Answers are grounded in this paper and cite their sources.
Plant disease recognition under a controlled-to-field domain shift suffers from source–target evidence misalignment. Uneven class coverage leaves some diseases weakly supported, while pseudo-label error accumulation can bias target memory. We propose MemAgri, a source-anchored dual-view target-memory framework that accumulates unlabeled field evidence. To address source–target evidence misalignment, target centroids capture field appearance while source anchors retain class references. For uneven class coverage, we combine complementary DINOv2 and Deep CORAL prototype predictions with empirical prediction-frequency correction to reduce evidence concentration. To limit pseudo-label error accumulation, a separate parametric posterior filters memory admission instead of allowing prototype diagnosis to determine its own updates. Each streaming prediction precedes the corresponding update to avoid immediate self-reinforcement. Batch adaptation refines target centroids before read-only testing. On 27 shared PlantVillage–PlantDoc classes with 236 test images, batch MemAgri raises Macro-F1 from 0.5872 for fixed source prototypes to 0.6689. External shared-class evaluations show further prototype gains under frozen development settings. Model selection uses target-development labels, and the main representation uses offline full-pool alignment. Persistent pseudo-label errors and the restricted evaluation scope limit broader field generalization.
No takes yet. Share an insight, caveat, or question.
Liao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: