India's Crop Estimates Are Model Outputs Now. The Ground Data Underneath Them Is the Part to Watch.
Two numbers steer a large share of the money that moves through Indian agriculture. The pre-harvest production estimate shapes procurement plans, stock decisions and trade policy months before a crop is cut. The yield figure at the end of the season settles crop-insurance claims and, with them, part of the repayment capacity standing behind crop loans. Both numbers are now produced, in part, by models: satellite imagery and weather data go in, an estimate comes out. The argument worth having is no longer whether the model is clever. It is about what ground data trained it, and how much weight the official system actually gives its output.
The forecast machine: FASAL
Under the FASAL programme, the Mahalanobis National Crop Forecast Centre generates satellite-based pre-harvest estimates for 11 major crops — paddy, wheat, jute, cotton, sugarcane, soybean, tur, gram, mustard, lentil and rabi sorghum. The coverage runs to 557 districts in 20 states.Uttar Pradesh alone accounts for 75 of those districts and Madhya Pradesh for 52; Kerala accounts for two. Same programme, very different ground operations underneath it. Hold that thought.
| State | Districts covered under FASAL |
|---|---|
| Uttar Pradesh | 75 |
| Madhya Pradesh | 52 |
| Bihar | 38 |
| Chhattisgarh | 33 |
| Telangana | 32 |
| Rajasthan | 32 |
| Odisha | 30 |
| Gujarat | 29 |
| Maharashtra | 29 |
| All 20 states | 557 |
Source: Ministry of Agriculture & Farmers Welfare, Lok Sabha reply, 2 December 2025. Chart: Farmioc.
The method stacks three layers. Multispectral and microwave satellite data map where each crop is growing — the microwave layer matters in the kharif, when cloud cover blinds optical sensors for weeks at a stretch. Satellite indices and weather data then feed weather-based and remote sensing models that estimate yield. Underneath both sits the unglamorous layer: state agriculture departments collect the ground data used to train the crop-mapping models, while the agriculture ministry and the Centre work with ISRO on yield estimates from a semi-physical model. A FASAL number is therefore not a satellite product. It is a satellite product resting on field data that somebody had to walk out and collect.
The claims machine: YES-TECH
The second system is younger, and it carries money more directly. Under the crop insurance scheme, YES-TECH has, since Kharif 2023, estimated yield at gram panchayat level for paddy and wheat, with a mandatory 30% weightage given to the technology-derived yield in the figure that settles claims. Soybean was added from Kharif 2024. The system runs in 10 states, claim payouts in all seven states that implemented it in its first season were made on a YES-TECH basis, and the stated direction is a gradual migration toward remote-sensing-based yield estimation.
The plumbing around the model matters as much as the model. Yield and crop-cutting data are captured through the CCE-Agri App and uploaded directly to the National Crop Insurance Portal; insurers can witness crop-cutting experiments as they happen; state land records are being integrated with the portal. Space-based inputs also decide where crop-cutting experiments are conducted and are used for addressing yield and area discrepancies. The ground experiment, in other words, has not been abolished. It has been repositioned — from being the estimate, to being the check on the estimate.
Where a model number can mislead
Start with the 30%. If the technology-derived yield were treated as settled truth, its weight in claim settlement would be 100%, not a mandated minority share. The blend is the official system pricing its own model error into the design. A commercial user who takes a model yield at face value is being less careful than the scheme that built it.
Training data is the second soft spot. FASAL's district counts say where the model runs; they say nothing about how much recent field data trained it in that district. Ground collection sits with state departments of differing staff strength and survey discipline, so a model output for a district with fresh, dense ground truth and one for a district running on older training data arrive wearing the same uniform. They should not be priced the same.
Timing is the third. A pre-harvest estimate is built to be revised — it is a reading taken before the weather has finished voting. Procurement cover, lending limits or trade positions set on a season's first estimate are set on the version carrying the least information. The revision pattern of a forecasting system, not its accuracy in one good year, is what a risk desk should study.
Aggregation is the fourth, and the one that bites lenders. Insurance settles at gram panchayat level; most commercial dashboards still read at district level. A district average can look sound while a cluster of panchayats inside it fails, and it is the cluster, not the average, that produces the claims and the overdue loans.
Four questions to ask of any crop number
Which system produced it? A FASAL forecast, a YES-TECH yield and a crop-cutting average are three different instruments built for three different decisions. Quoting one where another belongs is the most common error in agri data rooms.
What weight did the model carry? In claim settlement the answer is published: 30%. In a vendor dashboard it is usually unstated. Ask, and treat silence as an answer.
What trained it, and when? Ask for the ground-data vintage behind the estimate for your districts, not for the country. A national coverage statistic is not district evidence.
At what unit, and what is the spread inside it? A district figure quoted without the panchayat-level spread hides exactly the tail a lender or an insurer is exposed to.
Where Farmioc fits
Farmioc DataStore is built on the same pairing this post describes — satellite and ground data organised into yield, harvest and supply-demand outlooks by crop and region, for users in the financial sector and agribusiness. We would rather be judged by the four questions above than exempted from them. A model output that states its ground data, its weight and its unit is a number a credit committee can work with. A model output that states none of them is a chart.
India's crop statistics are not going back to clipboards; every official document in this area points one way, toward more model weight. The institutions that come out ahead will be the ones that learned to read a model output the way they learned to read a mandi price — as a signal with a source, a method, and a known way of being wrong.
Farmioc is India's agriculture data analytics and agri trading company. Explore DataStore for yield and harvest analytics, or PriceHub for commodity market intelligence across 200+ crops.