CropProphEU

CI Python ≥3.11 License: MIT GitHub stars Discussions Tests Latest Release

13 MCP Tools, 5 Crops, 26 EU Countries, 123 NUTS2 Regions — Yield forecasts, market values (€/ha), risk analysis, environmental risk scoring & portfolio optimization for European agriculture. Built for AI agents, by AI agents.

"How will wheat perform in Sachsen-Anhalt this year? What's my best €/ha allocation across 100 ha?"

pip install git+https://github.com/DasClown/CropProphEU.git

---

Features (13 MCP Tools)

#	Tool	What it does	V
1	yield_and_value	Yield + market value (€/ha) + plain-language summary (DE/EN)	V4.6
2	europe_yield_forecast	Pan-European RF forecast with Yield-at-Risk (P10/P50/P90) + NDVI correction	V4.3
3	crop_forecast	Current season: GDD, rain, soil moisture, drought index, frost warnings	V4.0
4	compare_regions	Batch-compare 20 regions × 5 crops with live market prices	V5.1
5	portfolio_optimizer	AI investment engine: budget → optimal allocation across regions × crops	V5.1e
6	season_comparison	Compare this season to historical years	V4.0
7	region_health	All crops for one region, single call	V4.5
8	weather_outlook	16-day weather forecast	V4.0
9	climate_scenario	What-if: +2°C, -20% rain, etc.	V4.4
10	yield_forecast	Analog-year yield matching (DE-focused)	V3.0
11	list_regions	123 NUTS2 regions across 26 countries	V4.2
12	list_crops	Crop parameters (GDD base, season, frost sensitivity)	V4.5
13	environmental_risk	NEW V5.4 — ERS (forest, erosion, storm, hail) + wild boar damage risk for DE	V5.4

---

Quick Start

1. Install

pip install git+https://github.com/DasClown/CropProphEU.git

2. Use as MCP Server

CLI (stdio):

crop-mcp

Python API:

from crop_mcp import predict_europe_yield

result = predict_europe_yield("DE11", "DE", crop="wheat", gdd=3050, precip_mm=650)
print(f"Yield: {result['predicted_yield_t_ha']} t/ha")
print(f"Revenue: ~{result['predicted_yield_t_ha'] * 239:.0f} €/ha")

3. Claude Desktop / Cursor

{
  "mcpServers": {
    "crop": {
      "command": "python3",
      "args": ["-m", "crop_mcp.server"]
    }
  }
}

4. HTTP Server (Remote / Smithery)

pip install crop-mcp[http]
crop-mcp --http --port 8080

Connect via SSE: http://your-server:8080/sse

5. Docker

docker build -t crop-mcp .
docker run -p 8080:8080 crop-mcp crop-mcp --http --port 8080

---

Verified Crops — Data Integrity ✅

Every prediction traces to a verified Eurostat crop code. No hallucinations, no silent wrong-crop training.

Crop	Eurostat Code	Samples	Countries	MAE (LOYO)	R²
🌾 Wheat	C1100	1,603	26 (🇪🇺+🇺🇦)	11.5%	0.87
🌽 Corn (Maize)	C1500	1,797	21 (🇪🇺+🇬🇧)	11.6%	0.72
🌿 Barley	C1300	1,885	26 (🇪🇺+🇬🇧)	11.2%	0.85
🌻 Rapeseed	I1110	1,825	25	10.8%	0.83
🌻 Sunflower	I1120	1,229	17	16.1%	0.74

⚠️ V5.1d Data Fix: Rapeseed + Sunflower were previously trained on RICE data (wrong Eurostat codes C2000/C2200). Now corrected to Industrial crop codes I1110/I1120. DE rapeseed prediction fell from 7.21t to 2.63t — real, not extrapolated.

---

Why CropProphEU?

Capability	CropProphEU	Open-Meteo MCP	Gro Intelligence
Yield forecasts	✅ 5 crops	❌	✅ $10K+/yr
Soil features	✅ 11 properties	❌	✅
Yield-at-Risk (P10/P90)	✅	❌	✅
Live market prices (€/ha)	✅ CBOT + MATIF	❌	✅
Climate what-if	✅	❌	✅
Frost warnings	✅	✅	❌
NDVI satellite correction	✅	❌	❌
Portfolio optimizer	✅	❌	❌
Multi-language (DE/EN)	✅	❌	❌
Environmental Risk	✅ V5.4	❌	❌
Price	Free	Free	$10K+/yr

Unique: Only free MCP server covering EU agriculture with soil → yield → market value → environmental risk → portfolio optimization in one pipeline.

