Smarter detection for global crops: new AI framework tackles cross-domain disease challenges

This breakthrough could significantly boost crop yields and help achieve food security goals.

Crop diseases not only reduce yields but also degrade soil health and disrupt ecosystems. Accurate, early detection is essential to managing these threats, but manual identification is slow, labor-intensive, and often ineffective. Modern approaches using deep learning have made detection faster and more accurate, but models trained in one setting often fail in another due to variations in lighting, resolution, climate, or equipment. These domain shifts cause performance drops, requiring costly data recollection and retraining. Overcoming this cross-domain challenge is vital for real-world deployment.

A study (DOI: 10.1016/j.plaphe.2025.100001) published in Plant Phenomics on 24 February 2025 by Fang Zhou’s team, Sun Yat-sen University, introduces the MGA framework to address the critical issue of feature misalignment between training (source) and testing (target) domains in crop disease datasets.

To evaluate the effectiveness of the proposed Multi-Granularity Alignment (MGA) framework for cross-domain crop disease detection, researchers conducted a series of experiments using the MGA-YOLOv8 model and compared it against four widely used object detection networks—Faster R-CNN, YOLOv5, SSD, and RetinaNet—across three domain-shift scenarios: different regions, environments, and image styles. The experiments used the mean Average Precision (mAP) metric at an IoU threshold of 0.5 to assess detection accuracy, with higher mAP values indicating better performance. A 5-fold cross-validation strategy was employed to ensure robust evaluation, and ablation studies were performed to determine the contribution of each domain adaptation module: omni-scale gated fusion (OGF), category-level (ClD), instance-level (IlD), and pixel-level (PlD) discriminators. In the regional dataset scenario (PlantVillage → CropDis), MGA achieved an mAP of 47.9%, outperforming the baseline model by 15.2% and even surpassing YOLOv5 by 14.8%. In the environment-shift scenario (PlantVillage → PlantDoc), MGA again led with a 46.8% mAP, whereas competing models remained under 30%. When the domain shift was reversed (PlantDoc → PlantVillage), MGA reached 48.3%, outperforming even the theoretical upper limit (Baseline_t). Image scale sensitivity tests further confirmed the robustness of PlD and the advantage of multi-granularity adaptation across variable resolutions. In the third scenario, involving style-transferred datasets simulating varied weather conditions, MGA achieved a remarkable 49.2% mAP, outperforming the baseline by up to 18.4%. Visual feature analysis using t-SNE demonstrated that MGA effectively aligned features between domains, reducing discrepancies and enhancing generalization. These results strongly validate MGA’s superiority in cross-domain adaptability, significantly improving detection accuracy across varied crop disease datasets and real-world conditions.

The MGA framework offers a powerful tool for practical agricultural management. By enabling disease detection models to generalize across regions and environments, MGA supports early diagnosis, targeted intervention, and increased crop productivity. Its performance across diverse conditions makes it especially valuable for developing countries, where annotated datasets are limited but disease prevention is critical to food security. Beyond agriculture, MGA shows promise in other domains such as autonomous driving, where models trained on daytime images could detect features in nighttime conditions. The researchers envision deploying MGA in drones or agricultural robots and integrating it into smart farm management systems.

###

References

DOI

10.1016/j.plaphe.2025.100001

Original Source URL

https://doi.org/10.1016/j.plaphe.2025.100001

About Plant Phenomics

Science Partner Journal Plant Phenomics is an online-only Open Access journal published in affiliation with the State Key Laboratory of Crop Genetics & Germplasm Enhancement, Nanjing Agricultural University (NAU) and distributed by the American Association for the Advancement of Science (AAAS). Like all partners participating in the Science Partner Journal program, Plant Phenomics is editorially independent from the Science family of journals. Editorial decisions and scientific activities pursued by the journal's Editorial Board are made independently, based on scientific merit and adhering to the highest standards for accurate and ethical promotion of science. These decisions and activities are in no way influenced by the financial support of NAU, NAU administration, or any other institutions and sponsors. The Editorial Board is solely responsible for all content published in the journal. To learn more about the Science Partner Journal program, visit the SPJ program homepage.