How the Farmland Atlas is measured

Every figure on the Farmland Atlas comes from a public federal or state file. This page names each file, the formula, the gate a number must pass before it is printed, and what is withheld and why.

Colours, ranks and next door

Each layer is coloured in up to seven classes, and each class holds about the same number of counties. The breaks between classes are the quantiles of every county with a figure. Each break is rounded, and the legend shows the rounded number the page actually classifies on. Where a seventh or more of the counties sit at zero (solar capacity, hot Julys), zero is a class of its own. A layer that runs through zero (a trend, a change) splits at zero: blue below, orange above, and grey for exactly zero. When the map is zoomed to a state with 20 or more counties, the classes are worked out again from that state's counties alone. The legend then says so, and the other states turn grey. Trends and changes keep the national classes, so an orange means the same thing everywhere. The two loss-ratio layers always carry a break at 1.00, where claims matched premium, so no colour mixes counties on both sides of it. A layer with nothing above zero (CRP against its own peak) shows blue for the fall and grey for a county at its peak.

A rank counts only counties with 10,000 acres or more in farms in the 2022 Census of Agriculture. Without that rule, a city such as Staten Island, with a few hundred acres of farmland, would rank first on land value. When counties share a value, the rank says they are tied. Past halfway, a rank is counted from the bottom. Rent is ranked only against rents from the same survey year, and a county last surveyed before the newest year gets no national rank.

Next door means the counties that share a stretch of boundary with this one, found from the county outlines. Counties that meet only at a corner are not counted. The next-door table compares rents only from the newest survey year; an older rent carries its year and is left out of the rank and the median. Across a state line, rent and yield come from each state's own survey.

Night heat

NOAA NCEI's nClimDiv county dataset gives, for every county, the monthly mean of each day's minimum temperature since 1895. I keep July, the silking month for corn across the Corn Belt, and August, the grain-fill month where the cited harm from warm nights is measured. The recent level is the mean of the last ten Julys (at least eight must be present). The trend is an ordinary least squares line through the Julys since 1976, reported in degrees per decade with a 95% confidence interval; a county gets a direction only when that interval, as printed, excludes zero, and no trend at all with fewer than forty of the years present. 1976 sits at the cold end of the record, which flatters a warming slope, so the slope through the whole record from 1895 is printed beside it in the arithmetic. Decade means run from the 1900s; a decade needs five Julys to print and ten to count as complete. A hot July is one whose average low was at or above 70 F: extension research puts the harm line for corn at nightly lows above 70 F, and a monthly average at that line means most nights were over it. North of Kansas that is rare, and the count says so; the decade means and the slider carry the Corn Belt story.

Water

Two official sources, neither produced by AGSIST. The irrigated share is irrigated harvested cropland over all harvested cropland from the 2022 Census of Agriculture; where NASS suppressed the irrigated figure it stays suppressed here, and where NASS published no irrigated row at all, which is what happens when no farm in the county reported any, the county shows none reported. The groundwater share is crop-irrigation withdrawals from groundwater over all fresh crop-irrigation withdrawals in the USGS county water-use estimates for 2015, the latest county release; USGS builds those from irrigated acreage and application rates, not meters, so they are estimates. Water withdrawn per irrigated acre comes from the same file and includes conveyance losses, so surface-water districts read high. The irrigated rent premium is separate: it is NASS's own survey of what tenants pay, and it needs at least six paired years before a trend is printed.

The loss record

USDA RMA's cause-of-loss files, one per crop year from 1989. Every crop indemnity dollar in the county is grouped by the peril that caused it and summed by period: 1989 to 1999, the 2000s, the 2010s, and 2020 on. Pasture, rangeland and forage rainfall-index policies are left out, because they pay on a rainfall index rather than a destroyed crop and in ranch counties they would swamp everything else; the dollars left out are printed. Area and index plans on crops (GRP, GRIP, ARPI, SCO, STAX) pay on a county outcome and RMA assigns them no peril; they are carried as peril not assigned rather than guessed into drought. Price decline stays in the denominator, so a revenue-protection county's 2014 and 2015 show up as price, not weather. Prevented planting sits inside excess moisture. Dollars are nominal and participation grew sharply after 1994 and 2000, so the shares by period are the comparable numbers, not the dollars. The heat-and-drought share by period is the number to watch. Irrigation failure is its own group: it is failure of the irrigation supply or equipment within the policy period, and aquifer decline is not an insured cause, so it is a narrow signal, not the Ogallala. The current crop year is partial until RMA closes it and is marked. The cause-of-loss files carry premium only for policies that had a loss, so no loss ratio is computed from them; that comes from the next layer.

