What grows at three Tennessee farms, and why not#

2026-10-08. The first working run of the plan in plant-conditions.md: the plant list turned into data, three sites profiled, the plants matched against them, and the bee forage calendar laid over the months. The sites are the three SCAN stations whose soil and sensors this project already knows best — McAllister Farm, Allen Farms and Eastview Farm, south-central Tennessee.

The three sites#

• NASA POWER is coarse. McAllister and Allen Farms, 27 km apart, share a grid cell and so an identical climate; only their soils differ. The grid also smooths extremes: it puts all three sites in hardiness zone 8a where the USDA's station-based map puts this area around 7, and its frost dates are optimistic. The profile says so on every run, and the match keeps a margin for it.
• Erosion is the loudest number. Bare, these silt loams would lose 11 to 37 times what the soil survey calls tolerable — Allen Farms most, on its 9% eroded slope. The practical meaning: on these soils, ground left bare over winter is the costly mistake, and cover crops, perennial forage and buffers are not optional.
• Heat stress for dairy cattle falls on a quarter to nearly half the year, depending on which estimate — the case for the heat-tolerant breeds on the animal list.

The match#

At McAllister Farm (the others differ only where topsoil pH does):

Ruled out: every tropical crop on the list — citrus, mango, papaya, pineapple, coffee, coconut, sugarcane, tamarind, breadfruit, and the tropical forage and agroforestry trees (napier and guinea grass, vetiver, leucaena, gliricidia) — by the frost-free season, and in most cases the winter as well; and a few coastal or deep-south species (radiata pine, bitter panicum, saltmeadow cordgrass).

Suits, or suits with cautions (most often rainfall above PLANTS' range, as for maize, blueberry and chestnut), by category, among others: wheat, oats, barley, rye, sorghum, maize; soybean, cowpea; alfalfa, red clover, vetches, birdsfoot trefoil, sunn hemp; tall fescue, orchardgrass, switchgrass and the native prairie grasses; pecan, black walnut, Chinese chestnut, American hazelnut; pear, persimmon, pawpaw, muscadine, blueberry, elderberry; white and red oak, black cherry, shortleaf and loblolly pine, sugar maple; Fraser fir and the other Christmas trees; all fourteen streambank plantings, the windbreak shrubs, the slope grasses; the bee trees — tulip poplar, basswood, sourwood, red maple, redbud.

Fixable: topsoil pH 6.2 is a little low for barley, red clover, sweet clover, partridge pea and big bluestem (lime), and a little high for shortleaf pine and bald cypress (sulfur). At Eastview Farm, pH 6.5, most of these pass.

What this run found wrong with the data#

USDA PLANTS is a conservation-planting database. It covers the plants on our list that conservation programs plant — erosion plantings (35 of 38 with characteristics), forage grasses (16 of 18), forage legumes (11 of 12), timber (24 of 29) — and almost none of the food: vegetables 5 of 59, spices and herbs none of 34, cut flowers 1 of 23, tropical fruit 11 of 32, pulses 6 of 23. Apple and peach have entries and no characteristics. So this first version answers conservation, forage and timber questions, and is mostly silent on food crops. Filling that is the case for ECOCROP, which FAO's data catalog lists under CC BY 4.0 (entry; FAO is being asked to confirm, and for a bulk file), or for another open crop source.

Its limits are guidance, not limits. PLANTS lists white clover as needing 250 frost-free days and dying below 22 °F, and bermudagrass, crimson clover and annual ryegrass much the same — all staples of Tennessee pasture and cover cropping, at 207 days and 11 °F. Crimson clover and annual ryegrass are winter annuals, for which frost-free days do not apply at all. Buckwheat, an eight-week summer annual, is listed at 365 frost-free days; aronia, hardy far colder, at 24 °F; mango, killed by a few degrees of frost, at 17 °F. The first version of the matcher took the values as limits and ruled out half of Tennessee's pastures. It now rules a plant out only on an overwhelming shortfall — under 75% of the season, or more than 15 °F below the minimum — which still excludes every tropical crop; anything less is a caution that quotes what PLANTS lists, and an implausible season on an annual is named as implausible and not used.

