top of page
Search

Automated Phenotyping and Spatial Data Analysis in Production Greenhouses

  • Writer: Vratislav Beneš
    Vratislav Beneš
  • 4 hours ago
  • 3 min read

In modern precision agronomy, Crop Registration remains the industry standard for monitoring plant health. However, traditional manual sampling is typically limited to a representative subset (<0.1% of total population). This heuristic approach is inherently prone to sampling bias and fails to capture the high-degree spatial variability of greenhouse microclimates.

The FRAVEBOT platform shifts the paradigm from sampling to census-scale data acquisition. Our objective is to generate a continuous, high-resolution biometric time-series for every individual plant within the facility.


1. Mechanical Stability and Data Acquisition

The fidelity of digital measurements taken in motion is directly correlated to the mitigation of mechanical vibrations.

  • Chassis Dynamics: The platform weighs 230 kg with a center of gravity situated 15 cm above the rail level. This high-inertia, low-CoG configuration minimizes oscillations in the roll and pitch axes, ensuring image sharpness at linear velocities of up to 2 m/s.

  • Sensor Configuration: We utilize an array of 10 fixed camera modules per side with vertical overlap. This multi-tier redundancy enables the reconstruction of the full vertical plant profile without the need for active mechanical positioning of the sensors, significantly increasing the system's MTBF (Mean Time Between Failures).


2. Localization and "Time-Scale Analysis"

Validating biometric changes over time—such as daily delta in apex thickness—requires absolute spatial consistency.

  1. Sensor Fusion: We integrate RTK-GNSS (centimeter-level global accuracy) with high-speed incremental wheel encoders (millimeter-level local resolution).

  2. World Frame Alignment: Measurements are anchored to a fixed coordinate system of the greenhouse. This allows for Time-Scale Analysis: the precise comparison of the same biological subject across different time windows by eliminating spatial drift.


3. Data Management and Edge Architecture

High-resolution acquisition across 20 cameras generates a massive data flow of up to 18 TB per 24 hours.

  • On-board Computing: The NVIDIA DGX Spark performs primary image analysis and sequence compression directly on the platform to reduce the transmission payload.

  • 60GHz Wireless Backbone: Data transfer to the central greenhouse hub is handled via 60GHz peer-to-peer units (MikroTik). This provides near-fiber throughput, bypassing the congestion common in 2.4/5GHz industrial environments.


4. Current Biometric Metrics (Vegetative)

The system currently validates and exports the following parameters for tomato, pepper, and cucumber crops:

  • Apex Thickness: Diameter of the stem in the terminal section (a critical indicator of the vegetative/generative balance).

  • Internodal Lengths: Distance between nodes to analyze growth dynamics and vigor.

  • Height Profile: Continuous vertical growth tracking across the entire facility.


5. Generative Metrics: Flower Counting and Classification

The next phase of our deployment focuses on Generative Performance. We are implementing automated bloom detection to track the reproductive potential of the crop:

  • Automated Counting: Identifying every flower cluster (truss) and individual bloom across the entire greenhouse.

  • Phenological Classification: Classifying flowers based on their developmental stage (e.g., bud, anthesis/full bloom, post-anthesis/wilting).

  • Pollination Success Tracking: By monitoring the transition from flower to fruit set, we provide growers with real-time data on pollination efficiency and potential fruit load.


6. R&D Roadmap: Predicting LAI using NeRF

Measuring the Leaf Area Index (LAI) in high-density environments is complex due to extreme occlusion. Our 2026 roadmap involves a predictive LAI model:

  • 3D Volumetric Reconstruction: We utilize NeRF (Neural Radiance Fields) to synthesize 3D scenes from stereo camera data, modeling complex foliage structures.

  • Ground Truth Correlation: The model is trained to find correlations between the robot’s 3D volumetric data and "ground truth" data manually collected by human experts.

  • Statistical Estimation: The system will provide a high-precision statistical estimate of total LAI based on the 3D plant architecture.


7. Agronomic Value and Yield Forecasting

The ultimate goal of census-scale monitoring is a Yield Forecast with a 6+ week horizon. By combining vegetative vigor (apex thickness) with generative data (flower classification), we enable:

  • Local Anomaly Detection: Real-time identification of stress zones.

  • Standardized Crop Registration: Replacing manual logs with automated, objective reports.

  • Precision Logistics: Integrating flower-stage data into growth models allows for the most accurate harvest planning in the industry.

 
 
 
bottom of page