Technical mentoring session · UC Berkeley
What Makes a Fertilizer
Recommendation Defensible
Diego Ramos · Agricultural Mechatronics Engineer,
Universidad Autónoma Chapingo · Software developer
diegotec13@hotmail.com
+52 55 6069 9009
Questions after the session are welcome — write or call.
Press → or swipe to advance.
01
The premise
There is no “right fertilizer.”
There is only a right combination of four decisions. Change one and the other three have to move.
The 4R framework
- Right Source — which product (urea, DAP, potash)
- Right Rate — how much per acre
- Right Time — when in the season
- Right Place — broadcast, banded, injected
Source: 4R Nutrient Stewardship — IPNI, The Fertilizer Institute, USDA‑NRCS. Taught by land‑grant extension across the US.
02
The analogy that fixes it
A recommendation is a prescription, not a match.
Compare
| Prescription | Fertilizer recommendation |
|---|---|
| Drug — amoxicillin | Source — urea, DAP, potash |
| Dose — 500 mg | Rate — lb per acre |
| Schedule — 3× daily | Time — planting, sidedress |
| Route — by mouth | Place — banded, injected |
Say only “amoxicillin” and you have said almost nothing useful — and possibly something dangerous. Your output is a prescription.
03
Where your app sits
You are matching on one of four dimensions.
That is fine for an MVP. It is not fine to not know it, and it is not fine to not say it in your UI.
04
Scoping
Three questions define your entire product.
- Who is the user? A US commercial grain grower, a home gardener, or a smallholder. These are three different products, not three settings.
- What crop? Every crop you add multiplies your coefficient tables and your liability. A legume like soybean usually needs zero fertilizer N.
- Do they have a soil test? This one is the master switch. Next slide.
Red flag
“It works for all crops and all users.” A recommendation engine that does everything has been calibrated for nothing.
05
The master switch
No soil test, no phosphorus or potassium number.
With a soil test
You can do real, defensible P and K math. Soil supply is known, the previous crop can be credited, and the result is checkable against an extension table.
Without one
The honest output is “get a soil test” — not a number. Phosphorus applied where soil is already high is wasted money and a water‑quality liability.
Design consequence
Treat “no soil test” as a hard gate that switches your output from numbers to advice. That is a branch in your code, not a disclaimer in your footer.
06
Scoping verdict
What is realistic in four weeks.
Build this
One crop (corn) · one user (US commercial grain grower) · soil test required for P and K.
This is the only segment where the data are free and authoritative, the units are consistent, the liability is manageable, and the math is genuinely implementable in four weeks.
- Home gardener — easy to build, low stakes, but the science barely matters. Intellectually thin.
- Smallholder — highest social value, hardest problem. Fragmented data, often no soil test, high risk of doing harm.
07
Inputs
Essential variables — and what breaks without them.
| Variable | What breaks if missing |
|---|---|
| Crop | No coefficients → no calculation at all |
| Yield goal | Removal scales ~linearly; over‑estimate and you over‑apply |
| Soil test P & K with units and extractant | No responsible rate exists. Refuse the number. |
| Soil pH | Source rules cannot fire — volatilization, P fixation, chelate choice |
| Previous crop | Legume N credit ignored → N over‑applied |
| Region / state | Wrong critical levels; recommendations are state‑calibrated |
08
Inputs
Desirable, and safe to default.
Desirable — improves precision
- CEC — sets the K critical level. Unknown? Default to the higher value to stay safe.
- Organic matter / manure history — manure supplies N‑P‑K credits; ignoring it is a classic over‑application.
- Tillage system — affects mineralization and urea volatilization risk.
Dispensable — default it
- Removal coefficients — published book values.
- N recovery efficiency — with a loud caveat; it varies enormously.
- Sampling depth / bulk density — standard 6‑inch layer.
Design rule
Every default must be visible and editable. A hidden default is an invented number wearing a lab coat.
09
The arithmetic
The mass balance, honestly stated.
Skeleton
Two honesty flags
1. This logic is most defensible for P and K. For nitrogen in US corn it is not the current scientific consensus — slide 15.
2. The recovery divisor is the largest source of uncertainty in the whole equation. Do not present it as precise.
10
Worked example — part 1
Corn, 180 bu/acre, maintenance case.
Given
Both sit between the critical level and the maintenance limit. That means: replace what the crop removes. No build‑up, no zero.
Removal
Coefficients & thresholds: Tri‑State Fertilizer Recommendations for Corn, Soybean, Wheat and Alfalfa (2020), Bulletin 974 — Culman, Fulford, Camberato & Steinke. Mehlich‑3.