Loss ratio and loss cost

The crop year RMA is still settling is left out of every ratio, count and share on the Atlas and shown apart as "so far": its premium is booked in full at sign-up and most of its claims are paid months later. Loss cost, printed as "claims paid per $100 of coverage", is indemnity per $100 of insured liability, a year. RMA prices premium from each county's own loss history, so loss ratios drift toward the same target everywhere; loss cost is what tells a risky county from a safe one. "Years over 1.00" is printed only for a county with ten or more crop years above the premium gate. RMA's Summary of Business county files (sobcov, one per crop year from 1989) carry every policy in the county whether it paid or not: liability, total premium, subsidy and indemnity by crop, plan and coverage level. Indemnity over that premium is the loss ratio as RMA defines it, and 1.00 means the county's book paid out exactly what it took in. It is summed over the whole record, over the last ten crop years, and by the same periods as the loss record; it prints only above $1,000,000 of premium, and a period prints only with at least five of its years present. Rainfall and Vegetation Index plans are left out here too, so the two RMA layers describe the same book, and the ratio of the two files' indemnity totals is printed as a check. The premium is the whole premium, farmer-paid plus federal subsidy, because that is what the ratio means to the program; a farmer's own premium is a fraction of it. Through 2012 RMA filed three big counties as two rating areas each (West Pottawattamie IA, West Otter Tail MN, West Polk MN); both halves are added onto the county. From 2024 RMA files Connecticut by planning region, which the 2019 boundaries on this map do not draw, so the old Connecticut counties' records end with crop year 2023.

Land value

Dry cash rent divided by this value is withheld where a tenth or more of cropland is irrigated (the 2022 census share, or USGS 2015 irrigated acres where the census suppressed it), because the census value then carries the irrigation and the ratio reads like a cap rate it is not. The Census of Agriculture asks every operator the market value of their land and buildings; NASS publishes the average per acre by county for 2012, 2017 and 2022. It is the only county land value on one method across every state, and it is what it is: operators' estimates, buildings included, five years apart. Beside it, the county's non-irrigated cropland cash rent for the census year divided by that value, printed as a gross percentage. The ratio is also withheld where cropland is under half of the county's land in farms (census cropland over owned plus rented acres, 2022): the value there is mostly pasture, woods and buildings, and before this gate the highest "yields" in the US were hill counties in Mississippi with 36 to 46% cropland. NASS county cash rent is the average of every cash-rented acre surveyed, family and long-standing leases included, and NASS publishes no standard error for it. No modelled value, ever, unless it ships with a backtest beside it; the one vendor that prints an accuracy for its model prints ±25%.

Tile drainage, tillage and cover crops

The Census of Agriculture asks every farm how many acres it drains with tile and with ditches, how many it farms no-till or with other conservation tillage, and how many it plants to cover crops. The Atlas reads those county totals from the 2017 and 2022 censuses through NASS Quick Stats and shows each as a share of the county's total cropland in the same census (idle, pastured and CRP cropland included, so the no-till and cover-crop shares run below a share of planted acres). The census counts drainage on all land in farms, not only cropland, so a share over 100% is withheld rather than cut to 100%. Where the census prints no row for a county and never prints a zero for that item anywhere, no row means no farm reported it, and the county shows none reported. No change in tile acres from 2017 to 2022 is shown. Iowa's statewide census total fell from 14.1 million to 13.0 million acres between the two censuses, and Ohio's and Indiana's fell too, though tile is almost never taken out; the county-to-county swings (a quarter of Iowa counties down more than 25%) measure who answered the census, not drainage put in or removed. Suppressed figures (D) stay suppressed. These are what farms reported to the census, not a map of the tile itself.

