Leaf-DETR sharpens dense leaf detection in crowded kiwifruit canopies
image: Motivation of leaf-deter. (a) The importance of leaves and the role of detecting leaves. (b) Differences in the DETRs architecture, Leaf-DETR has higher attention to the leaf edges and lower matching cost. (c) An example of linking object detection with downstream tasks, including disease classification and yield prediction.
Credit: Plant Phenomics
Accurate leaf monitoring is essential for smart agriculture because leaves reveal photosynthesis, growth, stress, and disease status. Dense leaf detection is especially difficult: overlapping leaves, similar textures, and heavy occlusion cause missed detections and unstable training. To address this, researchers developed Leaf-DETR, an end-to-end detection transformer (DETR) framework with a Progressive Feature Fusion Pyramid Network (P-FPN) and a Crowded Query Refinement (CQR) strategy. They also built a large dense kiwifruit leaf dataset containing 1,696 images and 85,375 annotated boxes. On this benchmark, Leaf-DETR improved mean average precision at intersection-over-union 0.50 (mAP@50) by 1% and average recall at 300 detections (AR@300) by 1.4% over its baseline, while converging quickly and generalizing to surveillance images and other crops.
Crop growth monitoring depends on timely, reliable information about leaves, the primary photosynthetic organs. Dense leaf detection supports leaf morphology analysis, disease identification, and yield estimation, but field scenes are far from simple. Leaves overlap heavily, share similar colors and textures, and often occlude one another, so generic detectors miss partially hidden leaves. Existing agricultural datasets mainly target single-leaf disease or prominent, unoccluded leaves, limiting real-world applicability. Traditional modular detectors and generic feature extractors also struggle with feature confusion, while one-to-one matching in detection transformers can become computationally expensive as crowded candidate boxes multiply and network convergence slows. Given these challenges, there is a pressing need for dedicated datasets and detection frameworks tailored to dense leaf scenarios.
Researchers from Guizhou University, Guangdong University of Technology, Suningfia Agriculture Group Co., Ltd., and Wayamba University published (DOI: DOI 10.1016/j.plapbe.2026.100182) the study in Plant Phenomics in 2026. The team introduced Leaf-DETR (detection transformer), a dense leaf detection framework combining a Progressive Feature Fusion Pyramid Network (P-FPN) with a Crowded Query Refinement (CQR) strategy. The work also presents a dense kiwifruit leaf dataset with 1,696 images and 85,375 boxes.
The technical core of Leaf-DETR is two-fold. The Progressive Feature Fusion Pyramid Network (P-FPN) fuses high-level spatial information with low-level texture detail through a top-down, progressive path. Its Adaptive Feature Aggregation (AFA) module uses dual spatial and channel branches to capture leaf position and edge cues, helping the model separate overlapping leaves that look alike. The Crowded Query Refinement (CQR) strategy adds a parallel training branch that generates extra queries. Class-agnostic non-maximum suppression (NMS) removes redundant original queries, reducing Hungarian matching cost, while a Top-K Hungarian matcher applies one-to-many matching to the remaining extra queries. This expands trainable positive samples, accelerates convergence, and is disabled at inference to avoid extra cost. An improved jointly trained auxiliary head further supplies diverse positive samples by assigning random confidence scores and applying NMS. On the kiwifruit benchmark, Leaf-DETR reached 93.2% mAP@50 and 79.5% AR@300, outperforming the Collaborative Hybrid Assignments Training (CO-DETR) baseline. Training stabilized around 3,000 iterations, and the model remained effective when query numbers grew beyond 600. It also generalized to surveillance images and cucumber, potato, sweet potato, and tea leaves.
The authors said the real advance is not simply a higher score, but a detection framework designed around the messy reality of fields. They said dense leaves create a double burden: features become confused, and matching becomes expensive. By progressively fusing features and refining crowded queries, they said, Leaf-DETR learns faster and keeps more occluded leaves in view. They added that the dataset and code are intended to support broader agricultural monitoring, while fine-tuning for growth stage, viewing angle, and crop type will remain important for deployment.
Leaf-DETR could support multi-stage agricultural monitoring, from leaf counting and morphology analysis to disease classification and yield forecasting. Its strong generalization to surveillance images and staple crops suggests value as a foundational detector that can be fine-tuned for specific fields, growth stages, or cameras. The authors caution that the current model performs binary leaf-versus-background detection rather than species or disease classification, and accuracy drops by 16.1% on oblique monitoring views. In extremely dense tea canopies, predicted boxes may not fully enclose every leaf. Future work will combine viewpoint augmentation, multi-view fusion, and downstream task integration to make field deployment more robust.
###
References
DOI
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100182
Funding information
This research was supported by the National Natural Science Foundation of China (No. 62506089), Guizhou Provincial Science and Technology Projects ([2024]002, CXTD[2023]027), Scientific and Technological Innovation Platform Research Project of Guizhou Province (CXPTXM[2025]024, CXPTXM[2025]026), Guizhou Province Youth Science and Technology Talent Project ([2024]317), and the Postgraduate Research Fund of Guizhou Province (2024YJSKYJJ090).
About Plant Phenomics
Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cutting edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.