DDAN-RGAL: A Dual-Domain Adaptive Network With Reality-Gap Aware Learning for Field-Deployable Plant Disease Detection

Abstract

Deep learning models achieve over 95% accuracy for plant disease detection in laboratory settings but suffer severe performance drops exceeding 30% when deployed under real agricultural conditions. This article presents dual-domain adaptive network with reality-gap aware learning (DDAN-RGAL), a unified framework that explicitly bridges this laboratory-to-field gap through three synergistic modules. The first, domain-invariant feature disentanglement, separates disease-intrinsic patterns from environmental artifacts using adversarial and mutual-information constraints. The second, cross-domain prototype learning with uncertainty, learns hierarchical laboratory and field disease prototypes while providing calibrated confidence through evidential deep learning. The third, reality-gap aware meta-classification, dynamically adapts expert classifiers based on domain characteristics inferred from each input. Experiments on PlantVillage, PlantDoc, and a real-world paddy field dataset collected using a field-deployed speed-dome camera and Raspberry Pi-based edge system demonstrate that DDAN-RGAL achieves 98.7% laboratory accuracy and 94.8% field accuracy, reducing the reality gap from 27.9% to 3.9%, a more than seven-fold improvement over existing methods. The framework provides highly reliable uncertainty estimates (expected calibration error: 0.042), inference (21.6 ms on GPU; 285 ms on Raspberry Pi 4), and strong data efficiency (85.3% accuracy with only 10% of the training data). These results establish DDAN-RGAL as a practical and field-ready solution for precision agriculture, enabling robust deployment of laboratory-trained models in real-world farming environments where both accuracy and reliability are essential.