README.md

metadata

license: apache-2.0
task_categories:
  - image-classification
  - object-detection
  - image-segmentation
language:
  - en
pretty_name: VLM Plant Sim
size_categories:
  - 10K<n<100K
tags:
  - plant-phenotyping
  - digital-twin
  - agriculture
  - synthetic
  - drone
  - remote-sensing
  - in-context-learning
configs:
  - config_name: default
    data_files:
      - split: synthetic
        path: manifests/synthetic.jsonl
      - split: real
        path: manifests/real.jsonl
      - split: initial_plant
        path: manifests/initial_plant.jsonl

VLM Plant Sim

Dataset for: Using Vision Language Foundation Models to Generate Plant Simulation Configurations via In-Context Learning

• 🌐 Project Page: https://heesup.github.io/projects/vlm-plant-sim/
• 📄 arXiv Paper: https://arxiv.org/abs/2603.08930
• 💻 Code: https://github.com/GEMINI-Breeding/Image2PlantArchitecture_v2

Abstract

This paper introduces a synthetic benchmark to evaluate the performance of vision language models (VLMs) in generating plant simulation configurations for digital twins. While functional-structural plant models (FSPMs) are useful tools for simulating biophysical processes in agricultural environments, their high complexity and low throughput create bottlenecks for deployment at scale.

We propose a novel approach that leverages state-of-the-art open-source VLMs — Gemma 3 and Qwen3-VL — to directly generate simulation parameters in JSON format from drone-based remote sensing images. Using a synthetic cowpea plot dataset generated via the Helios 3D procedural plant generation library, we tested five in-context learning methods and evaluated the models across three categories: JSON integrity, geometric evaluations, and biophysical evaluations.

Our results show that while VLMs can interpret structural metadata and estimate parameters like plant count and sun azimuth, they often exhibit performance degradation due to contextual bias or rely on dataset means when visual cues are insufficient. To the best of our knowledge, this is the first study to utilize VLMs to generate structural JSON configurations for plant simulations, providing a scalable framework for reconstructing 3D plots for digital twins in agriculture.

Authors

Heesup Yun, Isaac Kazuo Uyehara, Earl Ranario, Lars Lundqvist, Christine H. Diepenbrock, Brian N. Bailey, J. Mason Earles — University of California, Davis

Splits

Split	Count	Description
synthetic	2797	HELIOS-rendered cowpea plots (10/30/50/70/90 DAP) with full sidecar annotations
real	560	Real drone orthophoto patches (PNG only)
initial_plant	224	Annotated real-world field patches with bounding box labels (plots 1-16)

Data Sources

• synthetic: Rendered via Helios 3D Plant & Environmental Modeling Framework from 2025 Davis dataset generation pipeline (HELIOS_20260215). Each record includes JPEG image and sidecar files: scene JSON, camera JSON, params JSON, bounding box TXT, class label TXT, instance masks JSON, and per-plant XML architecture files.
• real: Orthophoto patches extracted from a real drone survey of cowpea field plots at UC Davis (2025). PNG images only; evaluation method outputs excluded.
• initial_plant: Manually annotated field patches (CVAT v5.0.1 format) with plant bounding box rectangles at [dap=10]. Only plots 1-16 contain JSON annotations paired with PNG images.

Notes

• Real split intentionally excludes method outputs (*.json and *.md) from VLM evaluation artifacts.
• Initial-plant split includes only samples where both *_orig.json and *_orig.png exist, covering field plots 1-16.
• Synthetic records include sidecar references (scene/camera/params/boxes/classes/masks/xml) in manifest files.

Manifest Files

• manifests/synthetic.{csv,jsonl}
• manifests/real.{csv,jsonl}
• manifests/initial_plant.{csv,jsonl}
• summary.json

BibTeX

@misc{yun2026usingvisionlanguagefoundation,
      title={Using Vision Language Foundation Models to Generate Plant Simulation Configurations via In-Context Learning},
      author={Heesup Yun and Isaac Kazuo Uyehara and Earl Ranario and Lars Lundqvist and Christine H. Diepenbrock and Brian N. Bailey and J. Mason Earles},
      year={2026},
      eprint={2603.08930},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.08930},
}