Below is a list of published work from the GEMINI project
Lundqvist, L., Ranario, E., Kamangir, H., Yun, H., Diepenbrock, C., Bailey, B.N., Earles, J.M.
Does Your VFM Speak Plant? The Botanical Grammar of Vision Foundation Models for Object Detection.
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026.
https://arxiv.org/abs/2604.09920
Summary: Vision foundation models can detect objects zero-shot without task-specific training data, but in complex agricultural scenes their accuracy depends heavily on how the text prompt is written. The paper presents a systematic prompt-optimization framework that substantially improves zero-shot detection in agricultural settings.
Uyehara, I.K., Berlingeri, J.M., Lo, S., Riggs, M., Desai, R., Rizzo, K.T., Earles, J.M., Diepenbrock, C.H., Bailey, B.N.
Smartphone-based proximal sensing as an accessible platform for improved physiological phenotyping.
The Plant Phenome Journal, Volume 9, Issue 1, 2026, e70107.
https://doi.org/10.1002/ppj2.70107
Summary: The authors tested whether a smartphone carried through a field could deliver accurate physiological traits cheaply, using image time series of 10 cowpea genotypes. The sensed traits matched labor-intensive manual measurements (R² ≥ 0.78), showed tight coupling between vegetative and reproductive phenology, and pointed to maximum pod length and flowering lag as traits linked to pod harvest index and yield.
Yun, H., Uyehara, I.K., Droutsas, I., Ranario, E., Diepenbrock, C.H., Bailey, B.N., Earles, J.M.
A vision language model for generating XML-based organ-level plant architecture representations of cowpea from simulated images.
Frontiers in Artificial Intelligence, Volume 9, 2026, 1844338.
https://doi.org/10.3389/frai.2026.1844338
Summary: The model was trained on synthetic cowpea images generated with the Helios 3D simulator, with each plant’s architecture stored in XML. A custom tokenizer turned that XML into sequences a vision language model could learn to predict from images. It estimated whole-plant traits that depend on the hidden 3D structure, such as leaf count and leaf area, with lower error than feature-regression methods.
Ranario, E., Mayanja, I., Yun, H., Bailey, B.N., Earles, J.M.
Thermal image segmentation in weedy fields via synthetic RGB-trained models and GAN-based cross-modality alignment.
Plant Phenomics, Volume 8, Issue 3, 2026, 100214.
https://doi.org/10.1016/j.plaphe.2026.100214
Summary: The study tackles plant segmentation in field thermal images, where weeds look similar to crop plants and frequently overlap them. Models were trained on more than 1,000 synthetic RGB images plus as few as 20 labeled field images, and masks were moved onto the thermal images through GAN-based RGB-to-thermal alignment.
Uyehara, I.K., Yun, H., Rizzo, K.T., Ranario, E., Earles, M.
AgRowStitch: A Leaf-Scale Image Stitching Pipeline for Ground-Based Agricultural Images.
Applied Engineering in Agriculture, 42(3), 307–317, 2026.
https://doi.org/10.13031/aea.16650
Summary: The paper presents an open-source pipeline for stitching close-range images along a crop row without needing GPS or other extra data. It was tested on 72 m rows imaged by two agricultural robots and a hand-carried camera. The resulting mosaics located real-world positions with a mean absolute error of 20 cm.
Yun, H., Uyehara, I.K., Ranario, E., Lundqvist, L., Diepenbrock, C.H., Bailey, B.N., Earles, J.M.
Using Vision Language Foundation Models to Generate Plant Simulation Configurations via In-Context Learning.
arXiv preprint, 2026.
https://doi.org/10.48550/arXiv.2603.08930
Summary: Open-source vision language models (Gemma 3 and Qwen3-VL) were asked to generate plant simulation settings as JSON directly from drone images. They handled structural information such as plant count and sun azimuth reasonably well, but often fell back on dataset averages when the image gave too few cues. Testing on real drone orthophotos showed how much they relied on context versus visual reasoning.
Bailey, B.N.
