Work

Which Plots Need a Field Visit? A Firm or Not-Firm Reading for Satellite Crop Yield Estimates

Year: 2026
Type: article
Abstract: Crop insurance in India is moving from manual crop-cutting experiments to yield estimates made from satellite data by AI programs. A field visit cannot be made to every plot, and today each plot gets one yield estimate from one pass, with nothing to show how firm that estimate is. This concept note describes a check that sits on top of any such AI program. The satellite data is processed once, and from that point the processed data is fixed. A small controller takes the three to four setting vectors that the maker has already validated. It moves the AI program's own non-learned settings a little below and a little above each value, sends the same processed data each time, and reads how much the yield estimates spread. Estimates that stay together are firm. A computer-decided shunting action then moves along the validated settings to one where the estimate is firm. Only if none is firm does that one plot go to the front of the field-check list, while all other plots carry on, without waiting for a person. The learned weights, the satellite data and the claim cutoffs are never changed. The note explains where this helps: field checks, claims, farmers, cloudy monsoon weeks, and flood and crop-damage maps. It also explains how the method differs from Bayesian networks and Monte Carlo dropout. Patent pending. Keywords crop yield, satellite, crop insurance, PMFBY, YES-TECH, remote sensing, firm reading, controller, shunting action, field verification
Language: English
Subfield: Ecology
Open Access status: green
OpenAlex ID: W7227370967
FWCI:
Cites:
Cited by:
Related to: 10