Quality-consistent purification and structure-conditioned diffusion augmentation with ConvSwin feature fusion for multiclass wheat leaf disease classification

Introduction Food security and sustainable crop production are essential for meeting the demands of a growing global population. Among staple crops, wheat is the second most consumed grain worldwide, and its sustainable production is essential for global food demand. The production of wheat is significantly influenced by foliar diseases, which may lead to lower yields of the crop, hence impacting sustainable production. Therefore, this work focuses on the classification of nine major wheat leaf diseases and one healthy leaf class. The multiclass nature of the problem is challenging, as several disease classes exhibit high visual similarity with only minor distinguishing features. Methods A diffusion-augmented ConvSwin Feature Fusion framework is developed for ten-class wheat leaf disease classification. The architecture fine-tunes a pretrained ConvNeXt-Tiny backbone and refines its spatial feature representation using a lightweight two-block Swin-style self-attention module before MLP-based feature fusion. The dataset comprised 10,641 images, of which 1,000 original images were reserved for validation and another 1,000 for testing. Of the remaining 8,641 training images, 5,839 were retained after the complete VQC-PS purification process. The purified training subset had a maximum class-count ratio of approximately 3.59:1 and was subsequently standardized to 1,000 images per class using structure-conditioned diffusion synthesis and controlled downsampling. The resulting VQC-PS ConvSwin-FF framework combines quality-aware training-data preparation, diffusion-based class balancing, convolutional feature extraction, transformer-based spatial refinement, and MLP-based feature fusion within a unified classification pipeline. The results are further interpreted using XAI-based techniques to enhance model transparency. Results Across five independent runs, the proposed framework achieved a mean test accuracy of 93.40 ± 1.25% and a macro-F1 score of 93.24 ± 1.39%, with the best run attaining 94.50% accuracy. Discussion These results indicate consistent performance across different random initializations and competitive performance relative to the evaluated baseline architectures under the adopted experimental protocol.