Wheat Acreage Under a Two-Acquisition Budget and Fragmented Holdings

Select two Sentinel-2 dates: tillering and heading.

The dominant risk here is not spatial resolution but separability. A single green date cannot reliably separate wheat from mustard, berseem or young sugarcane, all of which are green in the same window. Two dates capture phenological trajectory - wheat greens up and senesces on a distinct curve - and also provide a second chance against cloud, which a single tasked acquisition does not.

Addressing the resolution objection: at 10 m, a 0.8 ha field is only about eight pixels, and with fragmented bunds a large share of pixels are mixtures. Two responses make this tolerable: (a) accept that the map is a stratification and area-estimation tool, not a parcel map; (b) correct area with a ground-based regression estimator rather than counting pixels. Pixel counting of a smeared, mixed-pixel classification systematically biases area, while a regression estimator calibrated on measured field areas corrects for it.

LISS-IV at 5.8 m would help fragmentation, but with one uncertain date it cannot deliver separability, and a slipped acquisition past heading destroys the deadline. Reject it, or keep it only as a contingency if one Sentinel-2 date is lost to cloud.

Design:

• Frame and strata. Stratify by the two-date image strata (wheat-likely, other-crop, non-crop) within each of the 40 blocks. Stratification by map class is what makes the sample usable for bias correction.
• Allocation. Proportional allocation alone leaves thin cells. Use Neyman-type allocation with a floor: guarantee at least 6 fields per block (240 fields) so every block has a local check, and assign the remaining 160 to strata with the highest expected variance, especially the wheat/other-crop confusion stratum.
• Two-stage option. If travel time dominates, sample villages first, then fields within villages, and account for clustering in the variance.

Field protocol: measure the full field polygon by GPS walk, record crop, sowing date and irrigation status, and photograph. Do not measure only the visually pure core - that biases field size upward.

Estimator: compute a regression estimator of wheat area on image-derived wheat area within strata, with a stratified standard error and 95% confidence interval. Reserve roughly 15% of points as an independent check on the correction itself, so the correction is not validated on the same data that produced it.

Do not apply the correction mechanically. The 8-12% figure describes a different instrument (enumerators visiting and asking) on a different population over different years. Your map-derived estimate has its own error structure, so stacking an enumeration correction on top risks double-counting.

What to do instead:

1. Keep instruments separate. Report the satellite-based area estimate with its confidence interval first, as the primary figure, and treat enumeration as a separate series.
2. Test the bias locally wherever both exist: blocks or years with paired enumeration and ground-measured areas let you estimate the bias in your own setting rather than importing 8-12%. Report what you find, even if it contradicts the assumed range.
3. If a spliced series is required for comparison with history, apply the local bias adjustment as a range (e.g., a low and high scenario) and label the resulting series as adjusted, not observed. Carry the adjustment uncertainty into the confidence interval rather than presenting a single corrected number.
4. Document the instrument change in the release note; users comparing this year with last year must know that part of any difference is method, not cropping.