Bees: the bloom calendar#

Of the plants that suit the sites, 143 have a PLANTS bloom period. Laid over the months (the northern-hemisphere reading of PLANTS' seasons):

Bees can fly here from February to December (a mean daily maximum of 10 °C or more). September to November show nothing recorded in flower — but that is at least partly missing data: asters have no PLANTS characteristics, and PLANTS gives goldenrod "late summer". The true autumn gap is narrower than this; that it exists at all is the familiar pattern, and goldenrod, asters and late-sown buckwheat are the usual answers.

Food crops, from a crop model#

USDA PLANTS knows almost nothing about food crops, so a second source was brought in: the DSSAT crop models (BSD-3-Clause), whose parameter files give each crop's cardinal temperatures for development, its response to day length, and — per cultivar — how long each phase takes. dssat_crops.py ports DSSAT's own response curves and the CROPGRO, CERES-Maize, CERES-Sorghum, CERES-Millet, CERES-Rice and Cropsim-CERES wheat and barley phenology, and runs every cultivar of 28 crops through the site's thirty years of weather and day length: planted a week after each year's last frost (four weeks before it for cool-season crops, two weeks before the first fall frost for the winter cereals), does it reach physiological maturity before the first fall frost — or, for the winter cereals, survive the winter and mature? crop-parameters-dssat.csv carries the parameters, with DSSAT's notice.

At McAllister Farm:

Wheat and barley carry what the other models do not: vernalization and winter hardiness. Development runs on thermal time scaled by a vernalization factor (cold days counted through a −5/0/7/15 °C response until the cultivar's requirement, P1V, is met) and a long-day factor; and a kill temperature hardens from −6 °C toward the ecotype's limit (−20 °C for the US soft red winter ecotype) through cool autumn days and softens on warm ones. Newton, a US winter wheat, sown two weeks before the median first frost (October 20 here) flowers about April 30 and matures in 230 days — early June, when Tennessee wheat is cut, which is what leaves room for the double-crop soybean sown into its stubble. NASA POWER carries no snow, so the crown is never insulated; that errs toward winter kill.

Three rules go beyond DSSAT, each because a result was wrong without it. DSSAT stops counting cold once the spike grows, so a spring wheat sown in October "flowered" in late February and matured; now a daily minimum of −2.2 °C or below from boot on ends the season (K-State's C-646 puts injury at 28 °F at boot, 30 °F at heading), and the fall-sown spring wheats fail most years, as they would. A sowing swept through summer "matured" too, so a day over 31 °C while it flowers ends the season (above 31 °C pollination fails: Porter & Gawith 1999), and a season longer than nine months does not count — phenology alone lets a June-sown winter wheat wait out the summer for its cold. Five of the eight wheats, the winter ones, mature from a fall sowing 8 years in 10. Barley's three cultivars were calibrated in Syria and England; one matures, and its winter types flower in mid-March here, earlier than Tennessee barley heads. A local calibration would fix barley.

Rice runs DSSAT's direct-seeded path, as Mid-South rice is drilled rather than transplanted: germination paced by the sowing day's warmth, a juvenile phase, a short-day phase that also waits out a spell of cold nights before panicle initiation, then heading and grain fill. Lemont, a southern long-grain, sown May 1 heads in about 98 days and matures in about 122, the end of August, as the region's rice does. The season is no obstacle at McAllister — 42 of 55 cultivars mature 8 years in 10 — but rice there is a question of water, a flooded field on silt loam, which no phenology model asks.

Rye has Cropsim-CERES genotype files of its own (RYCER048: a test cultivar and the Italian Palazzo), in the wheat and barley format, so the same port runs it: both mature from a fall sowing at McAllister 9 years in 10 or better, in 217 and 251 days. Two cultivars is thin, and neither is a Southern cover-crop rye; it says the season fits, not which rye to buy. Oats have no DSSAT model at the pinned commit, and are left unchecked rather than run as barley.

Transplanted rice is ported too, as DSSAT runs a nursery crop (RI_Transpl_p.for): the seedlings develop in the nursery on the same weather, transplanting at 25 days lengthens the juvenile phase by the shock (P1 + 25 + 0.4 × the thermal time since emergence, at most 1.75 P1), and the crop goes on filling its tillers' grain to P5 + 40 degree-days where a drilled crop stops at 90% of P5. Sown on the same date, 35 of 55 cultivars mature 8 years in 10 against 42 drilled, and take a median 174 days from nursery sowing against 166: the shock and the longer fill cost a week. What transplanting buys in practice — a nursery sown weeks earlier under cover, so the field is occupied for less of the season — is not modelled, because the nursery here sits in the open on the field's own weather.

