Answers are grounded in this paper and cite their sources.
Accurate maize phenological information is crucial for improving yield prediction accuracy and optimizing field management decisions. However, traditional maize phenology detection methods largely rely on time-series data, which can only detect phenological stages after the growing season ends, making real-time monitoring challenging. This study provides a novel approach for real-time maize phenology monitoring using easily accessible and cost-effective unmanned aerial vehicle (UAV) RGB data. We identified maize plants at the seven-leaf, flowering, and maturity stages and used these data for model training to predict key phenological stages using the YOLOv6 model. We replaced the standard convolutional layers in the neck network of YOLOv6 with spatial and channel reconstruction convolutions (SCConv) to improve model efficiency and reduce the parameter count and floating-point operations. An efficient multi-scale attention mechanism (EMA) was added after fusing the high- and low-level features, enhancing the model’s ability to capture features at different scales and significantly improving detection accuracy. The improved model successfully identified key phenological stages in maize plants, achieving mean average precision (mAP) of 67.9%, mAP50 of 97.9%, mAP75 of 80.5%, and a recall rate of 98.57%. Incorporating the SCConv reduced the parameter count by 2.91 M and the floating-point operations by 2.68 G compared to the baseline model. The EMABlock improved detection performance for medium- and large-scale targets by 2.6% and 1.9%, respectively. The results demonstrate that the improved model provides superior recognition accuracy and a lightweight design, enabling potential deployment on mobile devices. Our results provide a theoretical foundation and technical support for the efficient real-time identification of maize phenological stages using UAV single-temporal imagery.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: