Leakage-guarded embedding compression and calibration for edge rice diagnosis

Reliable diagnosis of rice diseases from field imagery requires more than strong performance on randomly partitioned data because near-duplicate leakage, calibration drift under image corruption, and edge-device constraints can each obscure practical utility. This study presents a leakage-guarded pipeline for compact, edge-oriented rice disease diagnosis. Perceptually similar images are grouped by perceptual hashing before group-disjoint train, validation, calibration, and test partitions are formed. A frozen EfficientNet-B0 backbone then produces 1280-dimensional embeddings, and a lightweight multilayer perceptron (MLP) head is trained with shift supervision, consistency regularization, and embedding-space MixUp. Multi-objective optimization selects a sparse embedding mask and head hyperparameters by balancing macro-F1, expected calibration error (ECE), field-shift performance, and a deployment-cost proxy. Temperature scaling fitted on a held-out calibration split supports selective prediction and an illustrative cost-aware recapture policy. On an expanded Paddy Doctor benchmark containing 29,538 images from 10 classes, the selected model retains 242 of 1280 embedding dimensions (18.9%) and achieves a macro-F1 of 0.948 (95% confidence interval: 0.942–0.954) on the leakage-guarded in-distribution test set, with an ECE of 1.12%. Under the most severe synthetic field shift considered, macro-F1 remains 0.857. These results indicate that sparse embedding selection can preserve most of the predictive performance of a tuned full-embedding head while reducing head-level complexity. Severe blur remains a major failure mode, and broader multi-site validation is required before agronomic deployment.