DFA-YOLO: an enhanced YOLOv11-OBB and knowledge distillation-based maize stomata detection system

Stomata are vital gatekeepers of plants that regulate the fundamental trade-off between carbon gain and water loss. Precise, high-throughput identification of stomatal traits is therefore essential for assessing plant stress tolerance and water-use efficiency. However, conventional bounding box detection struggles to accurately localize densely distributed and arbitrarily oriented stomata.

This study proposes DFA-YOLO, an enhanced YOLOv11-OBB (Oriented Bounding Box) model for orientation-aware maize stomatal localization and preliminary OBB-derived trait extraction. Based on a maize stomatal dataset expanded from 1,053 to 3,597 microscopic images, DFA-YOLO integrates three task-specific components: (1) a cross-dataset MGD distillation strategy that transfers structural priors from single-stomata images to dense multi-stomata scenes; (2) a fixed-threshold Focaler-CIoU localization-loss reweighting strategy (u_stomata = 0.95, d_stomata = 0.00) to emphasize hard positive samples during OBB regression; and (3) a C3k2_AssemFormer feature-aggregation module that combines convolutional local feature extraction with linear attention-based context aggregation.

DFA-YOLO achieved 94.1% mAP50, 84.8% mAP75, 74.1% mAP50-95, and 90.0% recall, with higher recall, mAP50, mAP75, and mAP50-95 than the YOLOv11-OBB baseline under the same OBB evaluation protocol. When deployed on an automated platform, the system processed images at 44.9 FPS and supported stomatal localization, density estimation, and preliminary orientation-aware size description.

Under the tested maize microscopic imaging workflow, DFA-YOLO enables rapid extraction of detection-oriented stomatal traits and provides a prototype tool for high-throughput maize stomatal phenotyping.