11
Worked example — part 2
Nutrients are not products.
Allocation — satisfy the constrained nutrient first
The DAP you bought for phosphorus also delivered nitrogen. Miss that credit and you double‑apply N.
12
The trap
P₂O₅ is not P. K₂O is not K.
This is the error that will break your app, and it makes phosphorus come out about 2.3× wrong.
The mismatch
Fertilizer bags are labelled N – P₂O₅ – K₂O — oxides, by law.
Soil test labs usually report elemental P and K.
The molecule “P₂O₅” is not in the bag. It is a labelling convention, kept for legal and historical reasons.
Why: AAPFCO labelling rules require the guaranteed analysis to state N as total N, P as available P₂O₅, and K as soluble K₂O.
13
The trap, continued
Three more unit traps, all silent.
- ppm vs lb/acre. For the standard 6‑inch layer,
lb/acre ≈ ppm × 2. So 20 ppm P ≈ 40 lb P/acre. Depth‑dependent. - US vs metric.
1 lb/acre = 1.121 kg/ha.1 hectare = 2.471 acres. An agronomist in Mexico thinks in kg/ha; a US grower thinks in lb/acre. - Bushel weight differs by crop. Corn 56 lb, wheat and soybean 60 lb, barley 48 lb, oats 32 lb.
The engineering fix
Store one canonical internal unit. Convert only at the edges — input and display. Otherwise you will silently ship answers off by 2.29×, 2×, or 2.47×, and every one of them will look plausible.
14
Honesty
For corn nitrogen, the mass balance is superseded.
The old yield‑goal method — yield × a factor, minus credits — was standard until the mid‑2000s. Corn‑Belt land‑grant universities moved to MRTN (Maximum Return To Nitrogen), which sets the economically optimum rate from hundreds of N‑response trials.
The uncomfortable part
MRTN deliberately does not use a yield goal, because trials show a weak relationship between yield level and the optimum N rate.
And even for P and K there is no single US answer. A survey of 48 states plus Puerto Rico found land‑grant philosophies split: sufficiency 37%, build‑and‑maintain 19%, hybrid 20%, multiple 20%.
Sources: Corn Nitrogen Rate Calculator (cornnratecalc.org). Lyons et al. (2023), Soil Science Society of America Journal 87:985–998.
15
Source rules · 1 of 3
Urea left on the surface turns into gas.
Rule 1 — strongest evidence, biggest effect
IF urea or UAN is surface‑applied and left un‑incorporated on a high‑pH or calcareous soil, warm and drying —
THEN incorporate it, apply before rain or irrigation, or use an NBPT urease inhibitor —
BECAUSE urea hydrolyzes to ammonium and, at high surface pH, converts to ammonia gas that volatilizes away.
- Losses average ~16% of applied N worldwide, and reach 40–50% in hot tropical regions.
- A global meta‑analysis found up to 64% lost as ammonia (average 18%).
- NBPT works: cumulative loss 31.0% → 14.8% — a 52% reduction.
Sources: Cantarella et al. (2018). Pan et al. (2016), Agriculture, Ecosystems & Environment. Silva et al. (2016), Agronomy Journal — meta‑analysis.
16
Source rules · 2 of 3
Chloride and pH pick the product for you.
Rule 2 — chloride sensitivity
IF the crop is chloride‑sensitive (tobacco, potato, grapes, avocado, citrus) THEN prefer potassium sulfate over muriate of potash BECAUSE chloride degrades quality and adds salt.
For corn, soy, wheat and cotton, muriate is fine and cheaper. Sulfate carries a real price premium — do not prescribe it by default.
Rule 3 — iron chelates
IF correcting iron chlorosis THEN match the chelate to soil pH BECAUSE each one holds iron in solution only up to a characteristic pH; above it the iron precipitates and the product is wasted money.
17
Source rules · 3 of 3
Salt near the seed, and phosphorus that disappears.
Rule 4 — seed‑safe salt limit
IF placing fertilizer in direct seed contact THEN cap the salt load — for corn in 30‑inch rows, no more than about 10 lb/acre of N + K₂O in contact (K‑State cites a tighter 6–8 lb in‑furrow) — BECAUSE soluble salts raise osmotic pressure and pull water out of the seedling.
Sensitivity order: wheat (least) < corn < soybean < vegetables (most).
Rule 5 — phosphorus fixation
IF soil is strongly acidic or strongly calcareous THEN correct pH toward 6–7 and band the phosphate rather than broadcasting BECAUSE outside that window applied P precipitates as insoluble iron/aluminium or calcium phosphates.
Extension consensus; exact figures vary by state, crop and row spacing. Treat the seed‑safe number as guidance and verify it against your state's table.