Read-with-care values and the recent yield level

A county's 2022 census land value is marked "read with care" when it fell 10% or more from 2017 while the counties next door (at least three with both years) rose a median 15% or more, or when it rose 100 points or more faster than that median. 47 counties fell and 21 jumped on the September 2026 data. The falls, nearly all of them home to a city: McLean, Champaign and Macon in Illinois, Delaware and Howard in Indiana, Mecklenburg in North Carolina. The jumps are the same effect upward: Shasta CA rose 274% while its neighbours rose a median 61%, Loving TX 775%. The census value is the operators' own estimate of land and buildings over all land in farms, so a county with many small residential operations can move on the mix of farms rather than on the price of cropland. The value is printed as NASS published it; the change, the rent ratio and the rank are not built on it.

The recent corn yield level is the county's last five published years, ending at its newest year, with the highest and lowest dropped; with four years published, the middle two are averaged. Dropping only the low year leaned high: a missing NASS year is often a thin, bad one. With fewer than four of the five published there is no recent level. It is printed beside the long median because a median over 18 years sits below what the ground grows now when the trend runs upward. Dry rent per bushel is the non-irrigated cash rent over that recent level, printed only where under 10% of harvested cropland is irrigated (the 2022 census share; where the census suppressed it, USGS 2015 irrigated acres over harvested acres, and USGS "under 1,000 irrigated acres" counts as low), because the county yield mixes irrigated and dryland fields.

CRP

FSA's county statistics: acres enrolled by year since 1986, and acres whose contracts expire by fiscal year. The share is CRP acres over CRP acres plus 2022 harvested cropland, the county's arable base. Acres expiring in the current and next two fiscal years are summed and printed against the enrolled total. The average rental payment is FSA's own average annual payment per acre under contract in the latest fiscal year, from the same workbook; it is not shown for a county with under 100 acres enrolled. It averages every contract in force, some signed 10 to 15 years ago, so it is not the rate FSA offers today. The change from peak is not shown for a county whose peak never reached 100 acres. Counties are joined on the FIPS code FSA prints beside each name, so its split reporting districts (West Pottawattamie, West Polk) land on their county, and old codes are carried to the current one (Shannon to Oglala Lakota, Dade to Miami-Dade, the former Virginia cities to their counties).

Drought

The U.S. Drought Monitor's county statistics, one record per weekly map since January 2000. A week counts when at least half the county's area was in D2 (severe) or worse; D3 (extreme) or worse is counted the same way. Weeks are summed by year, over the record, and over the last five full years; a year needs forty weekly maps to count as full. The Monitor is a map drawn by people from many indicators, not a rain gauge, and it is the map that USDA's disaster designations and livestock forage payments key off, so it is the drought that the program world recognises.

Rain, July highs, freeze dates, heat units and storms

Rain and July highs come from the same NOAA nClimDiv county files as night heat: monthly precipitation (climdiv-pcpncy) and the monthly mean of daily highs (climdiv-tmaxcy), each a county average of NOAA's 5 km grid. The growing season is April through September; a season counts only when all six months are in, so the current year waits for September. The normal is the 1991–2020 average and needs 27 of the 30 seasons. The last ten seasons are compared with it as a percent. The year-to-year swing is the standard deviation of the 30 normal seasons over their mean: how far a typical season lands from average. Two counties with the same normal can differ a great deal in how often a season comes up short; the count of seasons under 75% of normal is the downside measure. Seasons under 75% of normal are counted from 1991. Where the normal is under 6 inches (the Central Valley, the desert Southwest) no percent of normal and no swing is printed: 300% of one inch is three inches, not a wet year.

