Indonesia knows how much rice was harvested because people go into the fields and check. It is careful work and it reports about two months after the fact. Meanwhile a radar satellite passes over the same fields every 6 days, in daylight or dark, through monsoon cloud that blinds an ordinary camera.
What this gives you: the harvest timing. Across 15 regency-years the satellite's monthly rhythm tracks the official one closely — a correlation of 0.77 — once you allow for its being consistently about 1 month early. If you want to know when the rice is coming in, this works, and it works weeks before the survey reports.
What it does not give you: the number of hectares. On the cells it flags it agrees with an independent rice map 83% of the time, but it only finds 51% of the fields that map calls rice. Month to month, its area figures barely track the official ones at all — 0.11 before you shift for the lag. Do not use this to count a harvest.
So the finding is a redirection. We set out to measure how much and can only tell you when. The date is the more useful product anyway: it is what a mill, a buyer, or a food-reserve office needs in advance, and it is the part the official survey is slowest to deliver.
Why radar, and why it half works
A flooded paddy is nearly flat water, and radar bounces off it away from the satellite, so the field reads dark. As the crop grows the stalks scatter the signal back and the field brightens. Flood, grow, harvest, flood again — each cycle leaves a dip and a rise. Count the dips and you have counted harvests.
That is the theory, and it is why the timing comes out well: the shape of the cycle is robust even when its exact size is not. Counting fields is harder, because a field only shows up if the satellite happened to look while it was flooded.
What breaks it, tested rather than argued
The published explanation for missing fields was that the measurement grid is too coarse. That turned out to be wrong, and the way it was shown is the part worth reading.
In Karawang, 96,603 known rice fields were re-examined with nothing changed — same fields, same method, same thresholds — except that satellite passes were thrown away to simulate a satellite that looks less often.
| A pass every… | Harvests found | Share of the full record |
|---|---|---|
| 6 days | 284,388 | 100% |
| 12 days | 22,635 | 8% |
| 18 days | 1,511 | 1% |
| 24 days | 233 | 0% |
Going from a pass every 6 days to one every 12 loses 92% of the harvests. The limit is not how finely the satellite sees. It is how often it looks — and a flooded field can drain in under a fortnight, so a satellite that visits fortnightly will simply miss it.
The check that had to come first
Before blaming the satellite for disagreeing with the official count, the official count had to be cleared. Indonesia changed how it measures rice area in 2018, and the figure moved by -19.4% — a big step. But that change is four years before this satellite record even begins, on 2022-07-01, so it cannot explain the gap. Ruling that out mattered more than any adjustment to the method.
How does this compare with other radar rice work?
Two answers, and they point in opposite directions. On agreeing with the annual official total, this case is ordinary: after calibration it tracks the yearly figure at 0.82 on a 0-to-1 scale, against 0.78 for a 2017 study using the same kind of radar over Myanmar, and 0.85 for the Indonesian portion of the 2025 product this case uses as its own rice map. Nothing unusual there.
On finding the fields, it is not ordinary at all. This record locates 47% of the rice and accounts for 35% of the official harvested area. The 2014 programme that mapped rice across six Asian countries reported classification accuracy of 85–95%; the 2025 product reports 98% overall accuracy. No published radar rice study reports missing half the crop.
The difference is the one named just above: how often the satellite looks. The 2017 comparison was built from 632 images over one country; this record has 1235 across three provinces. A field that floods and drains between two passes was never seen. That also explains why the annual figure survives and the monthly one does not — month by month, the agreement falls to 0.06, which is nothing. Averaging a year hides the misses; asking for a month exposes them.
What a government or an NGO can actually do with this
Three things follow, and the value is in the first one only.
- Buy time, not a replacement figure. The whole benefit is that the signal arrives while there is still a decision to make — imports, stock releases, procurement — rather than after the harvest it describes. The peak of the season is located to within about 1 month. An earlier estimate that is roughly right beats an exact one that arrives too late to act on, and those are different products.
- Distrust it selectively, using its own residuals. Calibration was done per regency and month rather than by one national factor, and the difference is not marginal: the worst regency error falls from 74.65% to 5.10%, and the average across regencies from 72.68% to 6.62%. Errors that are located can be discounted where they live instead of discrediting the whole map.
- Let a divergence trigger a check, not a headline. Timing bias is published per regency, so it is knowable in advance which way a place runs: Bojonegoro reads -5.24 weeks against the survey and Indramayu 4.97. An unexpected gap between radar and the official count means one of the two is wrong in a way that matters — that is a reason to send someone to look, and publishing the gap as a finding skips the step that would tell you which.
What you must not conclude. Nothing here says the official survey is wrong, and nothing here replaces it. Finding 51% of the fields is not a rival estimate of the harvest — it is a partial view with a known bias, and treating it as a production figure would be a serious error. The claim is narrower and firmer: the timing is recoverable early, and the area is not.
Could this run in daily operations?
FEASIBILITY ONLY — none
Every pipeline here is a batch backfill: it fetches a season or an archive, not a live feed. Anything operational means rebuilding the ingest for near-real-time arrival, whatever the verdict below says.
What stops it. All five checks failed, including both blocking ones. The decisive one is the temporal hold-out: does this work on a season it has not seen. Continuous ingest would ask that question weekly and in public. The reconciliation that looks strongest appears only after fitting to the same official statistics it is meant to predict.
What it would need. More than a pipeline. The constraint is the model, not the ingest.
If you want the detail
The methods, every failed check and the things still uncertain are on the technical article, and you can move the controls yourself on the dashboard. Every figure above is read from the same record as those, so they cannot drift apart.
Words on this page, in plain language 5 terms
- radar
- Instead of photographing reflected sunlight, the satellite sends its own microwave pulse and listens for the echo. It works through cloud and at night; it does not see colour.
- a look
- One pass of the satellite over a field. More looks means the crop's cycle is sampled more often; too few and a short stage — flooding, heading — can happen entirely between two looks and never be seen.
- hold-out
- Data deliberately kept away from the model while it learns, then used to test it. Without one, a model is graded on the answers it was shown.
- calibration
- Adjusting a model's output so it lines up with a trusted measurement. It can genuinely fix a scale error — or quietly force agreement and teach you nothing, which is why the before-and-after is always published here.
- regency
- The English name for a kabupaten: the administrative level below a province, and the level Indonesia publishes most official statistics at. A regency is not a city — cities are counted separately here, and on several of these cases they behave differently enough to be reported on their own.
Data vintage 2026-08-30. Written to be read without a background in remote sensing or statistics; nothing was rounded to make a point.
Found something wrong on this page? Report a correction — it opens a pre-filled issue — or email taufik.adi@openstudy.id. Corrections are credited by name in the errata, and one that changes a finding says so on the page.