AF-RT-DETR: Adaptive cross-scale feature interaction for real-time plant disease detection in complex field environments

Abstract

Introduction Accurate plant disease identification is of great importance for ensuring agricultural productivity and food security. However, complex illumination variations, leaf occlusion, and diverse disease spot scales throughout plant growth stages significantly increase the difficulty of real-time detection, leading to limited accuracy and robustness in existing approaches. Methods To address these challenges, we propose an improved RT-DETRv2-based plant disease detection model, termed AF-RT-DETR. A Bidirectional Cross Gate (BCG) module is introduced in the feature extraction stage to reduce channel redundancy and enhance discriminative feature representation through multi-level feature interactions. The original RepVGG structure is replaced with a Dynamic Channel Shift (DCS) module, effectively enlarging the receptive field and strengthening contextual feature fusion without additional computational overhead. Additionally, an improved Scale-aware Multi-level Loss (SML) emphasizes low-quality feature maps to improve detector robustness. Results The model achieves mAP50 and mAP50:95 of 93.6% and 67.2% on the Plant-Disease dataset, surpassing the baseline by 5.1% and 4.5%. Furthermore, the model was evaluated on multiple crops and growth stages under diverse field conditions, demonstrating robust performance and adaptability. Discussion These results indicate that AF-RT-DETR effectively enables real-time plant disease detection in complex field environments.

Plant phenotyping relevance

植物の病徴を画像から検出するモデルの開発と、複数作物・生育段階・圃場条件での性能評価が中心であり、植物病害状態の表現型計測手法に該当する。

Code and data availability

The paper evaluates AF-RT-DETR on three public Roboflow plant-disease image datasets, each cited with an explicit public URL. No author analysis code or trained model release is mentioned. The Ultralytics YOLOv8 repository is a generic third-party library, not a paper-specific asset.