The Agronomic
Knowledge Graph

A semantic triple-store connecting 140,000+ agronomic relationships — from soil microbiota to satellite indices — powering natural language queries, predictive alerts, and AEO-optimised reporting.

Triple-Store Architecture Natural Language Queries AEO / AI-Ready RDF / OWL Ontology
140K+
Agronomic Triples
38
Ontology Classes
<200ms
Query Latency
growsIn affectedBy hasSensor triggers sequestersCarbon requestsIntervention CropField Nebraska · 847ha Satellite Sentinel-2 Pest Soybean Aphid IoTSensor Soil·Moisture Alert Severity 3 Carbon Pool·tCO₂e Intervention Prescription
Foundation

What Is a Knowledge Graph?

VersaAgro's knowledge graph models the entire agronomic domain as a network of entities and relationships — enabling machines to reason over data the way experienced agronomists do.

Entities (Nodes)

Every agronomic object — a crop variety, soil horizon, weather event, IoT sensor, pest species, or intervention — is represented as a named entity node with typed attributes and a unique IRI identifier.

Relationships (Edges)

Named predicates define how entities relate: a CropField affectedBy Drought, a SoilSensor locatedIn ManagementZone, or an Alert triggers Intervention. Every edge has provenance metadata.

Inference & Reasoning

OWL-based ontology rules enable automatic inference: if soil moisture <20% AND crop stage = reproductive, the reasoner infers irrigation urgency = HIGH — without explicit programming for every crop × condition combination.

Architecture

RDF Triple-Store:
Subject → Predicate → Object

Every fact in VersaAgro is stored as an RDF triple. This standardised W3C format allows complex SPARQL queries, federated data integration with external registries (Verra VCS, GBIF species database, ESA Copernicus), and full auditability of every agronomic inference.

The table on the right shows a sample of live triples from the Nebraska soybean pilot — 847 ha, 12 management zones, covering crop physiology, satellite observations, IoT telemetry, and carbon accounting.

Subject → Predicate → Object

Each triple is immutable once written. Updates create new versioned triples with timestamps and provenance IRIs, forming a complete temporal audit log.

Subject Predicate Object
Crop & Field Identity
NE-Field-847 rdf:type va:CropField
NE-Field-847 va:cultivates va:GlycinMax_Pioneer_P39T69R
NE-Field-847 va:areaHectares 847.0^^xsd:decimal
NE-Field-847 geo:hasGeometry va:NE847_Polygon_GeoJSON
Satellite Observations
S2-Obs-2026-04-18-NE847 va:observes NE-Field-847
S2-Obs-2026-04-18-NE847 va:ndviMean 0.73^^xsd:float
S2-Obs-2026-04-18-NE847 va:cloudCoverPct 4.2^^xsd:float
IoT Telemetry
SENSOR-NE847-Z06-SM rdf:type va:SoilMoistureSensor
SENSOR-NE847-Z06-SM va:locatedIn va:Zone06_NE847
SENSOR-NE847-Z06-SM va:latestReading 17.4^^xsd:float
Pest & Disease
va:AphisGlycines va:threatensSpecies va:GlycinMax
va:AphisGlycines va:thresholdPopulation 250^^xsd:integer
ALERT-NE847-2026-05-03 va:causedBy va:AphisGlycines
Carbon Accounting
NE-Field-847 va:hasCarbonPool va:NE847_CarbonPool_2026
va:NE847_CarbonPool_2026 va:estimatedTCO2e 312.4^^xsd:decimal
va:NE847_CarbonPool_2026 va:verificationStandard va:VerraVCS_VM0042
Ontology

38 Entity Classes, One Unified Schema

The VersaAgro Agronomic Entity Ontology (AEO) defines every concept, attribute, and relationship type used across the platform — enabling cross-domain reasoning and interoperability.

Crop Entities
9 classes · va:CropOntology namespace
CropSpecies CropVariety GrowthStage PhenoPhase YieldEstimate CropField ManagementZone PlantingEvent HarvestEvent
9 classes · 42 object properties · 28 data properties
Soil & Terrain
6 classes · va:SoilOntology namespace
SoilHorizon SoilTexture SoilMoisture SoilOrganicCarbon ElevationBand DrainageClass
6 classes · 18 object properties · 34 data properties
Weather & Climate
5 classes · va:ClimateOntology namespace
WeatherObservation DroughtEvent FrostEvent PrecipitationRecord GDDAccumulation
5 classes · 14 object properties · 21 data properties
Pest & Disease
5 classes · va:BioticOntology namespace
PestSpecies DiseasePathogen RiskScore InfestationZone TreatmentThreshold
5 classes · 16 object properties · 19 data properties
IoT & Sensors
5 classes · W3C SSN namespace
SoilMoistureSensor TemperatureSensor NDVIObservation SensorReading AlertThreshold
5 classes · 22 object properties · 30 data properties
Carbon & Sustainability
4 classes · va:CarbonOntology namespace
CarbonPool SequestrationEvent EmissionFactor VerificationStandard
4 classes · 12 object properties · 18 data properties
Interventions
4 classes · va:AgronOntology namespace
IrrigationSchedule FertiliserApplication PesticideApplication PrescriptionMap
4 classes · 14 object properties · 22 data properties
Spatial & Temporal
3 classes · GeoSPARQL / OWL-Time
GeoPolygon SpatialFeature TemporalInterval
3 classes · 8 object properties · 12 data properties
Methodology

