Publish the loss method before the season, not after the storm

Every June, in the week the combines reach our wheat counties, I look at rainfall forecasts more often than at our system dashboards. Rain at harvest is the loss our agricultural book fears most. In June 2024 eight days of it fell across seventeen counties in Henan and Shandong, on wheat that was ready to cut. Within a fortnight 351,000 insured households had reported losses on 2.4 million mu, about 160,000 hectares.

Ten years ago an insurer facing that event would have sent surveyors into every village, measured a sample of fields with the village committee and settled county by county over two or three months. Today most Chinese crop insurers, Hesui among them, can also draw on optical and radar satellite imagery, drones flown by their own teams or by contractors, and models that turn pixels into estimates of damage. Used well, these tools make settlement faster and less arbitrary. I have spent most of my career building them, and I believe in them.

My argument is about timing. The industry tends to explain its remote-sensing method after a disaster, when payments are being calculated and every household has a reason to doubt them. By then the method looks like an excuse. It should be written down, agreed with the county agriculture and finance bureaus and explained in villages before the crop is planted, including a plain account of what the imagery cannot measure.

What the satellite cannot see

Imagery is very good at some things. It can map where wheat was planted, which fields are waterlogged, where the crop has lodged and how far harvest has progressed. Radar sees through the cloud that accompanies a week of rain. Combined with plot boundaries and household records, it can sort hundreds of thousands of plots into groups that probably suffered similar damage.

It is much less good at the loss that harvest rain actually causes. When wet wheat stands in the field or lies cut on the threshing floor, grain begins to sprout in the ear, mould develops, and the weight and quality of the crop fall. A buyer will pay far less for it, or nothing. Most of that damage is inside the grain. A field that has been cut and cleared looks, from orbit, exactly like a field whose harvest was perfect.

In 2024 our head of claims proposed paying each household on a plot-level loss rate estimated from imagery alone. Our analysts said we could do it in eighteen days. It was a serious proposal from a colleague who had watched farmers wait two months after the 2021 floods. I declined it because the model would have measured the visible part of the loss well and the larger part not at all. Households whose fields looked fine from above would have been underpaid, and they would have been right to complain.

The general lesson is that a model's confidence is not the same as its coverage. Our lodging classifier agreed with field measurement on about 85 per cent of plots in testing. That is a useful number for lodging. It says nothing about sprouting, and quoting it without that qualification would have misled everyone who heard it.

A method the village can check

My father was a technician at a township agricultural station in eastern Henan. Each June he estimated yields by laying a square wooden frame on sample plots and counting the ears inside it. When farmers disputed his figures he asked them to kneel beside him and count. The frame was crude, but the method was visible, and a visible method is easier to accept even when its answer disappoints.

We tried to keep that quality in 2024. Imagery divided the affected plots into strata. Within each stratum our survey teams measured sample plots, 1,840 in total, with a village representative and a technician from the county agriculture bureau present, and took grain samples for germination and weight tests. The loss rate measured on the samples set the payment for every household in that stratum. Each household was told its stratum, the rate and how to appeal through its village coordinator.

About 6,300 households appealed, 1.8 per cent of the total. Thirty-nine per cent of the appeals led to higher payments. Most came from two strata where a river terrace drained differently from the land around it, a difference our model had not distinguished. The appeals did not slow settlement much, and they improved the model for the following season. The median time from report to payment was 29 days, against 63 for agricultural claims after the 2021 floods.

The approach has costs. It took eleven days longer than paying on imagery alone. It required about 7,600 staff-days of survey work. It also opened the method to argument, and some branch managers would prefer that it were closed. I think the exchange is right. Farmers who can watch a method being applied can contest a figure within it, instead of contesting the insurer itself.

Three commitments before sowing

I would ask every agricultural insurer using remote sensing, and every board overseeing one, to make three commitments before each season.

The first is to publish the method by crop and peril. For each peril covered, the insurer should state which signals it will use, which losses those signals can measure and which they cannot, and how the gap will be filled. This should be agreed with the county agriculture and finance bureaus that co-fund the premium, and it should appear in the village notices that already list insured households.

The second is to fix the sampling design in advance: how many sample plots in each stratum, who witnesses the measurement, how grain is tested, and how strata will be redrawn if field results disagree with the imagery. A sampling design chosen after a disaster will always be suspected of serving the insurer, whatever its merits.

The third is to commit to a timetable and an appeal route, and to report afterwards on both, including the share of appeals that changed a payment. An appeal rate near zero is not good news. It usually means that households did not know how to appeal or did not believe it would help.

At Hesui this preparatory work now costs about RMB 9 million a year in survey time and printed materials. In 2026 we are extending it to maize and soybean in our three north-eastern provinces, where the perils are drought and early frost and the imagery is useful in different ways. Some colleagues regard the cost as an invitation to dispute. My view is that disputes come regardless. It is far better to have them about a method agreed in March than about one that appears to have been invented in June.

The subsidised programme already asks insurers to account for cover and claims household by household and to post them in each village. Publishing the method applies the same principle to the part of the process that technology has made least visible.