18
A cautionary case
A method that looks site‑specific and is nearly blind.
A widely taught Spanish‑language method solves the same crop at the same yield goal — strawberry, 70 t/ha — first without a soil test, then with one.
| Nutrient | No soil test | With soil test | |
|---|---|---|---|
| P | 84 | 84 | identical |
| Ca | 175 | 175 | identical |
| S | 35 | 35 | identical |
| B | 1 | 1 | identical |
| Zn | 1.8 | 1.8 | identical |
| K | 210 | 175 | moved |
| Mg | 63 | 70 | moved |
| N | 126 | — | dropped entirely |
Five of seven nutrients get the identical rate whether or not you tested the soil. The test only nudges the answer between two pre‑set factors.
19
Evidence
Tools built by professionals, validated in the field, still fail.
| Finding | Magnitude |
|---|---|
| Public corn‑N tools vs the economic optimum | within ±30 kg N/ha at only 21–47% of sites |
| Profitability loss from poorly calibrated tools | US$56–155 per hectare |
| Optical sensor algorithms, US Midwest | deviated 81–147 kg N/ha from optimum |
| Removal coefficients across cultivar and environment | uncertainty up to 52% |
| Critical soil test value, same data, different model | 46–66 mg/kg; false negatives 0–18% |
| Digital N strategies, Australian cereals | RMSE 15–57 kg/ha |
Sources: Ransom (2018). Bean et al. (2018), Agronomy Journal 110:2552. Ludemann et al. (2024), Nutrient Cycling in Agroecosystems. Slaton et al. (2024), SSSAJ. Colaço et al. (2024), Precision Agriculture.
20
Limits
What you cannot promise.
- Require a soil test before any P or K rate. No test, no number.
- Refer out for any crop or region outside your validated set, and for anything environmentally regulated.
- Never promise a yield. Not “apply this and get X bushels,” not a guaranteed return.
The credentialing reality
In the US, agronomic advice is given by Certified Crop Advisers — a certification established in 1992, held by more than 13,000 advisers, requiring exams, documented field experience, a code of ethics, and 40 continuing‑education units every two years.
Your app is not a CCA. Say so in the product, not just in the terms of service.
Also relevant: NRCS Conservation Practice Standard 590 and state nutrient‑management rules require plans built on a recent soil test. Recommending P on an already‑high‑P field can put your user out of compliance.
21
Limits
Over‑applying is not the safe direction.
There is a strong instinct — in users and in code — that a little extra fertilizer is harmless insurance. It is not.
Salt burn
Excess soluble salt near the seed raises osmotic pressure and pulls water out of the seedling. The plant dehydrates in wet soil.
Eutrophication
Nutrients that leave the field reach groundwater and surface water. Excess nutrient loading damages aquatic life — which is why phosphorus is regulated in sensitive watersheds.
Build this in
Above the maintenance limit, recommend zero. Not a small amount. Zero. That single guardrail prevents most of the harm your app could do.
Mechanism and regulation are well established. Loading thresholds are site‑specific — consult your state's water‑quality guidance rather than a general figure.
22
Next step
One concrete test, this week.
The validation
Take one crop. Take a published recommendation table from a state extension service. Feed your engine the same inputs. Compare the outputs.
- If they match — you have something real. You can now say your engine reproduces a published recommendation, which is a genuine claim.
- If they don't — even better. You know exactly where your error is, and it is almost certainly a unit conversion.
Add one more check: make your app refuse to answer when the soil test is missing. If it still produces a number, that is a bug, not a feature.
Where to look: Tri‑State Bulletin 974 (2020). Iowa State PM 1688. Corn Nitrogen Rate Calculator — cornnratecalc.org. FRST, the national soil‑test calibration database.
23
Reference
Minimum glossary.
| Español | English | Trap |
|---|---|---|
| extracción | removal | Not “extraction” — that is the lab chemistry |
| absorción | uptake | Uptake ≠ removal. Keep them distinct. |
| dosis | rate | US agronomy says rate, not dose |
| abono | fertilizer or manure | False friend — depends on context |
| análisis de suelo | soil test | — |
| análisis foliar | tissue test | A lab test, not visual diagnosis |
| unidad fertilizante | no US equivalent | 1 kg of nutrient/ha. Explain, don't translate. |
| quintal | quintal | Metric = 100 kg; traditional Mexican ≈ 46 kg |
| fertirriego | fertigation | One word in English |
| labranza cero | no‑till | — |
Deck and sources: this page.
Questions after the session:
diegotec13@hotmail.com ·
+52 55 6069 9009
24