When to sow. With --windows the tool sows every week of the year, with the fastest, a middle and the slowest of the cultivars that mature here 8 years in 10 (by simulated days, not by the size of their requirement, which ignores day length and once picked a short-day sorghum that never flowers in summer), and ends each season at the first freeze below that crop's own damage threshold (DSSAT's FREEZ1: −2.2 °C for tomato and beans, −6 for cabbage, −9 for faba bean, −29 for lentil). Sowing windows that mature 8 years in 10 at McAllister Farm: tomato Apr 9–Jul 2; pepper Apr 9–May 28; maize Apr 9–Jul 30; sorghum Apr 9–Jul 9; pearl millet Apr 9–Jul 30; soybean Apr 9–Aug 20 (the late end is the double crop after wheat); peanut Apr 9–Jul 2; dry bean Apr 9–Aug 13; cabbage Mar 12–Aug 6, spring and fall; chickpea Feb 19–Aug 27; canola Mar 5–Jul 30. The dates are sowing, not transplanting. Lentil and sunflower come out "all year" because DSSAT gives them damage thresholds of −29 and −15 °C; lentil's is a modern winter-hardy calibration, sunflower's is generous for young plants.

Sixteen food crops on the list move from no data to a verdict, and pigeon pea and velvet bean are ruled out for the right reason — day length, which PLANTS never checked. The matcher takes these with --crop-seasons and lets the crop model outrank PLANTS on the season, since it runs thirty real seasons rather than comparing one number. Simplified, and said so in the tool: hours are a sine between the day's minimum and maximum, soil temperature is taken as the air's, and water and nutrient stress are not simulated.

One command, in any language#

modules/advise/advise.py LAT LON --lang en,es --windows runs all of the above for a point and writes a report a farmer can read — the site, food crops with sowing windows, fruit trees by winter chill with bloom dates and the frost after them, what suits, what needs lime, sulfur or water, what does not suit and why, what is refused as invasive, bees, livestock heat — plus the data behind it. Its words come from the language packs, so the same report renders in Spanish (dates, decimal commas and the reasons included) and in any language a pack is added for. Plant names come from AGROVOC and Wikidata where they have one — 198 of 477 in Spanish, 272 in French — and stay English beside the scientific name where they do not (plant_names.py). Example: advise-mcallister.en.md, advise-mcallister.es.md.

Invasive listings by state. USDA PLANTS lists invasive and noxious status by state for 13 jurisdictions — Tennessee is not one of them, Alabama and North Carolina are — plus the federal noxious weed list. A plant listed in the site's own state, or federally, is refused; one listed elsewhere carries a caution naming the states. So the report refuses water spinach (a federal noxious weed) and kudzu (on the project's do-not-plant list), and flags paulownia as listed in Alabama, Connecticut, North Carolina and Wisconsin. The Tennessee Invasive Plant Council's list would close the local gap; its terms are to check.

The 2050s#

advise.py --warm 1.0 --future-label 2050s runs everything twice, the second time on the same thirty years of weather with every day 1.0 °C warmer (the delta method), and writes what changes. (The example reports, advise-mcallister-2050s.en.md and advise-mcallister-2050s.es.md, now come from the CMIP6 run described below; this uniform shift's numbers are kept in the comparison there.)

Why +1.0 °C. The Berkeley Earth series for Tennessee in ~/dev/climate-analysis puts the 2050 level at +1.89 °C against 1951–1980 and 2060 at +2.18 °C. The 2011–2020 mean sits at +1.149 °C and the trend runs at 0.264 °C a decade, so the 1991–2020 baseline this record represents (centred near 2005) is about +0.89; 1.89 − 0.89 ≈ +1.0 °C for the 2050s (and ≈ +1.3 °C for the 2060s). The figure comes from that repository's data and method; none of its code is imported.

What moves at McAllister. Frost-free days 8 years in 10 go from 207 to 213, growing degree days from 2774 to 2985, days over 35 °C from 13 to 21, and the dairy heat-stress upper estimate from 172 to 183 days; the hardiness zone stays 8a (1.8 °F is under half a zone). Every modelled crop's sowing window opens about two weeks earlier (maize and soybean from April 9 to March 26), and many close a week later. The longer season lets more sorghum (51 → 52 of 55) and pearl millet (48 → 53 of 60) cultivars mature; no crop loses one. The plant matcher moves one plant (saltmeadow cordgrass now fits). Chill barely moves (1048 to 1023 hours on average), because warming turns hours below freezing into chill hours about as fast as it lifts others above 7.2 °C — see below.