Freeze dates, frost-free days, growing degree days and hot days come from NOAA's nClimGrid daily county averages (EpiNOAA, on AWS open data), 1996 on, quality-controlled months only. The last spring freeze is the last day from March 1 to June 30 with a county-average low of 32 °F or lower; the first fall freeze is the first such day from August 1 to November 30. A year with no freeze in a window is recorded as before March 1 or after November 30, and a median that falls on one of those is printed that way rather than as a date. A year with no freeze inside a window ran longer than the window can measure; it is ranked above every measured year rather than dropped, and when the median lands next to such a year it is printed as "at least" (or "on or before", "on or after" for a date). When half or more of the years run past the windows, no length is printed. Dates are the day a county-average low reached 32 °F, roughly the day half the county froze: low ground and frost pockets freeze later in spring and earlier in fall, so a station freeze table reads later in spring and earlier in fall than these. Corn growing degree days use the 86/50 method (the high capped at 86 °F, both floored at 50 °F), May 1 to September 30. A season is kept only when every day of its window is present. A county average smooths frost pockets, river bottoms and hilltops; in mountain counties it says little about any one field.

Storms are NOAA Storm Events reports filed against a county, from 2010, the year the National Weather Service raised the severe hail line from 0.75 to 1 inch, so one rule covers every year. A hail day is a day with at least one report of 1 inch or more (and, separately, 2 inches or more); a wind day has a thunderstorm wind report of 50 knots (58 mph) or more; tornadoes are counted by track segment through the county. Wind filed against a forecast zone cannot be tied to a county and is left out. A report is where someone saw and reported the storm: counties with more people and more spotters report more, and a zero is no report, not no hail.

Livestock and pasture

The 2017 and 2022 Censuses of Agriculture through NASS Quick Stats: cattle and calves, beef cows, milk cows, cattle on feed, hogs and pigs, and permanent pasture and rangeland (not cropland, not woodland). Items are picked by their exact NASS wording; a changed wording stops the run rather than zeroing a layer. A (D) is a figure NASS suppressed to protect one operation and stays suppressed. Where the census never prints an explicit zero for an item, a county with no row had no farm reporting it and reads none reported. Cattle per 1,000 acres uses the county's 2022 land in farms (owned plus rented). The change from 2017 is shown only where the 2017 count was 50 head or more.

Solar and wind

EIA's Form 860 lists every generator of 1 MW or more at a grid-connected plant, with the plant's county and coordinates; the Operable tab gives what runs and the Proposed tab what developers have reported with a planned year. Solar, wind and battery nameplate MW are summed by county, with the year the first solar and first wind generator came online. A county with no plant in the file reads zero, not missing, because the file is a census. Proposed is not the interconnection queue; much of it is never built, and the queue itself is a later layer.

Wells and water rights

Two states publish an open register that this page has read and verified: Nebraska's registered wells (every well, with use, status, depth, static level at completion and completion date) and Kansas's WIMAS points of diversion (every point, with use, active flag, source of supply and the right's priority date). Nebraska counties show wells, irrigation wells not decommissioned, median depth, median static level and completions by decade; Kansas counties show points, active irrigation rights, the median priority year and the source split.

From September 2026 the Atlas also reads every other state register it could open, check and classify, where that month's pull succeeds (the map's Sources and vintages list names the registers in the current build): Minnesota's County Well Index, Michigan's Wellogic, Ohio DNR's well logs, Wisconsin's active high-capacity wells, the Oklahoma Water Resources Board's wells, California DWR's well completion reports (new wells), Arizona's Wells55 registry (not cancelled), Idaho DWR's wells, Oregon WRD's well logs (new wells) and Nevada DWR's driller reports (new wells). Each is filtered on the register's own irrigation code, which was read from the service's list of values before it was used. Every well is placed in a county by its own coordinates, not by the county name the register prints; a row with no point is counted and left out. A county shows irrigation wells on record, the median depth, the median water level when drilled (Oklahoma records first water instead, and says so), completions by decade and, where a pump or test rate is recorded, how many were rated at 100 gallons a minute or more (Michigan's register includes small lawn wells). Registers differ in what they count and how far back they go, so a county is compared with its own state. A state whose pull fails keeps its last good pull. Washington (no use field), Montana, Utah and New Mexico (water-right points, or a code table not yet verified), Texas and Colorado (a zip or a key), and the rest (no open irrigation code, or a host that did not answer) read "no open register read yet". Water-level change over the High Plains aquifer (USGS) is a layer to add once its file can be fetched from a runner.