Agronomic Entity Optimisation (AEO)

AEO is VersaAgro's approach to structuring agronomic knowledge for AI consumption — analogous to SEO for search engines, but designed so that large language models, RAG pipelines, and AI agents can precisely retrieve, cite, and reason over agronomic facts.

Every entity is tagged with schema.org-compatible markup, SPARQL-accessible IRIs, and human-readable summaries — making the knowledge graph simultaneously machine-queryable and AI-grounded.

1
Entity Extraction & Normalisation
Raw data from sensors, satellite APIs, and weather feeds is parsed into typed entity nodes. Crop variety names are resolved against the CGIAR GRIN database; pest names validated against GBIF taxonomy.
2
Relationship Inference
The OWL reasoner infers implicit relationships — e.g., a zone with NDVI < 0.45 AND soil moisture < 20% is automatically linked to va:DroughtStress and assigned to the relevant alert class.
3
Semantic Embedding & Indexing
Each entity summary is vectorised using an agronomy-domain fine-tuned embedding model. Vectors are indexed in a dedicated HNSW store for sub-200ms semantic retrieval across 140K+ triples.
4
Natural Language Query Interface
User queries are parsed with intent classification, mapped to SPARQL templates, and executed against the triple store. Results are re-ranked by freshness, geospatial relevance, and severity — then formatted as a cited, evidence-backed response.
AEO Query Pipeline
Natural Language Input
"Which zones show aphid pressure above threshold?"
Intent Classification
entity=Pest · predicate=RiskScore · filter=threshold
SPARQL Query Generation
SELECT ?zone ?score WHERE { ?zone va:riskScore ?score FILTER(?score > 65) }
Triple-Store Execution
140K+ triples · GeoSPARQL spatial filter · <200ms
Re-ranking & Evidence Linking
Severity · Freshness · Spatial proximity · Source IRI
Cited Natural Language Response
Zones 4, 7, 11 · Scores 82, 74, 68 · 3 sources cited
Live Examples

Natural Language Queries in Action

The AI Query box in the VersaAgro dashboard translates plain-English questions into SPARQL against the live knowledge graph. Here are four real example queries from beta pilot operators.

"Which zones in Nebraska show early signs of Soybean Aphid pressure above economic threshold?"
Pest
VersaAgro AI Response
Zones 4, 7, and 11 currently exceed the 250 aphid/plant economic threshold. Zone 7 has the highest risk score (82/100), driven by 3 consecutive days of temperatures in the 21–27°C optimal range for Aphis glycines combined with a 6-day rainfall gap. Recommended action: scout Zones 7 and 4 within 48h; consider foliar application trigger at 500 aphids/plant.
Zone 7: Risk Score 82 · Last sensor ping 2h ago · NDVI drop −0.06 vs. 7-day avg
Zone 4: Risk Score 74 · Aphid pressure index elevated since 2026-04-29
Zone 11: Risk Score 68 · Threshold recently crossed · Monitoring recommended
IoT·SENSOR-NE847-Z07 S2-Obs-2026-05-01 va:AphisGlycines
"What is our total verified carbon sequestration across all active pilot fields this season?"
Carbon
VersaAgro AI Response
2,140 tCO₂e sequestration has been modelled across all 6 active pilot fields (3,579 ha combined) for the 2025–26 season. Of this, 1,847 tCO₂e is currently eligible for Verra VCS VM0042 verification. At the current EEX spot price of $68.50/tCO₂e, this represents $126,520 in pending verified carbon credit revenue.
Nebraska Soybean (847 ha): 312 tCO₂e · Verification status: In progress
Kansas Wheat (1,240 ha): 890 tCO₂e · Verification status: Eligible
Tanzania Coffee (560 ha): 468 tCO₂e · Verification status: Submitted
Verra VCS VM0042 va:CarbonPool_2026 EEX Carbon API
"Alert me if any field's soil moisture drops below critical field capacity in the next 72 hours."
Water
VersaAgro AI Response
2 fields are projected to breach critical field capacity thresholds within the 72-hour window. Jalisco Blue Agave (Zone 3) is forecast to reach 16.8% volumetric water content by 2026-05-14 06:00 — below the 18% trigger for agave drought stress. Bolivia Hay (Zone 2) shows 19.1%, marginally above the 19% Pennisetum purpureum wilting point.
Jalisco-Z3: 17.8% now → 16.8% in 72h · Forecast confidence 84% · Irrigation recommended
Bolivia-Z2: 19.6% now → 19.1% in 72h · Monitor only · No precipitation forecast
Open-Meteo Forecast SENSOR-MX220-Z03 va:AgaveTequilana
"Compare yield potential across our three North American grain fields and rank by expected ROI."
Yield
VersaAgro AI Response
Ranked by expected ROI, Kansas Winter Wheat (1,240 ha) leads at an estimated 4.8 t/ha · $312/ha net margin, followed by Nebraska Soybean (847 ha) at 3.6 t/ha · $284/ha, and Iowa Corn (312 ha) at 11.2 t/ha · $198/ha. Kansas wheat benefits from above-average NDVI uniformity (CV 0.08) and optimal GDD accumulation of 1,240°C-days since emergence.
KS Wheat: NDVI 0.81 · GDD 1,240 · Uniformity CV 0.08 · ROI rank #1
NE Soybean: NDVI 0.73 · GDD 912 · 2 active pest alerts · ROI rank #2
IA Corn: NDVI 0.68 · GDD 788 · Irrigation deficit 22mm · ROI rank #3
Sentinel-2 Multi-Date USDA NASS 2025 GDD Model v2.1
AgroLD Integration