---

V5.4 — Environmental Risk Score + Wildschaden 🌍🐗

Feature	Beschreibung
Environmental Risk Score	Komposit aus Waldanteil, Maisfläche, Bodenerosion, Sturm- + Hagelrisiko → 🟢🟡🔴
Wildschaden DE	DJV-Jagdstreckendaten + Waldrandindex + Maisflächenanteil → €/ha-Verlustschätzung
Ampel-System	🟢 low (<35), 🟡 moderate (35-65), 🔴 high (≥65)
MCP Tool	environmental_risk(region='DE26') → sofortige Analyse inkl. Wildschaden

Beispiel Maßbach (DE26 Unterfranken):

{
  "overall_risk": "🔴 high",
  "ers_level": "🟡 moderate",
  "wild_boar_risk": {"level": "🔴 high", "loss_eur_ha": 158},
  "management": ["Waldrandstreifen 3-6m", "Drückjagd Nov-Dez", "8-Tage-Anzeigefrist §36 BayJG"]
}

V5.4 — Testing & CI 🧪

Maßnahme	Status
pytest	16 Tests, alle passing (tests/test_crop_mcp.py)
CI (GitHub Actions)	✅ Aktiv bei jedem Push — Badge im Header grün
Git LFS	*.pkl + große .json via LFS (aus Git-Tree entfernt)

Run tests:

pip install -e ".[test]"
pytest tests/ -v

---

Model Accuracy

Metric	Value
LOYO MAE (Wheat)	0.599 t/ha (11.5%)
Forward Validation (Train ≤2022, Test 2023-24)	0.794 t/ha (15.0%)
R² (LOYO)	0.871
R² (Forward)	0.628

Most accurate for core EU (DE, FR, BE, NL, AT, CZ) where training data is dense.

Per-Crop Performance (V5.2)

Crop	Algorithm	Top Feature	Key Insight
🌾 Wheat	RF 200 trees	solar_kwh (35%)	Nord/Süd gradient dominates
🌽 Corn	RF 200 trees	clay_pct (42%)	Maize is extremely soil-sensitive
🌿 Barley	Ridge	clay_pct (27%)	Best coverage of all crops
🌻 Rapeseed	RF 200 trees	coarse_pct (28%)	Corrected — now 1,825 real samples
🌻 Sunflower	Ridge	silt_pct (24%)	17 countries (post-fix)

---

Live Market Prices

Crop	Source	€/t (Mai 2026)	Market
Wheat	✅ CBOT ZW=F + MATIF premium	239	Euronext MATIF
Corn	✅ CBOT ZC=F + MATIF premium	189	Euronext MATIF
Barley	✅ Reference (AMI regional)	190	AMI regional exchanges
Rapeseed	✅ Reference	470	Euronext MATIF (ECO)
Sunflower	✅ Reference	420	ICE / Black Sea

Production costs are country- and crop-specific (V5.3+). See market_prices.py → COUNTRY_PRODUCTION_COSTS (28 EU countries × 5 crops). Sources: FADN, KTBL, ARVALIS. Examples:

Country	Wheat	Corn	Barley	Rapeseed	Sunflower
🇩🇪 DE	650	700	600	780	650
🇫🇷 FR	700	650	600	750	600
🇵🇱 PL	550	600	500	650	550
🇷🇴 RO	450	500	400	550	450
🇪🇸 ES	600	650	550	700	—

ℹ️ Full table: COUNTRY_PRODUCTION_COSTS dict in crop_mcp/market_prices.py

---

Data Sources

Source	Data	Access
Eurostat	Crop yields (apro_cpshr) — 25+ years, verified codes	Free, no key
NASA POWER	GDD, precip, solar, soil moisture	Free, no rate limits
Open-Meteo	16-day forecast, GDD	Free, no key
SoilGrids v2 (ISRIC)	11 properties: SOC, pH, N, CEC, clay, sand, silt, bdod (bulk density), cfvo (coarse fragments), AWC, coarse	Free REST API
LUCAS Soil (ESDAC)	Texture ~20K field points + coarse fragments	Free download
Sentinel-2 NDVI	Vegetation index (Copernicus STAC + Planetary Computer fallback)	Free, no auth
Yahoo Finance	Live CBOT wheat/corn futures, EUR/USD	Free, no key

Zero API keys required — all sources are free and public.

