Maize tassel detection with CA-YOLO for UAV images in complex field environments

Abstract

Maize detasseling is a key step in maize cross-pollination and is a key part of maize seed production, and the accurate identification and detection of maize tassels is a prerequisite for maize detasseling. The traditional manual method has the disadvantages of strong subjectivity, high workload, and poor timeliness. An artificial intelligence algorithm based on deep learning provides a new method for maize tassel detection. In this research, an uncrewed aerial vehicle (UAV) collected data for 10 consecutive days, covering the entire maize tasseling period. Data collected under different weather conditions and UAV flight altitudes were used to construct a diverse dataset for maize tassel detection. A preprocessing strategy suitable for the tassel dataset was developed through a dataset analysis. Then, a coordinate-attention(CA) mechanism was added to the backbone network of You Only Look Once version 5 to enhance the regression and localization capability of the model by embedding location information to extract important features. The improved model is called CA-YOLO. Comparative experiments revealed that CA-YOLO provides an effective method for detecting maize tassels, with an average precision of 96 %, surpassing the YOLO series models, such as YOLOv5m and YOLOv7, and classical detection models, such as the faster region-based convolutional neural network and single-shot detector. Moreover, CA-YOLO is robust and can effectively detect early-stage, leaf-obscured, mutually obscured, and complex-background tassels. This study provides a theoretical basis for accurately monitoring tassels and mechanical detasseling systems.