Open-Source Knowledge Graph
Source of Truth

VersaAgro's agronomic intelligence is grounded in AgroLD v2 — an INRAE/Montpellier SupAgro SPARQL-queryable RDF knowledge graph integrating 8 major ontologies. Below is our curated entity dataset used for real-time NLP alert enrichment and search attribution.

AgroLD v2
INRAE · Montpellier SupAgro
NCBI Tax · GO · TO · PO · CO · AGROVOC
v2.agrold.org API Docs
CORS-constrained endpoint: The AgroLD SPARQL endpoint (https://v2.agrold.org/sparql) does not send Access-Control-Allow-Origin headers — direct browser fetch() is blocked by Same-Origin Policy. This interface queries a curated local static snapshot of 36 AgroLD entities with canonical IRIs preserved. In production, VersaAgro routes queries through a server-side SPARQL proxy — see the Proxy Architecture reference.

Loading AgroLD entity dataset…

Developer Reference

SPARQL Proxy Architecture

How VersaAgro bridges a CORS-constrained academic SPARQL endpoint to browser-based search, alert enrichment, and NLP pipelines — without exposing API keys or violating Same-Origin Policy.

Production Request Flow
Browser Client
knowledge-graph.html · NLP query input · Alert enrichment UI
VersaAgro API Gateway
POST /api/kg/query · CORS-enabled · Auth validated
NLP → SPARQL Translator
Intent classification · Entity resolution · Template hydration
AgroLD SPARQL Endpoint
https://v2.agrold.org/sparql · Server-to-server (no CORS issue)
VersaAgro Triple Store
Internal sensor + satellite triples · Federated JOIN
Cited Response + Source IRIs
Merged result · AgroLD IRI attribution · <200ms total
Ontologies Federated via AgroLD
Gene Ontology (GO)
Biological processes · Molecular functions · Cellular components
Trait Ontology (TO) + Plant Ontology (PO)
Agronomic trait descriptors · Plant structures & developmental stages
Crop Ontology (CO) + AGROVOC
CGIAR crop-specific traits · FAO multilingual agricultural thesaurus
NCBI Taxonomy + Gramene
Species taxonomy · Gene-to-trait genomic associations
CORS constraint: browser fetch() blocked at https://v2.agrold.org/sparql
Static Layer (current)
Proxy Route (production)
SPARQL Template
// Static layer — CORS-safe, zero latency // Queries curated data/agrold-entities.json const searchAgroLD = async (query) => { const { entities } = await loadEntities(); return entities.filter(e => e.label.toLowerCase().includes(query) || e.tags.some(t => t.includes(query)) || e.description.toLowerCase().includes(query) ); }; // Each entity preserves canonical IRI for // full provenance: e.iri → AgroLD resource
// Production: browser → VersaAgro proxy // Client call (browser) const res = await fetch('POST /api/kg/query', { body: JSON.stringify({ nlp: query, field_id }) }); // Server proxy (Node.js) const sparql = nlpToSparql(nlp); const agrold = await fetch( `https://v2.agrold.org/sparql?query=${sparql}` ); // server-to-server: no CORS block const local = await queryTripleStore(sparql); return mergeAndCite(agrold, local);
# SPARQL query template — Soybean Aphid traits # Resolved from AgroLD entity IRI PREFIX to: <http://purl.obolibrary.org/obo/TO_> PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> SELECT ?trait ?label WHERE { <https://v2.agrold.org/resource/ Aphis_glycines> to:associated_with ?trait . ?trait rdfs:label ?label . } LIMIT 20
Static Dataset vs. Live SPARQL
Static Layer (current prototype)
36 curated entities · Canonical IRIs preserved · Zero latency · No CORS issues · Demonstrates architecture for investors and product owners
Live SPARQL (production roadmap)
Full AgroLD corpus · 82+ entity types · Federated joins with VersaAgro sensor data · NLP→SPARQL translation pipeline · <200ms response
Use Cases

