Temporal generalization evaluation of deep semantic segmentation models for UAV-based wheat lodging monitoring under transfer learning

Temporal domain shift is a major barrier to the reliable application of agricultural artificial intelligence (AI) models in long-term field monitoring. Although deep learning models have achieved high accuracy in UAV-based crop lodging monitoring, their performance often declines across years, phenological stages, illumination conditions, and sensor platforms. This study developed a temporal generalization evaluation framework for deep semantic segmentation models for UAV-based wheat lodging monitoring under transfer learning. Five years of UAV imagery were organized with 2023 as the source domain and 2019, 2020, 2021, and 2024 as temporal target domains. Class-wise Bhattacharyya distance, RBF maximum mean discrepancy (RBF-MMD) of frozen ResNet50 features, and Wasserstein distance of lodging proportions described appearance, deep-feature, and label-distribution shifts, respectively. Nineteen representative models were first evaluated for source-domain convergence; nine converged models were retained, and eight further passed within-year generalization for direct cross-year deployment and target-domain fine-tuning. Target year strongly affected reliability: 2019 had the largest appearance shift and greater between-run variability, whereas 2024 had a smaller appearance discrepancy. Transformer models, especially Swin Transformer, generally showed stronger temporal transferability than convolutional networks. On the 2024 prediction subset corresponding to 20% fine-tuning, Swin-L achieved 87.05% lodging Dice, 89.92% lodging-free Dice, 84.37% lodging recall, and 92.20% lodging-free recall, with a Gap(a-m) of 0.38%. Across three independent runs, its 2024 micro lodging Dice was 0.882 ± 0.010, compared with 0.614 ± 0.099 in 2019. Limited newly annotated target-domain samples can support adaptation, but the outcome is constrained by target-year difficulty and sample composition.