CROSS-DOMAIN ROBUSTNESS OF COTTON LEAF CURL DISEASE CLASSIFICATION

Abstract

High internal performance does not guarantee that a plant-disease classifier will behave similarly under new backgrounds, lighting, cameras or visually overlapping conditions. This paper evaluates the frozen cross-domain robustness of eight Cotton Leaf Curl Disease (CLCuD) classifier configurations developed under a common ResNet34 training protocol. External evaluation used 257 Healthy Leaf and 400 Curl Virus original field images from SAR-CLD-2024 after exact and perceptual overlap auditing against the 3,462-image internal cohort. Twenty-three available checkpoints were evaluated without retraining, external threshold fitting or calibration under raw and automatically leaf-isolated conditions. Adaptive disease-preserving enhancement (A1) achieved the highest seed-ensemble balanced accuracy in both settings: 0.802 raw (95% source-group bootstrap CI 0.770 0.834) and 0.817 isolated (0.787–0.847). A prespecified stress extension added all 1,449 images from five other repository conditions, producing 2,106 images and 96,876 checkpoint predictions. Best full-cohort balanced accuracy was 0.597 raw and 0.620 isolated. Herbicide Growth Damage was the strongest category-specific confounder. Quantitative attribution and cue-intervention audits showed that stronger-looking foreground localization did not necessarily imply stronger cross-domain classification. The study provides a frozen, overlap audited account of acquisition shift and class-boundary stress, identifies bounded adaptive preprocessing as the most consistent component in the two-class field comparison, and establishes explicit multiclass differential diagnosis supervision as the next system requirement.