How the Knowledge Graph Powers the Platform

Every intelligent feature in VersaAgro — from automated alerts to carbon MRV reporting — draws directly from the live triple-store rather than isolated data silos.

Intelligent Alert Prioritisation
The reasoner evaluates 12+ graph properties per zone simultaneously — NDVI trend, soil moisture, weather forecast, pest history, and crop growth stage — to assign severity scores that eliminate false-positive alert fatigue.
67% reduction in false-positive alerts vs. threshold-only systems
Carbon MRV & Audit Trail
Every carbon sequestration estimate is backed by a full provenance chain in the graph: satellite observation IRI → soil sensor reading → GHG Protocol calculation → Verra VCS standard reference — providing an immutable audit trail for third-party verification.
100% auditable provenance per tCO₂e estimate
Zero-Training Agronomic Search
Field managers query the knowledge graph in plain language without learning SPARQL or dashboard navigation. The AEO pipeline resolves agronomy-domain terminology — including local crop names, regional pest variants, and legacy measurement units — into precise graph queries.
<200ms average query response · Zero training required
Prescription Map Generation
The graph traversal engine walks from CropField → ManagementZone → IoTSensor → WeatherForecast → CropStage in a single query to generate variable-rate prescription maps, respecting regulatory constraints linked as named entities in the graph.
12-zone variable-rate prescriptions in <4 seconds
Temporal Trend Analysis
Immutable triple versioning enables historical queries across any time window: "How did Zone 7's NDVI trend compare to the same week last season?" — answered by traversing the temporal graph without any additional data processing.
Full temporal coverage from field registration date
Federated Data Integration
GeoSPARQL federated queries span VersaAgro's internal triple store and external SPARQL endpoints simultaneously — pulling live crop prices from USDA, species threat data from GBIF, and carbon prices from EEX without ETL pipelines.
6 live federated data sources in beta
Knowledge Graph Powers All 5 App Screens
Field Dashboard
Crop Intelligence
Alerts & Actions
Carbon MRV
Access & Roles
View All Screens
Knowledge Extension · Forestry & Arboriculture

Tree Issue Glossary
Diseases · Pests · Nutrient Deficiencies · Abiotic Stresses

A reference compendium of common tree health problems drawn from professional arboricultural practice. Each entry is structured for VersaAgro's ML training pipeline — visual symptom imagery, diagnostic confidence scores, and prescriptive intervention protocols grounded in field-tested arborist knowledge. As farmers report field observations, this dataset feeds our symptom-matching model for precision tree health recommendations.

ML Training Dataset Visual Search Ready Prescriptive RX
J. J. Mauget Co.
Arcadia, CA · Est. 1940
Glossary content adapted from Mauget's professional arborist reference library — a leading authority in micro-injection tree care since 1940. Used under educational fair-use for prototype ML pipeline demonstration.
Prototype Disclaimer
This content mirrors the Mauget glossary for concept demonstration only. VersaAgro intends to license or partner with Mauget for production use. Not for commercial deployment without written permission.
View Original Source
Roadmap: Visual Symptom Search

In the production ML pipeline, farmers will photograph leaf discolouration, bark lesions, or defoliation patterns with their phone camera. VersaAgro's CNN-based visual classifier matches the image against this glossary's symptom profiles and returns a ranked diagnosis with confidence score, treatment protocol, and the relevant ontology IRI — all grounded in the Mauget reference dataset.

📸 Photo → Diagnosis 🧠 CNN Symptom Classifier 💊 Prescriptive RX Output 🌳 36 Arborist-Grade Terms
Beta Pilot — Limited Slots Available

Experience the Knowledge Graph Live

The AI query box in the VersaAgro dashboard runs directly against the live triple-store. Try natural language queries on real Nebraska soybean and Tanzania coffee data.