Microscopic hyperspectral image segmentation of wheat stripe rust based on a dual-model semi-supervised framework

Microscopic hyperspectral imaging can simultaneously capture spatial structural information and continuous spectral signatures of biological tissues. Microscopic morphology with spectral reflectance is often required for disease detection at the cellular level. However, conventional imaging cannot identify subtle pathological changes. Hyperspectral images are typically characterized by extremely high dimensionality, strong inter-band redundancy, and complex datasets. In addition, pixel-level annotation of microscopic hyperspectral imaging has greatly restricted the direct application of fully supervised segmentation, due to its highly labor-intensive and time-consuming. Especially, expert labeling is costly and limited in plant diseases. In this study, a dual-model framework of semi-supervised segmentation was proposed with dynamic high-confidence pseudo-labeling for microscopic hyperspectral images of wheat leaves infected by stripe rust. The framework was designed to reduce the need for large-scale manually annotated datasets, particularly for the segmentation accuracy and robustness. Specifically, convolutional neural network (CNN) and a Vision Transformer (ViT) were employed to serve as two complementary teacher-student branches. The CNN branch was utilized to capture local textures and fine-grained spatial patterns. While tThe ViT branch was used to simulate global contextual dependencies and long-range semantic relationships. Moreover, the two models iteratively exchanged reliable predictions using bidirectional cross pseudo-label supervision. The framework was realized designed to learn more discriminative representations from limited labeled samples and abundant unlabeled data. Furthermore, a dynamic threshold update mechanism was introduced to adaptively adjust the pseudo-label selection threshold using validation-set performance during training. The pseudo-label filtering was allowed to change with model performance, unlike fixed-threshold strategies. There was a balance between pseudo-label quantity and quality during learning. At the same time, a high-confidence filtering was incorporated to retain only those pixels with sufficiently reliable teacher predictions for student model optimization. Noisy pseudo-labels were suppressed to reduce error propagation during iterative training, thereby enhancing both segmentation stability and generalization. Experimental results demonstrate that the segmentation performance was significantly improved using microscopic hyperspectral images of wheat stripe rust. The U-Net model achieved an mDice of 0.829 0 and an mIoU of 0.707 9, while the ViT model achieved an mDice of 0.820 9 and an mIoU of 0.696 2 after semi-supervised optimization. The framework enhanced segmentation precision under limited annotation. Overall, a feasible and effective solution was provided to reduce manual labeling dependence in segmentation of microscopic hyperspectral images. Complementary model architectures were integrated with dynamic pseudo-label optimization. The framework alleviated pseudo-label noise, for the segmentation accuracy and stability. The findings can provide technical support for wheat tissue structure analysis and host-pathogen interaction phenotyping. A useful reference can offer be offered for semi-supervised segmentation of the microscopic hyperspectral imaging in agricultural and biological applications.