CPD-YOLO: A cross-platform detection method for cotton pests and diseases using UAV and smartphone imaging

Abstract

Cotton is a major industrial crop and a primary raw material for textiles worldwide. Intelligently detecting pests and diseases is crucial for automating cotton cultivation management. This study proposed a cross-platform detection method for cotton pests and diseases using unmanned aerial vehicle (UAV) and smartphone imaging. The method employed a self-built improved YOLO, namely CPD-YOLO. The bidirectional feature pyramid network with the reparameterization vision transformer was constructed to achieve an optimal balance between multi-scale feature fusion and inference efficiency. Secondly, the head network, featuring four dynamic detection heads, was designed to enhance the multi-dimensional dynamic awareness capability. Finally, the Inner-Shape intersection over union was proposed as the bounding box regression loss function, which can improve positioning accuracy and accelerate convergence. The improvement strategy significantly improved the detection accuracy of multi-scale objects in cross-platform scenarios. Compared with the original model, the F1-score and mean average precision (mAP) were increased by 7.44 % and 7.08 %, respectively. The CPD-YOLO also outperformed typical object detection models, with an F1-score of 88.86 % and a mAP of 90.42 %. Moreover, the CPD-YOLO model had superior generalization capability, with a 1.30 % decrease in F1-score and a 0.55 % decrease in mAP. Furthermore, a mobile application named Doctor Cotton, utilizing the CPD-YOLO, was developed. It promises to be a new farming tool for the real-time and accurate detection of cotton pests and diseases. These findings provide a valuable reference for the cross-platform application of consumer-grade UAVs and smartphones in detecting other crop pests and diseases.