Lodging is one of the primary factors that reduce wheat yield; therefore, rapid and accurate monitoring of wheat lodging helps to provide data support for crop loss and damage response and the subsequent settlement of agricultural insurance claims. In this study, we aimed to address two problems: (1) calculating the wheat lodging area. Through comparative experiments, the SegFormer-B1 model can achieve a better segmentation effect of wheat lodging plots with a higher prediction rate and a stronger generalization ability. This model has an accuracy of 96.56%, which realizes the accurate extraction of wheat lodging plots and the relatively precise calculation of the wheat lodging area. (2) Analyzing wheat lodging areas from various growth stages. The model established, based on the mixed-stage dataset, generally outperforms those set up based on the single-stage datasets in terms of the segmentation effect. The SegFormer-B1 model established based on the mixed-stage dataset, with its mIoU reaching 89.64%, was applicable to wheat lodging monitoring throughout the whole growth cycle of wheat.
Keywords: area calculation; deep learning; growth stages; unmanned aerial vehicle (UAV); wheat lodging.