Abstract
Deep learning models for plant disease detection achieve near-perfect accuracy on controlled datasets such as PlantVillage, yet their performance degrades substantially when deployed on real-world field images due to domain shift. This gap between laboratory and field conditions limits practical applicability in agricultural settings. In this paper, we investigate whether multi-task learning (MTL), combined with CORALbased domain alignment, can improve cross-domain generalization for plant disease recognition. Our framework jointly predicts plant species and disease class, encouraging the learning of structured and domain-invariant representations. Domain shift is explicitly addressed through covariance alignment between source and target feature distributions using unlabeled target-domain data. All models are trained on PlantVillage and evaluated on PlantDoc in a strict cross-domain setting. Experimental results show that MTL consistently outperforms single-task baselines, while CORAL-based alignment further improves target-domain accuracy and macro-F1 score, with the strongest gains achieved by a self-adaptive multi-scale MTL variant.