Water and fertiliser are applied uniformly or by habit on most fields, although soil, crop stage and weather vary within and between fields. Precision systems fuse soil sensors, weather forecasts, satellite imagery and crop models, then compute when and how much to irrigate and where to apply how much nitrogen. Typical results are lower input use at equal yield and better documentation for nutrient and water regulation; gains depend strongly on crop, soil and baseline practice.
Problem & value
- Today: irrigation follows fixed rotations or experience; nitrogen is applied by field-level rules of thumb or by calendar. Over-application causes nitrate leaching and wasted cost, under-application costs yield and quality.
- AI value: (1) forecast crop water demand (evapotranspiration, soil water balance) and schedule irrigation against it, (2) estimate in-season crop nitrogen status from satellite/drone indices and generate site-specific, split-dose application maps, (3) respect hard constraints (permit volumes, pump capacity, nutrient caps) through an optimiser.
- Pressure in the EU: the Nitrates Directive and national fertiliser ordinances (in Germany the Düngeverordnung, including red zones with reduced N limits) restrict nitrogen; the CAP requires documentation of nutrient and water management; drought years raise the cost of water [1][2].
- Evidence: the FAO AquaCrop and irrigation-scheduling literature show that deficit and sensor-based scheduling can reduce water use in many crops; magnitudes are trial-specific, so measure your own baseline rather than relying on vendor averages [3].
Variants
- Arable (cereals, maize, sugar beet, potatoes): variable-rate nitrogen from satellite indices, split dosing, plus centre-pivot or reel irrigation scheduling.
- Horticulture and vegetables: drip irrigation and fertigation with soil moisture probes and greenhouse climate coupling.
- Orchards and viticulture: regulated deficit irrigation for quality; zone-based management.
- Cooperatives and advisory services: service provider runs a platform for hundreds of farms and delivers recommendations as an advice subscription.
- Public sector: water authorities allocating irrigation volumes in water-scarce basins.
- Private use: smart garden irrigation (weather-aware controllers) is a simplified variant.
How it works
- sensors (soil moisture, EC, weather) + satellite/drone + soil map + machine logs
- data harmonisation per field/zone (time series, geodata)
- state estimation (soil water balance, crop stage, N status via vegetation indices)
- forecast (ET0, rain, crop demand, N uptake) using weather forecasts
- optimiser (schedule / dose under water permit, pump, nutrient-cap constraints)
- recommendation (map or schedule)
- farmer/advisor review
- actuation (irrigation controller, ISOXML task for spreader/sprayer)
- feedback from yield and sensors
- Crop and soil models: mechanistic models (AquaCrop, DSSAT, APSIM) simulate water and N dynamics; ML models (gradient boosting, LSTM/temporal transformers) correct or replace parts using local data. Hybrid "physics + ML" is the pragmatic standard.
- Remote sensing: Sentinel-2 indices (NDVI, NDRE, chlorophyll proxies) give crop vigour every few days under clear sky; drones fill gaps and add resolution.
- Optimisation: linear/mixed-integer programs or model-predictive control for schedules; reinforcement learning is mostly in research, validated in simulators first.
- LLM layer (optional): an assistant explains recommendations, drafts documentation for authorities, and translates natural-language questions into queries (see OPT-018).
- Human in the loop: the farmer approves application maps and irrigation plans; autonomous actuation only inside hard limits (maximum volume per day, permit cap).
Data preparation
Data needed: field boundaries, soil maps and samples, yield maps, application records, weather, satellite imagery and machine telemetry. It typically lives in the farm-management software, machine terminals and manufacturer portals (often in proprietary formats), paper field books, Excel sheets, and with advisers or the cooperative. Typical gaps: field boundaries that do not match between systems, missing or inconsistent application records, yield maps with sensor artefacts, lab results on paper, different data owners per machine brand, and unclear rights to machine data.
Preparation steps: 1. Fix field boundaries once as a master layer and reuse it everywhere. 2. Digitise paper field books and lab reports (PREP-006, PREP-008). 3. Collect machine data through agrirouter or manufacturer exports (PREP-005, PREP-007). 4. Clean yield and sensor data: remove outliers and overlaps, tidy spreadsheets (PREP-011). 5. Clarify who owns and may use the data, in line with the EU Data Act.
Tools: QGIS, Python (geopandas, rasterio), agrirouter, farm-software exports. Rough share of effort: 40-60 percent, and several seasons of history improve results; start collecting now.
