight [1] and radar-shadow observations are discarded. Because the compositing is what suppresses speckle, the product commits its user to a monthly cadence.
ALOS PALSAR/PALSAR-2 mosaics
Beyond Sentinel-1, another widely used SAR data source is the ALOS PALSAR/PALSAR-2 annual mosaic, a seamless global L-band SAR image created by mosaicking strips of PALSAR and PALSAR-2 imagery. Each strip is orthorectified and slope-corrected using the ALOS World 3D–30 m digital surface model, and polarization data can be converted to \(\gamma^0\) values. Because these mosaics are delivered as annual composites with terrain and radiometric corrections already applied, they require less user-side preprocessing than Sentinel-1 GRD products.
| Sentinel-1 monthly mosaics (CDSE) | PALSAR/PALSAR-2 yearly mosaic (JAXA) | |
|---|---|---|
| Thermal noise | Thermal noise removal applied. | Not documented as a separate step; radiometric balancing between adjacent paths is derived from their overlap regions, over land only. |
| Radiometric calibration | Calibrated to \(\beta^0\), then flattened to \(\gamma^0_T\) by integrating the illuminated facet areas over the DEM. | Radiometric slope correction, with \(\gamma^0\) recovered from the stored digital numbers as \(10\log_{10}(\mathrm{DN}^2) - 83.0\) dB. |
| Speckle reduction | Temporal averaging: a weighted monthly mean, weights set by each observation’s local resolution [1], radar-shadow observations discarded. | Power averaged over neighboring pixels; the annual composite is assembled from strips selected by visual inspection, favoring those with the least surface-moisture response. |
| Terrain correction | Radiometric correction as above, plus Range-Doppler geometric correction, both against the Copernicus DEM — 10 m over Europe, 30 m elsewhere. | Ortho-rectified and slope-corrected against the ALOS World 3D 30 m surface model (AW3D30). |
| Grid and resolution | 20 m for IW (VV, VH), 40 m for DH (HH, HV); 16-bit cloud-optimized GeoTIFF on a UTM grid. | 25 m; HH and HV (VV and VH over Japan); \(1^\circ\times1^\circ\) tiles in geographic coordinates, with layover and shadow flagged in a quality band. |
| Cadence and coverage | Monthly, global, 2020 onward. | Annual, global, 2007–2010 and 2015 onward; no mosaic exists for 2011–2014, between the two missions. |
What is inherited
The case for starting from either is the case for analysis-ready data generally: the work is done once, by people who know the sensor, and the result is consistent across space and time in a way an in-house pipeline rarely is. What is given up is granularity. The cadence is fixed, so the temporal decision discussed in Temporal distribution is made on the user’s behalf and individual acquisitions — with their orbit direction and incidence angle — are no longer addressable. The normalization is fixed too, and the two products differ in kind rather than in quality here: the Sentinel-1 mosaics integrate facet areas over the DEM, while the PALSAR mosaic applies a slope correction and a fixed calibration factor.