Abstract
Cucumber foliar symptoms reduce functional leaf area and complicate greenhouse scouting. This study presents CuFormer U-Net, a Transformer-enhanced multi-task network for three-class image-level classification on the 3,754-image Kaggle benchmark (healthy, powdery mildew, and downy mildew) and binary visible-symptom segmentation on a separate 400-image greenhouse cohort (200 healthy and 200 affected leaves). The greenhouse annotations comprise 2,705 rectangles (1,741 color-abnormality and 964 tissue-damage instances); rectangle-seeded lesion masks were manually refined for the dense-prediction experiments. The reported fixed-split point estimates were 96.7% accuracy, 96.3% macro precision, 95.8% macro recall, and 96.0% macro F1 for classification, and 92.4% mean image-wise Dice, 83.1% macro IoU, 93.2% foreground precision, and 91.7% foreground recall for segmentation. The complete internal-validation ablation configuration reached 93.1% Dice and 87.3% IoU. The available experiment archive does not contain independent-seed reruns, per-image predictions for inferential testing, annotator-level masks, quantitative results for the separate 100-image greenhouse stress-test cohort, or reproducible hardware profiling. Accordingly, all between-model differences are interpreted descriptively, the external cohort is not presented as formal external validation, and robustness and deployment claims have been moderated. CuFormer U-Net provides a technically motivated basis for joint recognition and spatial characterization, but confirmatory multi-run and multi-site evaluation is required. boundary refinement, cucumber leaf disease, multi-task learning, plant disease, precision agriculture, semantic segmentation, vision transformer