What the delta method cannot say. It moves today's weather and keeps its variability, so it says nothing about wetter springs, longer droughts or stronger storms. It is uniform over the day and the year, while the Mid-South's record shows nights warming faster than days (the "warming hole" in summer maxima): with warmer nights and the same days, chill would fall faster and heat stress at night rise faster than this shows. Rain is left unchanged because the record gives no defensible direction for it here.

The 2050s from CMIP6 models (since 2026-10-09)#

A uniform +1.0 °C keeps the level honest and loses the shape. advise.py --cmip6 ssp245 replaces it (cmip6.py): five CMIP6 models from NASA's NEX-GDDP-CMIP6 (CC0; ACCESS-CM2, GFDL-ESM4, MPI-ESM1-2-HR, MRI-ESM2-0, NorESM2-MM), read for McAllister Farm's own grid cell, each model's 2041–2070 under SSP2-4.5 against its own 1991–2014, the median taken month by month and applied to the site's thirty years — days and nights separately.

Larger than +1 °C everywhere, and largest from October to January — the chill season. So the answers move further than the uniform shift said, and in a different place:

The uniform shift had said warming hardly touches chill here, because a warmer winter turns hours below freezing into chill hours nearly as fast as it lifts others above 7.2 °C. With the winter months warming by 2–2.5 °C rather than 1, that offset no longer holds: high-chill fruit — most apples, the later cherries — loses its chill. Bloom comes about a week earlier and frost after it holds steady or falls. Five models and one scenario, applied by the delta method, are a central estimate, not a range; the per-model numbers are written beside each report (cmip6-deltas.json) for anyone who wants the spread.

Stone fruit by winter chill#

A peach, plum or cherry that does not get its winter chill flowers late and unevenly, and chill was the one condition for fruit trees with no open source behind it. FAO ECOCROP is non-commercial, and CITA Aragón's 143-cultivar dataset is CC BY-NC-ND. The source that works is the review by Fadón, Herrera, Guerrero, Guerra and Rodrigo (2020, Agronomy 10:409, CC BY 4.0), which tabulates every published requirement it found: chill-requirements.csv, 662 requirements for 531 cultivars of almond, European and Japanese apricot, peach, nectarine, flat peach, European and Japanese plum, and sweet and sour cherry. Each row carries its model (chill hours, Utah units, chill portions, and growing degree hours for heat), method, place, and the original study.

Re-keying it was the work: the tables centre each cultivar's name across its rows, which text extraction cannot follow (a four-row cultivar's name lands beside its third row). chill_requirements.py reads the PDF's word positions instead and splits each page's rows so every name sits nearest its rows' midpoint. The result reproduces the review's own cultivar count for every crop. One cherry is extra because the review's Table 8 itself is one short. Every one of the 658 rows a text extraction shows is present, plus 4 rows the PDF prints with no location.

Counting chill. A chill hour is an hour between 0 and 7.2 °C — the review's definition, and chillR's — so hours below freezing do not count, though the site profile counted them at first. That correction cut McAllister's figure by a third (1648 to 1048 hours a winter, November to February). Chill is counted over full winters only (the record's half-winters at each end had pulled the mean down a few percent), and for fruit through the end of March, since a tree goes on gathering chill until it has enough: 1090 hours 8 years in 10. The advisor holds each cultivar's requirement to the top of its published range, and a cultivar measured twice to the higher figure. At McAllister that covers every almond, nectarine, flat peach and Japanese plum, 115 of 126 peaches, 26 of 41 European apricots, 25 of 40 cherries and 4 of 9 European plums: the high-chill end of the list falls short. Warming hardly changes it (1047 in the 2050s; 108 peaches, 24 apricots), because a warmer winter here turns hours below freezing into chill hours nearly as fast as it lifts others above 7.2 °C — a mild-winter site would lose chill faster. The requirements were measured in a dozen countries — Spain for 43% of rows, then China, the US and Korea — by methods that do not always agree, so the report calls them a guide.

