Work

CAM-Threshold-Guided Weakly Supervised Lesion Segmentation and Relative Lesion-Burden Stratification of Wheat Stripe Rust Using CF2- WRD and Light-UNet

Year: 2026

Type: preprint

Abstract: Accurate localization and quantification of wheat stripe rust lesions are important for phenotyping and disease management, yet dense pixel-level annotation is costly. We present a Grad-CAM-guided weakly supervised framework that reuses a dual-branch image classifier, CF2-WRD, to generate pseudo-labels for a compact Light-UNet segmenter. The dataset contained 34,694 images in 87 image-level categories, split into 27,834 training, 3,430 validation, and 3,430 test images. CF2-WRD achieved validation and test accuracies of 0.8776 and 0.8711. Among four CAM thresholds, 0.40 produced the highest agreement with CAM-derived validation pseudo-labels (mIoU 0.6584, Dice 0.7101, lesion IoU 0.5555), whereas 0.45 was used as a conservative operating point. Raising the segmentation probability threshold from 0.50 to 0.80 reduced mean predicted lesion-to-leaf ratio from 55.02% to 29.63% in a 300-image subset. The 3,430-image burden distribution was stratified by empirical quartiles. Because no independent manual lesion masks or expert-calibrated severity labels were available, overlap metrics quantify agreement with weak-supervision targets and burden groups represent dataset-relative rankings rather than agronomic severity grades.

Source: Research Square

DOI: 10.21203/rs.3.rs-10959751/v1

Authors: Bojian Guo, Xin LI, Temirbaeva Nazgul Ismanovna, Asel Kartanova

Institution: Kyrgyz State Technical University

Language: English

Topic: Smart Agriculture and AI

Subfield: Plant Science

Field: Agricultural and Biological Sciences

Domain: Life Sciences

SDG: Zero hunger

Open Access status: green

Funder: Gansu Agricultural University