For years the pitch for agricultural remote sensing was resolution: finer pixels, sharper boundaries, better classification. Resolution helped. It was not the binding constraint. In most of the world's major cropping regions, the growing season coincides with the cloudiest months of the year, and an optical satellite cannot see through cloud at any resolution.
A monsoon-season crop can go thirty or forty days without a usable optical scene. That gap falls precisely across sowing and early vegetative growth — the window in which sown-area estimates and insurance sowing verification are needed.
What radar contributes
Synthetic aperture radar sees through cloud and works at night. It does not measure greenness; it measures surface roughness and moisture, and how those interact with the polarisation and geometry of the transmitted signal. A flooded, freshly puddled rice field returns a distinctively low signal. A developing canopy changes the ratio between co-polarised and cross-polarised returns in a way that tracks biomass accumulation.
This makes radar complementary rather than substitutable. Optical time series remain better for crop type discrimination in clear conditions and for stress detection through vegetation indices. Radar provides continuity through the gaps and a direct handle on moisture that optical cannot give.
The useful question is not optical or radar. It is what each sensor can answer in the week you need the answer.
Building a fused series
In practice, a workable plot-level series is built by harmonising both. Optical scenes are cloud-masked and gap-filled; radar series are speckle-filtered and terrain-corrected. Both are aggregated to the plot rather than the pixel, which suppresses noise and matches the unit that decisions are actually made about.
The fusion step benefits from being explicit about uncertainty. A sown-area estimate produced during a long cloud gap, resting mainly on radar, should carry a wider confidence interval than one produced in a clear fortnight. Publishing that uncertainty is what makes the estimate usable for a decision with financial consequences.
Ground truth is still the limiting factor
Neither sensor removes the need for field observation. Classification models are only as good as the labelled examples that trained them, and those examples need refreshing as varieties, practices and cropping calendars change. The most reliable programmes we have worked on treat a modest, well-designed ground campaign — a few thousand geotagged observations spread across agro-climatic zones — as a permanent annual line item rather than a one-off calibration.
The satellite record is free or cheap and improving. The scarce input is knowing what was actually growing in a given plot on a given date, and that still comes from someone standing in it.