Bloom and frost. Late frost at bloom is the commoner loss here, and 145 of the requirements come with a heat requirement too, so bloom can be simulated: fruit_bloom.py counts chill from November 1 until the cultivar's requirement is met, then growing degree hours (Anderson et al. 1986: base 4, optimum 25, ceiling 36 °C) until its heat requirement is met — bloom — and calls a year frosted when the minimum reaches −2.2 °C (28 °F, 10% bud kill at full bloom for sweet cherry; 27 °F for peach and apricot, in Washington State's Prosser measurements) within two weeks after. At McAllister:

That is the shape growers know: almonds are not grown in Tennessee because they bloom into February frosts, peaches lose a crop to spring frost some years, late-blooming cherries escape. In the 2050s every crop blooms about a week earlier and the frost count holds or falls by one — the frosts move earlier too. The grid's minimums are milder than an orchard floor on a still night, so these counts err low, and an orchard on a slope will do better than one in a frost pocket.

Apples and pears#

The stone-fruit review leaves out the commonest orchard trees. Two more CC BY 4.0 papers cover them, re-keyed by chill_requirements_pome.py into chill-requirements-pome.csv:

• Apples: the review by González Noguer, Delgado, Else and Hadley (2023, Frontiers in Horticulture 2:1217689). Its Table 2 holds 173 published requirements, each with the cultivar's use (dessert, cider, crab or rootstock), place, method and study. It is read from the article's HTML, where a cell spanning several rows carries its citation key down to each, and every key is checked against the author and year the cell prints. One cell writes a thousand with a point (Raxao, "1.005" chill hours), read as 1005.
• Pears: Fadón, Espiau, Errea and Alonso Segura (2023, Agronomy 13:518), 16 traditional European pears over twenty years at Zaragoza, in all three chill models and heat.

Both define a chill hour as the stone-fruit review does, 0 to 7.2 °C. The advisor counts the dessert and cider apples. It leaves out rootstocks, crab apples and the 16 cultivars the table gives no use for, since none of those says what will fruit. That leaves 78 apples with a chill-hour requirement, 54 of them from one study (Hauagge and Cummins 1991).

At McAllister, 55 of 78 apples and all 16 pears get their chill. The pears need 825 to 894 hours at the top of their spread, well inside the site's 1090. In the 2050s the apples fall to 39. Apples, unlike stone fruit, lose ground here, because Hauagge and Cummins put sixteen common cultivars between 1049 and 1090 hours: Golden Delicious, Gala, Fuji, McIntosh and Granny Smith among them. That is right where McAllister's 8-in-10 winter sits, so the 2050s' 43 fewer hours carry all sixteen across the line. The cliff is in the table more than in the orchard. Those requirements come from one place and one method, and a cultivar 20 hours short still blooms, only less evenly. But it marks which apples a warming winter here reaches first.

Only 24 apple estimates come with a heat requirement too, and 21 of them are local apples from northwestern Spain, 19 of those for cider (Delgado and others 2021). So the apple bloom below speaks mostly for those:

Two apples bred for warm winters, Anna and Dorsett Golden, need about 250 hours in the Egyptian study and bloom here at the start of March. Frost follows two years in three, which is why they are planted in Florida and not Tennessee. The estimates disagree as much as they do for stone fruit. Hauagge and Cummins give Dorsett Golden 814 hours to Egypt's 238, and a Moroccan statistical estimate pairs Gala's 537 hours with 22,402 degree-hours of heat, which simulates bloom on May 19. Estimates from warm-winter sites trade chill for heat and do not travel. The pear estimates, from one orchard and one method, barely differ from each other, which says more about the method than about the pears.

Bees per plant, from GloBI#

plant-bees-globi.csv: for every plant on the list, the bees recorded visiting or pollinating it in GloBI — 327 of 473 plants, from 49 source datasets credited in plant-bees-globi-sources.txt (CropPol, the global crop-pollination database; DoPI; EuPPollNet; iNaturalist; museum collections). The managed bees on the animal list turn up where they should: the squash bee on squash, the blue orchard bee and its relatives on apple, the alfalfa leafcutting bee on alfalfa, the southeastern blueberry bee on blueberry, the sugarbag bee on macadamia, stingless bees on coffee.

Two cautions. These are records, not dependence: maize has 337 bee records because bees collect its pollen, but it is wind-pollinated. And records follow where people looked — Europe and North America, crops and gardens — so few records is not few bees. Queries are limited to bee families and capped at 4,000 records; 26 plants reached the cap, and their counts are lower bounds. The first version of the query took every interaction and filtered after: a well-studied plant's first few thousand records were its pests and diseases, and apple came back with no bees at all.

