ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Game Engine Meets VLM: A synthetic-driven open-vocabulary diagnosis framework for tomato disease monitoring

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600469.(doi:10.13031/aim.202600469)
Authors:   Pengyao Xie, Xingjian Li, Weilong He, Inga Meadows, Lirong Xiang
Keywords:   Contrastive learning, domain adaptation, fine-grained classification, plant disease diagnosis, sim-to-real transfer, vision-language model, zero-shot inference.

Abstract. Accurate plant disease characterization is pivotal for precision agriculture, yet current deep learning approaches are constrained by costly real-world annotations, data scarcity, and the inherent “closed-set” limitations of traditional classifiers. Moreover, existing systems typically struggle to simultaneously resolve local lesion details and global canopy distribution patterns under zero-shot settings. To address these challenges, we present PD-CLIP, a plant disease contrastive language-image pre-training framework. Utilizing Unreal Engine 5 3D physical simulation, we constructed a synthetic dataset encompassing 10 tomato foliar disease categories across 7 severity levels and 4 spatial spread types, which were automatically annotated with fine-grained semantic descriptions by multimodal large language models. Our model employs a dual-stream vision architecture and a parameter-efficient fine-tuning strategy via low-rank adaptation to map multi-scale visual and textual features into independent diagnostic subspaces. To overcome sim-to-real distribution shifts, we developed a composite domain adaptation module integrating Fourier transform-based domain adaptation, real-world anchor data guidance, and entropy minimization. Evaluated through a leave-one-out cross-validation strategy, PD-CLIP demonstrated strong cross-domain zero-shot capabilities. On unseen real-world field images, the domain-adapted model achieved a disease type classification accuracy of 0.575, a spread type identification accuracy of 0.992, and severity grading with a mean absolute error of 0.818. By effectively characterizing disease types, severity, and spread types without requiring explicit real-world annotations for the target pathogens, this approach bridges the sim-to-real gap, offering an open-vocabulary solution for crop health monitoring.

(Download PDF)    (Export to EndNotes)