DTM-Unet: A deep learning framework incorporating DenseNet and transformer for precise identification and quantitative segmentation of cucumber leaf disease

• ‘DTM-Unet achieves spot segmentation of cucumber leaf disease with 98.71 % PA. • The MRDC module enhances the Unet network feature extraction capability. • A composite loss function solved the class imbalance problem in segmentation. • The feature reuse improves extraction efficiency of deep features of fuzzy lesions. Precise identification and accurate segmentation of disease lesions on cucumber leaves are crucial tasks in smart agriculture for early detection and effective management. Taking common foliar diseases of cucumber as an example, cucumber downy mildew, bacterial angular leaf spot and cucumber phytophthora blight are characterized by indistinct lesion boundaries and fuzzy appearance during early infection stages, poses significant challenges to traditional segmentation methods. To overcome these issues, we propose a novel semantic segmentation framework, named DTM-Unet, which effectively integrates DenseNet and Vision Transformer architectures into an enhanced U-Net model. The DenseNet backbone facilitates efficient feature reuse and propagation, capturing rich hierarchical features, while the Transformer encoder captured long-range contextual dependencies, enhancing global feature representation. Additionally, we designed multi-residual dilated convolution (MRDC) modules within skip connections to effectively expand the receptive field and capture fine-grained lesion details. The model was trained and validated on a dataset encompassing multiple cucumber leaf diseases, augmented with various image transformations to improve generalizability. Performance evaluation indicated that the DTM-Unet model achieves superior segmentation performance, attaining a Pixel Accuracy (PA) of 98.71 %, Recall of 91.27 %, F1-score of 86.96 %, and Mean Intersection over Union (mIoU) of 87.26 %. Comparative evaluations against existing segmentation models further verify its robustness and generalization ability, highlighting the potential of DTM-Unet for broader applications in intelligent crop disease diagnosis and precision agriculture management, offering valuable support for automated plant disease phenotyping and precise agricultural interventions.