Seeing sudden NDVI dips over irrigated pivots near Yuma this morning after the latest Sentinel-2 pass, even though the true color looks clear. In GEE I set s2cloudless at 40% and apply QA60 for cirrus; are you lowering that threshold or running a haze correction before NDVI to avoid false negatives? NAIP from 2023 shows consistent vigor, so I’m leaning toward misclassification.
I’d drop the ‘QA60 cirrus’ mask on S2_SR and instead join s2cloudless (Sentinel-2: Cloud Probability | Earth Engine Data Catalog | Google for Developers) at about 60% with SCL cloud+shadow classes; thin haze over bright desert often trips QA60, so also require B10 > 0.01 before masking. If a few pivots still dip, try a 3-day rolling median before NDVI — QA60 can be a smoke alarm for toast — do you have a sample tile/date to check?
Quick sanity check: instead of lowering the probability, I screen NDVI by AOT and B10 — mask pixels where AOT > 0.2 or B10 > 0.005, then compute NDVI; that stopped the sudden dips over Yuma pivots for me when ‘s2cloudless at 40%’ + QA60 was overzealous. If you’ve got a short tolerance, a 3‑day rolling median on NDVI cleans up the rest — can you afford a 3‑day window or do you need same‑day?
Seeing the same in Yuma lately; my workaround is to fill pixels flagged by QA60/s2cloudless using a qualityMosaic over the prior 1–2 clear S2_L2A images (pick lowest cloud prob or highest NDVI), then run NDVI on that composite. It smooths out the sudden dips from a ‘40%’ run, but if you need strictly single-date metrics that’s a caveat. Would a small 2–3 image blend be acceptable for your pivots?