Authors

Abstract

Crop semantic segmentation plays a crucial role in precision agriculture, enabling applications such as growth monitoring, yield prediction, and pest control. However, deep learning methods, such as U-Net, rely heavily on large amounts of labelled data, which are costly and time-consuming to obtain in agricultural settings. To address this limitation, a semi-supervised maize segmentation method based on an improved Vision Transformer within a student-teacher framework is proposed. The model leverages limited labelled data and abundant unlabelled data through consistency training and confidence-based self-training. Experimental results demonstrate that the proposed method achieves a mean Intersection over Union (mIoU) of 0.661, representing a 14.3% improvement over U-Net. These results confirm its effectiveness in reducing annotation costs while achieving superior accuracy in complex farmland environments.

Abstract in Chinese

作物语义分割对于精准农业至关重要，有助于生长监测、产量预测和病虫害控制。然而，像U-Net这样的深度学习方法严重依赖大量标记数据，在农业环境中获取这些数据既昂贵又耗时。为此，我们提出了一种基于改进版Vision Transformer和师生框架的半监督玉米分割方法。该模型通过一致性训练和基于置信度的自我训练，利用有限的标记数据和大量的未标记数据。实验结果显示其mIoU为0.661，比U-Net高14.3%，表明其在减少注释负担的同时，同时在复杂农田场景中实现了更高的精度。