Detecting the maize silage harvest through the clouds

Every autumn, agribusiness runs into the same wall. The maize harvest — the moment supply becomes real — happens under the cloudiest skies of the year. And the tool most of the industry relies on to watch crops from space, optical satellite imagery, can't see through clouds.

NDVI, the vegetation index derived from optical satellites like Sentinel-2, is the workhorse of crop monitoring. It works beautifully under clear summer skies. But it's a camera: every cloudy pass leaves a gap. The usual fix — interpolating across those gaps — smooths the signal and lags reality. A field cleared in three days gets blurred into two weeks. So the harvest, the single event with the most commercial value, is detected late, or not at all.

That's the autumn blind spot. And it shows up precisely when timing is money — when a trader is sizing supply, when a cooperative is planning storage and logistics, when a commercial team is trying to read where the season really is.

The contrarian move: stop fighting the clouds

The instinct is to patch the optical data better. We took a different route: use a signal that ignores clouds entirely.

Radar does that. Sentinel-1's radar passes straight through cloud cover and works at night — the weather is simply not its problem. The catch is that radar doesn't natively give you NDVI. So we built a model, Crop-SAR, that reconstructs the vegetation index from the radar signal, one value per field, per week.

Two Sentinel sources go in: radar (Sentinel-1, VV and VH polarizations) as the all-weather backbone, and optical NDVI (Sentinel-2) as the reference the model learns to reproduce. For each field we resolve the radar's orbit geometry to keep a clean series, then align it week by week with the optical history. The model itself is a ConvTransformer — a spatial encoder reads the radar structure, a temporal transformer learns how the signal moves across the season, and a decoder turns that into NDVI.

The part that matters is how it's trained. We don't feed it clean data and hope. We deliberately degrade the inputs during training — masking optical observations, widening the largest cloud gaps, hiding the most recent weeks — so the model learns to hold the signal steady precisely when reality goes dark. Teaching it to tell “no crop” apart from “no data” is what makes the reconstruction trustworthy inside a gap, not just at its edges.

That yields two ways to use it: an extrapolation (nowcast) mode that estimates this week's NDVI when clouds block the optical view, and an interpolation mode that rebuilds past gaps — refined every week, so each estimate improves with time instead of freezing on a bad guess.

Same destination. Radar just gets there first.

It also beats the standard fix

Most crop monitoring fills cloud gaps with linear interpolation — a straight line drawn between the last clear image and the next. It's simple, and it fails exactly when it counts: during rapid growth or senescence, the crop doesn't move in a straight line.

Head to head on French fields, Crop-SAR cuts the share of large errors — NDVI off by more than 0.1 — from about 27% with linear interpolation to 16%, and tightens the overall error spread by about 30%. Fewer moments, in other words, where the filled-in curve quietly lies to you.

The test: all of Bavaria, one season

A model is only as good as its proof. So we ran Crop-SAR across every classified maize field in Bavaria for the 2025 campaign — 207,776 fields, 544,000 hectares — and checked it against the official numbers.

The rule was simple: a field counts as harvested when its NDVI drops durably below 0.40, the signature of the canopy coming off. We tracked that week by week, radar versus optical, and compared the outcome to Destatis, the German federal statistics office.

Cumulative maize harvest, Bavaria 2025. The radar signal crosses the official silage area (~406 kha) about a week ahead of optical.

Two things stand out.

First, the harvest splits cleanly into two waves — silage maize (about 406,000 ha) coming off in one block from mid-September to around 19 October, then grain maize (about 123,000 ha) following from late October into December. The curve tells the story of the season on its own.

Second, radar is faster. Its signal crosses the official silage area — 405,900 ha per Destatis — right around 19 October, inside the agronomic silage window and a full week ahead of the optical reference, which is slowed by exactly the cloud interpolation we set out to beat.

And the reconstruction holds up over the whole cycle: across the season, radar NDVI tracks the optical reference with a mean error of just 0.015 on a 0–1 scale — while staying continuous where optical went blind.

Why a week matters

A week doesn't sound like much until you're the one deciding on it. In-season, a week of lead time is the difference between acting on the field and reacting to the report. It's earlier visibility on how much silage is actually off, where grain still stands, and how fast the season is moving — while there's still time to do something about it.

That's the whole point of in-season foresight: not a prettier dashboard after the fact, but an earlier, field-level read on what's happening in the fields — robust enough to trust when the weather is working against you.

And the ground is getting less predictable, not more. Climate change is pushing harvest windows around from one year to the next, and growing practices are shifting with them — earlier or later sowings, a changing silage-versus-grain split, new hybrids and rotations. The more the calendar moves, the less last year's averages and end-of-season statistics tell you about this year. A signal that reads the field in-season, week by week, is how you keep up when the patterns won't sit still.

Bavaria was one region, one season. But Crop-SAR isn't tied to it: the same radar-based reconstruction runs on maize anywhere in the world. Wherever the crop grows and the autumn clouds roll in, the method holds.

Sources: Destatis (Federal Statistical Office of Germany), 2025; Bavarian State Research Center (LfL); Sentinel-1/2 satellite data. Hyperplan analysis, July 2026.