AGRIFARM-AI: Intelligent Management System for Climate-Resilient Smart Agriculture

AGRIFARM-AI is an autonomous agricultural management platform integrating hybrid VLC/RF networks, multi-agent reinforcement learning, and agentic AI for precision farming in controlled environments.

---

🚧 Status: Under Active Development

Current phase: Architecture Design

---

Project Overview

AGRIFARM-AI is an ambitious research project developing an Intelligent Management System (IMS) for climate-resilient smart agriculture. The system integrates cutting-edge technologies to enable autonomous precision farming in controlled environments.

Key Innovation Areas

Area	Technology	Impact
Communication	Hybrid VLC/RF Networks	73% RF reduction, crop-safe
Decision Making	Multi-Agent RL (MARL)	Autonomous coordination
Intelligence	Agentic AI + LLM	Natural language farming advice
Privacy	Federated Learning	Cross-border knowledge sharing

---

System Architecture

Core IMS Components

The system extends our proven IMS architecture (ACM MobiCom 2023) with four specialized agricultural agents:

🔄 Agri-IHC (Intelligent Handover Controller)

• Function: Seamless VLC/RF network transitions
• Agricultural Adaptation: VLC-priority in grow zones; RF for mobile robots
• Target: <250ms handover latency

📍 Agri-ETL (Edge Tracking & Localization)

• Function: Federated learning-based positioning
• Agricultural Adaptation: Drone/robot tracking; worker localization via WiFi Doppler
• Target: <30cm accuracy in greenhouse

💧 Agri-ARS (Adaptive Resource Management)

• Function: ML-based resource allocation
• Agricultural Adaptation: Precision irrigation scheduling; sensor prioritization
• Target: 40% water savings, 30% fertilizer reduction

⚡ Agri-EHM (Energy Harvesting Management)

• Function: LED brightness optimization
• Agricultural Adaptation: Dual-purpose grow lights (PAR + VLC)
• Target: 25% energy cost reduction

---

15-Dimensional Agricultural State Space

Our MARL agents operate on an expanded state space validated with agronomic literature:

Environmental Variables (5)

SoilMoisture       [0-100%]     Volumetric water content
AirTemperature     [15-40°C]    Metabolic rate driver
RelativeHumidity   [20-100%]    Disease pressure indicator
SolarRadiation     [0-1000 W/m²] Photosynthesis driver
CO2Concentration   [400-1000 ppm] CEA optimization

Crop State Variables (4)

GrowthStage        [BBCH 0-9]   Seedling → Harvest
LeafAreaIndex      [0-8 m²/m²]  Canopy development
CropStressIndex    [0-1]        NDVI-derived stress
DaysToHarvest      [0-120]      Countdown timer

Soil Variables (3)

SoilpH             [4-9]        Nutrient availability
ElectricalConductivity [0-8 dS/m] Salinity indicator
NPKLevel           [Low/Med/High] Composite nutrient index

Network Variables (3)

RSSIcurrent        [-90 to -30 dBm] Signal strength
APType             [VLC=1, RF=0]    Access point type
SensorLoad         [0-100%]         Network utilization

---

Climate Resilience Causal Chain

Complete pathway from forecast to validated outcome:

┌─────────────────┐
│ 1. Data Collection │ IoT sensors + ERA5 climate data
│    (Real-time)      │ Every 15 minutes
└────────┬────────┘
         ▼
┌─────────────────┐
│ 2. Prediction      │ SEACLID regional model + LLM
│    (48-72 hours)   │ Heat stress, drought, pest risk
└────────┬────────┘
         ▼
┌─────────────────┐
│ 3. Decision        │ LLM + RAG knowledge base
│    Support         │ "Reduce irrigation 20% now"
└────────┬────────┘
         ▼
┌─────────────────┐
│ 4. Action          │ IMS-ARS automated execution
│    Execution       │ OR farmer SMS notification
└────────┬────────┘
         ▼
┌─────────────────┐
│ 5. Outcome         │ Treatment vs. control plot
│    Validation      │ Quantified yield improvement
└─────────────────┘

---

MARL Coordination Mechanism

Agent Interaction Model

Coordination: Centralized Training, Decentralized Execution (CTDE)

Value Decomposition: QMIX algorithm

Convergence: Nash equilibrium via fictitious play (<1000 episodes)

Reward Functions

Agent	Reward Formula
IHC	r = RSSI_quality - 0.1 × handover_cost
ARS	r = yield_proxy - 0.05 × water_cost - 0.1 × energy_cost
EHM	r = energy_stored - 0.2 × light_impact_on_crop

---

Federated Learning Framework (AgriFL)

Privacy-preserving cross-border agricultural intelligence:

Data Schema

Sensor Type	Frequency	Volume/Season
Soil moisture	15 min	50 MB
Temperature/humidity	15 min	20 MB
Camera (pest/disease)	1 hr	500 MB
Weather station	1 hr	10 MB
Total		~580 MB

Privacy Guarantees

• Differential Privacy: ε ≤ 0.5
• Secure Aggregation: No raw data leaves farm
• Cluster-based FL: Tropical (ASEAN) + Temperate (EU)

---

Target KPIs

Metric	Baseline	Target
Yield improvement	-	+15% under climate stress
Water use efficiency	-	+40%
Nitrogen use efficiency	-	+30%
Pest/disease reduction	-	-50%
Energy consumption	-	-25%

---

Technology Stack

Hardware

• Edge Gateways: Jetson Nano (<50ms inference)
• VLC Platform: OpenVLC 1.3 + BeagleBone Black
• Sensors: Low-cost nodes (€50/unit vs. €200+ commercial)

Software

• MARL Framework: QMIX with PyTorch
• Foundation Model: LLaMA-3.1-8B (5M agricultural tokens)
• FL Framework: AgriFL with Byzantine-robust aggregation
• Climate APIs: ERA5, SEACLID

---

Development Timeline

2026        2027        2028        2029
  │           │           │           │
  ├──Phase 1──┼──Phase 2──┼──Phase 3──┼──Phase 4──┤
  │           │           │           │           │
  │ Design    │ Integration│ Validation│ Transfer  │
  │ Farmer    │ MARL Train │ 4 Trials │ Open-Source│
  │ Co-Design │ Edge AI    │ Impact   │ Commercial │
  └───────────┴───────────┴──────────┴───────────┘

---

Publications & Resources

Related Publications

• ACM MobiCom 2023: IMS Architecture for Hybrid VLC/RF
• LNEE 2024: Offline RL for Agricultural IoT (90.4% success)
• IEEE ICCC: VLC/WiFi MAC Optimization

Links

• 📂 GitHub: kotobuki09/AGRIFARM-AI
• 📄 Documentation: Under development

---

🌱 Vision: Every smallholder farmer equipped with AI-powered climate resilience, accessible through open-source technology and natural language interfaces.

---

👨‍💼 Author

NGÔ TRUNG KIÊN 🌐 kngo.netlify.app 📧 kiennt@hsb.edu.vn 🏫 Hanoi School of Business and Management (HSB) 📱 Faculty: Non-Traditional Security

---