UANP-MT based semi-supervised image segmentation method for identifying weeds in cabbage field

Weeds pose a great threat to modern agricultural production. The unrestrained growth of weeds has also a significant impact on crop yields and quality. Therefore, it is a high demand to effectively manage and control weeds in recent years, in order to optimize agricultural production activities. The specific area of weed management can be expected to serve as the promising research direction for weed control operations, with the rapid development of unmanned aerial vehicle (UAV) platforms and artificial intelligence (AI). Accurate and efficient identification of weeds can be one of the most key steps to realizing automated weed management in the field. However, the current efficient models of weed identification are often required a significant amount of labeled agricultural data. The major bottleneck can be also limited in the model. In this study, a semi-supervised semantic segmentation network was proposed, called UANP-MT (uncertainty aware and network perturbed mean teacher), in order to reduce the dependence on labeled data. The PSPNet network model was also selected to leverage the teacher-student network concept from MT (Mean Teacher). Several key improvements were then incorporated during this time. Firstly, the improved model reduced the errors that caused by random network factors. The teacher network output was also augmented to take the mean, thereby ensuring the robustness of the network prediction. Secondly, the improved model constructed the uncertainty constraints for the different uncertainties between networks and assigns weightings, in order to balance the differences in the output between different networks. This approach enhanced the confidence and reliability of the prediction, ultimately leading to the higher recognition accuracy of an improved model. A series of ablation experiments were conducted to evaluate the effectiveness of the improved model. These experiments included the setting values of network parameters, the selection of feature extraction network backbones, and the model performance testing on datasets with varying volumes of data. Some optimal parameter settings were identified for the improved model, with the ResNet50 as the preferred choice for the model backbone. The comparison experiments showed that the UANP-MT model outperformed the original supervised network, in terms of the three evaluation indicators of F1 score, pixel accuracy (PA), and intersection over union (IoU), when the labeled data was less than that of the original supervised network, indicating the superiority of the semi-supervised network. Furthermore, the effectiveness of the improved model was further validated to compare with the classic semantic segmentation networks, including SegNet, U-Net, and Deeplabv3+ on the CaiXin weed dataset. UANP-MT model achieved better performance with an F1 score of 81.83%, a PA of 95.84%, and an IoU of 90.70% when only 1/4 of the labeled data was used for training. These evaluation metrics were superior to those of the Deeplabv3+ model by percentage point of 4.71, 7.94, and 8.27, respectively. The UANP-MT model was achieved in the high-quality weed detection and recognition with the low labeled data, thus significantly reducing the dependence on labeled data with the manpower and time cost saving. The findings can provide valuable implications for the subsequent weed recognition research and the development of automated weeding operations on UAV platforms.