Reliability-Aware Fusion of Lesion Segmentation and Agronomic Context for Rice Disease Diagnosis under Domain Shift
Abstract
Image-based rice disease models often lose reliability under domain shifts caused by varying backgrounds, symptom scales, and acquisition conditions. This study presents a linked two-stage design to investigate whether observational agronomic context can mitigate such transfer degradation. In Stage A, four instance-segmentation architectures are benchmarked on 6,465 source images across five disease classes; YOLO26m-seg is selected with 35.35% validation mask average precision and reaches 37.59% on the held-out test set. In Stage B, the frozen visual branch is transferred to 4,270 Paddy Doctor records spanning four overlapping diseases: instance confidences are aggregated into image-level visual evidence and combined with CatBoost context predictions (from crop age and variety) through class-specific reliability weights. Record-stratified five-outer/four-inner nested cross-validation yields 21.19% macro-F1 for vision-only, 47.92% for context-only, and 52.04% for the proposed fusion (55.46% accuracy), a 4.12-point gain over context-only diagnosis (95% bootstrap interval: 3.03–5.19). These findings suggest that class-dependent multimodal fusion can exploit complementary evidence within the evaluated Paddy Doctor records; grouped and open-set sensitivity analyses reveal limited unseen-context and out-of-distribution generalization.