Enhancing model performance in detecting lodging areas in wheat fields using UAV RGB Imagery: Considering spatial and temporal variations

Abstract

Fast and accurate assessment of wheat lodging holds significant importance for disaster prevention and agricultural insurance claim settlement. However, lodging detection is often influenced by the spatial and temporal heterogeneity of fields, leading to inconsistent lodging features in images captured at different times and locations. This inconsistency poses challenges for evaluating model performance and obtaining consistent results. In this study, we proposed a comprehensive lodging area detection model performance evaluation method for multi-year and multi-phenological periods. Three evaluation indicators—accuracy, potential, and stability—were introduced to assess model performance from different perspectives. Additionally, a weighted method is employed to combine the results of multiple experiments. To evaluate the performance of lodging area segmentation in wheat fields, we selected three feature adaptive models (CBAM-unet, SE-unet, and Swin-transformer) along with the Unet model. The experimental results are as follows: (1) Swin-transformer achieved the highest weighted average accuracy ( Acc cp ) of 83.53% among the four models evaluated. (2) Swin-transformer exhibited the highest upper limit of segmentation accuracy, reaching 97.98%. However, its average accuracy in the prediction experiment was 85% of the upper limit, suggesting potential for optimization. (3) CBAM-unet demonstrated the lowest overall weighted segmentation accuracy variance (35.92) compared to the other three models, indicating higher stability. Based on the experimental results it can be seen that our proposed method and model can address the challenge of evaluating lodging area detection models under spatial and temporal heterogeneity.