---

Example Output

German (default):

Weizen – Region DEE0 (DE)
Ertrag: 7.35 t/ha (Spanne 6.50–8.20)
Temperatur: warm (2950°C Wärmesumme)
Niederschlag: ausreichend (480 mm)
Bodenfeuchte: feucht (48%)
Modellabweichung: ±11.5% (1603 Samples, 26 Länder)
Vergleich zu 2024: +0.15 t/ha (im Rahmen des Vorjahres)
Marktwert: 1.757 €/ha @ 239 €/t
Kosten: 650 €/ha → Deckungsbeitrag: 1.107 €/ha

English (with language="en"):

Wheat – Region DEE0 (DE)
Yield: 7.35 t/ha (range 6.50–8.20)
Temperature: warm (2950°C GDD)
...

---

Architecture

crop-mcp/
├── crop_mcp/
|├── server.py                 # 857 Zeilen — Pydantic-Modelle + Tool-Registry + MCP-Init
│   ├── tools/                  # **V5.4e** — Handler in logische Module aufgeteilt
│   │   ├── weather.py          # weather_outlook, crop_forecast, season_comparison, region_health
│   │   ├── yield_tools.py      # yield_forecast, europe_yield_forecast, yield_and_value, climate_scenario
│   │   ├── market.py           # compare_regions, portfolio_optimizer
│   │   ├── info.py             # list_regions, list_crops
│   │   ├── environmental.py    # environmental_risk (V5.4)
│   │   └── helpers.py          # Shared utilities (NDVI correction, frost, language)
│   ├── europe_model_api.py       # RF (200 trees) + Yield-at-Risk + NDVI correction
│   ├── environmental_risk.py     # V5.4 — ERS + Wildschaden DE
│   ├── ndvi_correction.py        # Sentinel-2 NDVI correction factor (±30%)
│   ├── market_prices.py          # Live CBOT/MATIF via Yahoo Finance
│   ├── feature_cache.py          # Sub-second historical queries
│   ├── simulate_yield.py         # Analog-year matching
│   ├── auto_update.py            # Monthly retrain cron
│   ├── core/regions.py           # 123 NUTS2 regions
│   └── sources/                  # Weather, soil, NDVI, Eurostat, FAOSTAT fetchers
├── models/                       # .pkl files ➜ download from Releases
├── data/                         # Training data ➜ download from Releases (or generated by build)
├── tests/                        # V5.4 — 16 pytest tests
├── .github/workflows/ci.yml      # V5.4 — CI workflow (lokal, benötigt PAT mit workflow-Scope für Push)
├── pyproject.toml
└── README.md

Key design principles:

• No hallucination — every yield prediction traces to verified Eurostat data
• Live prices — CBOT wheat/corn via Yahoo Finance, updated hourly
• Self-updating — monthly cron rebuilds models with latest Eurostat data
• Zero API keys — all data sources are free and public
• AI-for-AI — built for agents, no dashboards

---

Building & Training

# Build training data (25 min per crop)
python3 build_europe.py --crop corn

# Train model (2 min)
python3 train_europe_fast.py --crop corn

# Auto-update monthly (cron: 1st of month at 06:00)

---

Commercial Use Cases

• Agri-trading desks: "What's wheat worth in Picardie at current MATIF prices?"
• Farm advisory: "How does this season compare to the last 5 years?"
• Insurance / Risk: Yield-at-Risk (P10/P50/P90) per region + crop
• EU policy analysis: Climate scenario impact on national yields
• Investment: Portfolio optimizer for 100+ ha allocation decisions

---

🤝 Getting Help & Contributing

Channel	Purpose
💬 GitHub Discussions	Questions before coding, feature ideas, community chat
🐛 GitHub Issues	Bug reports, confirmed feature requests
📖 CONTRIBUTING.md	Development setup, branch naming, commit conventions, code style

New contributors welcome! See CONTRIBUTING.md to get started.

---

License

MIT — free to use, modify, and distribute.

Built with ❤️ for AI agents that need real, verifiable crop intelligence.