Implementation paths
Which path for whom
- SMB (farms): A, because mature farm-management and field platforms (xarvio, 365FarmNet, machine-maker portals) exist at farm prices and integrate with machines via standard interfaces.
- Enterprise (cooperatives, agribusiness, large estates): B, because models must be integrated with own field, machine and trading data and calibrated per region.
- Developer: C, because open data (Copernicus, DWD) and open-source crop and geo stacks run on HW1 and avoid lock-in.
- Public (advisory services, research institutes): C, because sovereign, reproducible open-source tooling on German or EU infrastructure suits public advice and funding rules.
Path A — Buy
- Farm management and field platforms: xarvio (BASF, DE), 365FarmNet (DE), agrirouter- connected tools, Climate FieldView (Bayer), John Deere Operations Center, CLAAS/Fendt/Krone telematics, Trimble Ag, Raven/CNH.
- Irrigation and sensors: Netafim (Orbia) and Lindsay (FieldNET), Valmont, Pessl Instruments / METOS (AT), Sentek, Arable; Hortimax/Priva for horticulture (NL).
- Satellite and N-advisory: Farmers Edge, Cropwise, Agricolus (IT), EOS Data Analytics, Sentinel Hub / Planet, national tools such as N-Sensor ALS (Yara, NO) and Yara Atfarm.
- Choose A when: standard crops, machinery with ISOBUS terminals, no data-science team.
Path B — Build on cloud APIs
- Data inputs: Copernicus Data Space Ecosystem and Sentinel Hub APIs (EU), Copernicus Climate Change Service ERA5, DWD and ECMWF open weather data, openEO for cloud-based processing [4][5].
- Compute: EU regions of AWS, Azure, Google Cloud or EU providers (STACKIT, OVHcloud, IONOS); managed time-series and geospatial databases (PostGIS, TimescaleDB).
- Solver: OR-Tools, Gurobi or HiGHS for scheduling; pipelines orchestrated with Airflow.
- LLM: a European or frontier model with EU region (Mistral, Claude, Gemini) for explanation, report drafting and tool-calling over farm data.
- Choose B when: you are an ag-tech vendor, cooperative or consulting company building a service.
Path C — Self-host (open source)
- Crop models: FAO AquaCrop (and AquaCrop-OSPy), DSSAT, APSIM, WOFOST via PCSE (Wageningen) [3].
- Irrigation tools: FAO CROPWAT and pyfao56 (FAO-56 dual crop coefficient) for ET-based scheduling.
- Geo and EO: GDAL, rasterio, STAC clients, openEO, Sentinel-2 via the Copernicus data space; farm platforms like farmOS for records.
- ML: scikit-learn, LightGBM, PyTorch; time-series foundation models such as Chronos or TimesFM can serve as baselines for sensor forecasting.
- Edge: sensor gateways on Raspberry Pi/industrial controllers (ESP32 nodes, LoRaWAN, MQTT, ChirpStack) running local control loops so irrigation continues offline.
- Hardware: HW1 suffices; GPU only for drone-image segmentation (HW2).
Path D — Custom training
Worth it for large estates, seed or fertiliser companies and research institutes with multi-year yield maps and trials. Needs several seasons of labelled data (yield, soil analyses, application logs) per crop and region; models do not transfer well across soils and climates, so start with simulators and fine-tune on local data.
State of the art (2026)
- Hybrid models (crop-model state plus ML correction) outperform pure ML on small field datasets; satellite-derived indices with weather forecasts drive most in-season N advice.
- Sentinel-2 plus Sentinel-1 radar give regular crop monitoring through clouds; Copernicus data is free and open under the EU programme, and Copernicus EO foundation models (for example IBM/NASA Prithvi-EO and ESA-backed models) are used for crop-type and stress mapping [4].
- Sensor-driven irrigation is mainstream in horticulture and high-value crops; in broadacre arable it is spreading with LoRaWAN probes and virtual soil-moisture products.
- Foundation and time-series models reduce the need for per-field training for forecasting.
- Limits: satellite indices saturate in dense canopy and miss N effects in cloudy periods; soil probes represent only a small volume; yield response to N depends on weather after application, which no model foresees; recommendations vary by field and year. Data quality (sensor drift, wrong field boundaries) is the main failure mode, and vendor claims on savings are rarely independently audited.
Shared resources & combinations
- Same stack: remote-sensing and time-series capabilities shared with PRED-013 (yield prediction), VIS-021 (satellite imagery analysis) and PRED-017 (AI weather forecasting).