A generalized framework for procedural generation of three-dimensional static and dynamic plant model geometries.
arXiv technical report, 2025.
https://doi.org/10.48550/arXiv.2512.17966
Summary: The report describes a new framework, built into the Helios modeling system, for procedurally generating 3D plant geometries that can be static or change over time.
Ranario, E., Lundqvist, L., Yun, H., Bailey, B.N., Earles, J.M.
AGILE: A Diffusion-Based Attention-Guided Image and Label Translation for Efficient Cross-Domain Plant Trait Identification.
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 5392–5401.
https://arxiv.org/abs/2503.22019
Summary: AGILE uses pretrained diffusion models and public agricultural datasets to translate images between domains while keeping object labels accurate. It does this by optimizing text embeddings and guiding attention maps during denoising. The translated images improved object detection in the target domain.
Mayanja, I.K., Yun, H., Bailey, B.N.
Automated calibration of stomatal conductance models from thermal imagery by leveraging synthetic images generated from Helios 3D biophysical model simulations.
Journal of Experimental Botany, 2025, eraf420.
https://doi.org/10.1093/jxb/eraf420
Summary: A machine learning model was trained only on synthetic thermal images from Helios and then applied to real thermal images to estimate stomatal conductance and calibrate stomatal model parameters. Compared with independent porometer measurements, predictions reached an R² of 0.7, roughly matching the porometer’s own accuracy. The pipeline separates genotype-specific stomatal traits from environmental effects and allows predictions in new environments.
Berlingeri, J., Fuentes, A., Ranario, E., et al.
Integration of crop modeling and sensing into molecular breeding for nutritional quality and stress tolerance.
Theoretical and Applied Genetics, 138, 205, 2025.
https://doi.org/10.1007/s00122-025-04984-y
Summary: This review argues for combining crop modeling, AI-enabled sensing and molecular breeding so that nutritional quality is bred for alongside yield and stress tolerance. It surveys the current state of these tools in grain legumes, rice and leafy greens. It then outlines how integrating them can improve prediction of breeding values and the study of genotype-by-environment-by-management interactions.
Lei, T., Graefe, J., Mayanja, I.K., Earles, M., Bailey, B.N.
Simulation of Automatically Annotated Visible and Multi-/Hyperspectral Images Using the Helios 3D Plant and Radiative Transfer Modeling Framework.
Plant Phenomics, Volume 6, 2024, 0189.
https://doi.org/10.34133/plantphenomics.0189
Summary: The study extends Helios with radiative transfer modeling to simulate RGB, multi- and hyperspectral, thermal and depth images, each with fully resolved labels. The synthetic images closely resembled real ones, which supports using them to train machine learning models for phenotyping.
Yun, H., Lo, S., Diepenbrock, C.H., Bailey, B.N., Earles, J.M.
VisTA-SR: Improving the Accuracy and Resolution of Low-Cost Thermal Imaging Cameras for Agriculture.
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 5470–5479.
https://arxiv.org/abs/2405.19413
Summary: VisTA-SR combines RGB and thermal images in a deep learning network to improve the temperature accuracy and resolution of low-cost thermal cameras. The work also includes temperature calibration and validation and a paired image dataset.
Mayanja, I.K., Diepenbrock, C.H., Vadez, V., Lei, T., Bailey, B.N.
Practical Considerations and Limitations of Using Leaf and Canopy Temperature Measurements as a Stomatal Conductance Proxy: Sensitivity across Environmental Conditions, Scale, and Sample Size.
Plant Phenomics, Volume 6, 2024, 0169.
https://doi.org/10.34133/plantphenomics.0169
Summary: Using a high-resolution 3D model, the study examined how well leaf temperature tracks stomatal conductance across environments, measurement scales and sample sizes. Leaf temperature responded most to stomatal conductance under high radiation, high vapor pressure deficit and low boundary-layer conductance. Small-scale measurements such as thermocouples were the most sensitive but could need 50–75 samples to tell genotypes apart. Thermal cameras needed fewer samples but picked up low-sensitivity shaded leaves and soil.