Multi-disease identification and severity estimation in crop leaves using panoptic segmentation

Abstract

Pixel-level lesion segmentation enables direct localization and quantification of symptomatic sugarcane tissue, but semantic overlap alone does not show whether individual lesions are separated reliably or whether predicted lesion area is quantitatively useful. This study evaluates Panoptic Feature Pyramid Network (Panoptic-FPN) as an instance-aware lesion-analysis framework and benchmarks it against convolutional and transformer-based segmentation models. The experimental corpus contains 588 image-mask pairs with binary lesion annotations. Source-image identities were partitioned before augmentation into 411 training, 88 validation and 89 test images (69.90%, 14.97% and 15.14%); augmentation was restricted to the training subset, producing 2,055 additional records, and no source identity overlapped across subsets. U-Net, DeepLabV3+, Mask R-CNN, Panoptic-FPN, SegFormer-B2 and Mask2Former-Swin-Small were evaluated on the same untouched 89-image test set. Mask2Former-Swin-Small achieved the highest Dice (84.01%) and mean intersection-over-union (mIoU; 74.52%), whereas SegFormer-B2 achieved the highest precision (86.70%) and accuracy (95.31%). Panoptic-FPN achieved Dice 79.28%, mIoU 69.02%, precision 85.37%, accuracy 94.79% and the highest specificity (97.52%). Friedman tests followed by paired Wilcoxon signed-rank tests with Holm correction confirmed significant model-level differences. Panoptic-FPN retained an overall robustness score of 96.29% across six controlled perturbations, estimated lesion coverage with mean absolute error of 3.07 percentage points and R2 of 0.866, and achieved micro instance F1 of 62.71% at IoU ⩾ 0.50 with mean matched-instance Dice of 84.38%. These findings show that transformer segmenters provide the strongest semantic overlap, while Panoptic-FPN adds complementary instance-aware and severity-oriented information for quantitative lesion analysis.