- Upstream: VIS-005 (plant disease and crop health detection from imagery), PRED-027 (flood/hydrology) and PRED-044 (livestock sensing) share sensor infrastructure on mixed farms.
- Solver and energy: OPT-011 (storage dispatch) combines when pumping runs on PV and battery; OPT-018 supports natural-language what-if analysis on the plan.
- Hardware: one edge gateway tier (HW1) per farm, one cloud geospatial pipeline shared across many fields.
Effort, cost & KPIs
- Path A: weeks to onboard; subscription per hectare plus sensors and ISOBUS-compatible hardware; variable-rate spreaders may need retrofit.
- Path B: 3–9 months for a production service including data connectors and agronomy review.
- Path C: one season to calibrate crop models per region; low licence cost, high agronomy effort.
- Cost drivers: sensor installation and maintenance, connectivity in rural areas, agronomist validation, machinery compatibility.
- KPIs: water use per tonne of yield, nitrogen use efficiency and N surplus (kg/ha), yield and quality versus control fields, energy for pumping, share of recommendations accepted, nutrient balance compliance.
People & change
The owner decides on scope and budget; the farm or area manager is the process owner; machine operators and employees are the daily users and know the fields best, so involve them in validating maps and recommendations. IT or the machinery dealer handles interfaces; suppliers (machine makers, platform vendors, contractors) must be asked about data access and rights. Works council: normally no, because small farms rarely have one; where one exists, machine telemetry that records operator behaviour can fall under §87(1) no. 6 BetrVG, so inform it under §90(1) no. 3 and avoid operator-level evaluation. Data protection: field data of a person can be personal data; sign an AVV (Art. 28 GDPR) with the platform and involve the DPO where one exists. A DPIA is usually not required.
AI literacy: support staff with short hands-on sessions on how recommendations are derived and where they fail (AI Act Art. 4). Pilot on two or three fields with a control strip and compare against the usual practice. Typical concern: "the algorithm knows my soil less well than I do". Address it by treating output as a proposal; the farmer decides and can override, and variable-rate maps are checked before application.
See GUIDE-009, GUIDE-008 and GUIDE-012.
Risks, regulation & compliance
- EU AI Act: typically minimal risk; not in Annex III. Closed-loop control of safety-relevant infrastructure (for example public water supply) would change the assessment. AI literacy (Art. 4) applies. Art. 5 prohibitions are not relevant.
- GDPR: mostly non-personal, but farm data linked to a named farmer, plots, subsidies and machine telemetry are personal data for sole proprietors; use contracts and DPAs, prefer EU hosting.
- EU Data Act (applicable since September 2025): gives users of connected machinery access to and control over generated data and eases switching vendors; check portability of machine and sensor data. The EU Code of Conduct on agricultural data sharing guides contracts [6].
- Environmental rules: Nitrates Directive, Water Framework Directive, national fertiliser ordinances and CAP conditionality; recommendations must never exceed legal caps; keep an auditable record of what the model suggested and what was applied.
- Risks and mitigations: over-reliance on model advice (agronomist review, field checks), vendor lock-in (open formats ISOXML, agrirouter), sensor failures (plausibility checks, fallback to ET-based schedule), cyber-security of irrigation controllers (network segmentation, updates).
Getting started
- Collect field boundaries, soil data, past yields and fertiliser/irrigation logs; fix baseline KPIs.
- Pick one crop and 3–5 pilot fields with matched control fields.
- Install soil moisture probes and a weather station (or use a dense public network) and connect Sentinel-2 indices through a Path A platform or Copernicus services.
- Run recommendations in advisory mode for one season; compare with farmer practice.
- Enable variable-rate or automated irrigation only inside permit and nutrient limits; document it.
Sources
- European Commission — Nitrates Directive — https://environment.ec.europa.eu/topics/water/nitrates_en
- European Commission — CAP and farm sustainability tools — https://agriculture.ec.europa.eu/common-agricultural-policy_en
- FAO — AquaCrop crop model — https://www.fao.org/aquacrop/en/
- Copernicus Data Space Ecosystem — https://dataspace.copernicus.eu/
- openEO — open API for Earth observation processing — https://openeo.org/
- EU Code of Conduct on agricultural data sharing by contractual agreement — https://copa-cogeca.eu/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2023/2854 (Data Act) — https://eur-lex.europa.eu/eli/reg/2023/2854/oj