Why radar, not optical
The argument for the whole method, counted rather than asserted. Every Sentinel-2 scene over these six regencies, 2023–2025, sorted by whether cloud left anything usable behind.
The harvest wave
Staggered planting is what makes Indonesian rice supply continuous rather than a once-a-year flood. Scrub the crop calendar and watch it move across the kecamatan — this is the only view that shows it as motion.
Our number against theirs
Radar-derived harvested area beside BPS's own KSA figures, month by month and regency by regency, with the residuals charted rather than netted out.
BPS replaced a generation of eye-estimates with the KSA area-frame survey from the 2018 reference year. 2016 and 2017 are a moratorium with no data at all. The two series are not comparable, and any chart that runs one line through the hole is telling a lie about a trend that is a definitional change.
The calendar is moving
When the rain is late the crop is late, and the harvest lands in a different month from the one the procurement plan assumed. Monsoon onset against our own detected transplanting dates, with ENSO and the Indian Ocean Dipole underneath.
Six weeks early
The held-out 2025/26 season replayed as it would have been issued: our forecast at transplanting, and then the official figure arriving weeks later on the same timeline. The product, demonstrated rather than described.
Explore
Six regencies, chosen by where the benchmark is monthly rather than by where the rice is prettiest. Five of them have a 96-month KSA panel; Grobogan is the deliberate rainfed contrast and is validated annually.
Method & gates
Thresholds were fixed before any result was seen, and a gate that fails is published red with a diagnosis rather than quietly moved.
Speckle. SAR is coherent imaging, so a single-look pixel carries noise the size of the signal. Cells are multi-looked, and the average is taken in linear power and converted to dB afterwards — averaging decibels biases every cell low by an amount that depends on its own variance, which correlates with land cover and therefore looks like a real spatial pattern.
Orbits are never mixed. γ⁰ depends on incidence angle, so each relative orbit is a separate series until it has been normalised onto a common per-cell reference. Combining them raw produces a sawtooth that mimics phenology.
Flow, not stock. Harvested area is a flow: a field yielding three crops contributes its area three times. Paddy extent, harvested area and planted area are named and shown separately, everywhere.
Yield is not modelled. Production, where shown, is our harvested area multiplied by BPS's published productivity, with the arithmetic on the page.
These are literature values and they are not fitted. Fitting them on the KSA benchmark would guarantee agreement with the benchmark and destroy the gate that is supposed to test it. The sensitivity of the headline to each one is exported.
This panel used to say the gap was a resolution problem with a known fix: a smaller analysis cell. The review tested that and it does not hold. The controlled experiment is the one that settles it — take radar looks away from the densest regency's own record, leave the fields, the detector and every threshold untouched, and re-run. Widening the gap between looks from — to — days costs — of its crop cycles and — of its fields, and reproduces East Java's observed shortfall from sampling alone.
The same experiment runs forwards in the record. As Sentinel-1C and 1D came into service, —'s median gap fell — and the share of its rice fields we find rose — — no change to the crop, the plots or the code. Cell size is a real second-order effect and the case's damping term measures something real, but revisit binds first, and the two are separable: both are testable on these same six regencies, and the answer decides what a pilot should buy. That is the honest scope of the next step, and it is why this page ships red rather than tuned.
We had this instrument reviewed, adversarially.
An independent read of the same data against the published literature. It factors the shortfall against the official survey and finds that three-quarters of it is fields never found, not crops never counted — so the diagnosis this page used to give was the wrong way round. It then takes radar looks away from Karawang's own record and reproduces East Java's shortfall from revisit alone.