What mixes with what#

interactions.csv: 506 records of one thing affecting another, each with its mechanism and source. 154 come from USDA PLANTS fields that already encode a link — 40 palatable forages, 35 browses, 31 toxic plants, 28 nitrogen fixers, 15 marked allelopathic, 5 bloat risks; 327 from GloBI's bee records; 25 curated from the plan, each now carrying its own citation (a paper, an extension guide or a regulation) and worded to what that source supports. Checking them changed five: poultry following cattle is unproven (a controlled trial found broilers left dung pats and face-fly larvae alone); ducks in rice are mixed (weeds down, planthoppers up, because the ducks eat the spiders too); squash bees prefer tilled soil, so it is deep tillage over nests already dug that harms them; nothing supported "pigs tolerate acorns" as a toxicology claim; and kudzu and mowing during bloom were dropped where no source carried them.

The PLANTS toxicity flags line up with known agronomy — sorghum's prussic acid, black cherry and chokecherry's cyanide, sweet clover's dicoumarol, leucaena's mimosine, buckwheat's photosensitisation — while "bloat: medium" on maize is another PLANTS oddity.

Not checked yet#

For the 28 crops with a crop model, temperature, day length and time to maturity — and for wheat and barley, vernalization, winter kill, freeze after boot and heat at flowering — are checked together, by simulation; for 531 stone-fruit cultivars, 78 apples and 16 pears, winter chill, and for 167 of them, bloom date and frost after it. Still not checked: chill and bloom for blueberries, heat damage to fruit set in any crop but tomato, a rice nursery raised under cover, oats (no DSSAT model), and everything for the 255 plants with neither source — most fruit, herbs and flowers.

Heat at fruit set is checked for tomato, the one crop with a threshold a field can be held to: Sato, Peet and Thomas (2000) found four of five cultivars set no fruit at 32/26 °C day and night, where ten days' relief before flowering restored it (fruit_set.py). At McAllister the days get there — ten-day mean highs of 32 °C on up to 77 days a summer — but the nights never do: the warmest ten-day stretch of nights in thirty years averaged 23.6 °C. So heat does not stop tomato fruit set here, and in the 2050s, a degree warmer, it still would not. Pepper and bean were considered and left out: the pepper study held a constant 33 °C, which no field's day is, and the bean threshold could not be read from an open copy of its study.

The plants with neither source, tried against NRCS Plant Guides (2026-10-09). The guides are public domain and written for exactly these questions, so they were the recommended next source. Of the 285 plants with no PLANTS characteristics, 22 have a guide (lentil, pea, faba bean, sesame, camelina, taro, goldenseal, ginseng, milkweed and others), and their prose gives a pH, rainfall or hardiness range that a pattern can find for 4 — lentil with two different pH ranges in one guide. That is not a source to judge a plant by, so they stay as the report lists them, not judged, rather than judged on a guess. What would change it is a tabulated source: ECOCROP, listed under CC BY 4.0 in FAO's data catalog, once FAO confirms that and a bulk file is found.

Blueberries, looked for and not found (2026-10-09). No openly licensed table of chill by cultivar turned up — not in MDPI or Frontiers, where the apple and stone-fruit tables came from; extension guides carry figures (southern highbush 150–800 hours, rabbiteye 350–650, northern highbush 800 and more) under terms that do not allow reuse. Two routes remain. US plant patents are public domain and each new blueberry's states its chill requirement, but the patent office serves them as scanned images, so a table needs OCR of each patent and a list of them. USDA GRIN-Global holds the national Vaccinium collection at Corvallis and may carry chill evaluations, public domain, but has no bulk download and its accession pages were down when tried. Until one of those is built, the advisor says nothing about blueberry chill rather than borrow numbers it cannot cite.

Reproducing#

python3 modules/advise/plant_conditions.py modules/advise/data/plant-list.csv --cache CACHE \
    --out modules/advise/data/plant-characteristics-usda.csv
python3 modules/advise/site_profile.py 35.06445 -86.59137 --cache CACHE --json site.json
python3 modules/advise/plant_match.py modules/advise/data/plant-list.csv \
    modules/advise/data/plant-characteristics-usda.csv site.json --out match.csv

Climate: NASA POWER, CC BY 4.0. Soil: USDA NRCS SSURGO, public domain. Plants: USDA NRCS PLANTS, CC0. Bees: GloBI, CC BY 4.0, and the studies it cites.