Irrigated against dry rent

Irrigated cash rent divided by non-irrigated cash rent, same county, same survey year. The direction is called on that multiple, fitted through every paired year, and only when the 95% interval excludes zero. A dollar gap would widen with general rent inflation even when nothing about the water changed. Irrigated rent also pays for the well, pump and pivot, so the multiple is not the price of water alone.

Rent and yield

County cash rent and county corn yields are the same NASS series that drive Cash Rent by County. The yield trend here is fitted on every published year since 2008 and labelled with its window; that page's own trend uses the last fifteen years, which as of 2026 starts on the 2012 drought, and the two are different windows, not two answers. The rent change is nominal dollars over about ten years, using the closest published year at or before ten years back, because NASS published no county estimates for 2015 or 2018 and some counties skip years. The worst year is an observation, not a fit: the lowest published yield since 2008 as a share of the county's median over the same years.

Who farms it

Read from the same files as Rented vs Owner-Farmed and Foreign-Owned Land, so the pages cannot disagree. The rented share is land rented from others over land owned plus rented, as farms reported it to the Census of Agriculture; a county NASS suppressed has no figure, never an estimate. Foreign-held acres are USDA FSA's AFIDA disclosures summed by county for the latest year: self-reported, leaseholds of ten years or more included, and much of it forest. A county with no filing disclosed none, so it is coloured as zero rather than left blank; codes with no boundary on the map are left out.

What is never done

No number is interpolated from a neighbour. No layer is scaled, weighted or added to another. No owner name appears on this page. No land value is modelled. A county with nothing published shows a dash and the reason. The Atlas holds every county in the fifty states, and each layer covers only the counties its own source publishes: some cover all of them, some are still being fetched a state at a time, and the wells layer is two states by design. The county panel says which is which.

The night-heat yield test

The Atlas tests one claim: that counties whose July nights are warm, or warming, gain corn yield more slowly. Each county's yield trend is fitted on every published year since 2008. Four night-heat measures are compared with it, each across all counties and again within each state after the state mean is taken out. The within-state rows are the ones that count, because across the country the south is warmer and different in every other way too. Eight tests are run and all eight are shown; each chance figure is multiplied by eight. Neighbouring counties move together, so treat the chance figures as generous.

Sources

NOAA NCEI, nClimDiv county monthly minimum temperature (climdiv-tmincy) · USDA NASS, Census of Agriculture 2022 and Cash Rents Survey · USGS, Estimated Use of Water in the United States, county-level data for 2015 · USDA RMA, Summary of Business, cause of loss and state/county/crop with coverage level · USDA NASS, Census of Agriculture 2012, 2017, 2022, land and buildings value per acre · USDA FSA, CRP statistics by county · National Drought Mitigation Center, U.S. Drought Monitor county statistics · EIA, Form 860 · NOAA NCEI, nClimDiv county precipitation and maximum temperature (climdiv-pcpncy, climdiv-tmaxcy) · NOAA, nClimGrid-daily county averages (EpiNOAA) · NOAA NCEI, Storm Events Database · USDA NASS, Census of Agriculture 2017 and 2022, cattle, hogs and pasture · State well registers: Minnesota MDH/MGS, Michigan EGLE, Ohio DNR, Wisconsin DNR, Oklahoma OWRB, California DWR, Arizona DWR, Idaho DWR, Oregon WRD, Nevada DWR · Nebraska DNR, registered groundwater wells · Kansas DWR / KGS, WIMAS points of diversion · Night-heat harm line: Ohio State C.O.R.N. 2019-27 · USDA NASS, Census of Agriculture land in farms by tenure, 1997–2022 (via Rented vs Owner-Farmed) · USDA FSA, Agricultural Foreign Investment Disclosure Act holdings (via Foreign-Owned Land) · County boundaries: US Census cartographic boundary file, 2010 vintage.
Data files: atlas.json (every county, every layer) · per-county detail under /data/atlas/counties/<fips>.json.