Leakage-Safe Cross-Domain Adaptation for Bangladeshi Rice Leaf Disease Classification
Students & Supervisors
Student Authors
Supervisors
Abstract
This paper evaluates and improves cross-domain rice leaf disease classification using a leakage-safe experimental protocol. Two Bangladeshi real-field datasets, RiceLeafBD and RiceyLeafDisease, are used and reduced to a shared three-class problem: bacterial leaf blight, brown spot, and healthy. The protocol separates source training data, target adaptation data, and the final untouched target test set to avoid data leakage. ResNet50 is used as the main backbone model, and three settings are compared: source-only transfer, Deep CORAL, and 10% labeled target fine-tuning (FT10). The results show that adaptation can improve cross-domain performance, but the improvement depends on the transfer direction. In the RiceLeafBD to RiceyLeafDisease direction, FT10 gives the clearest improvement, while Deep CORAL does not provide a reliable gain. In the RiceyLeafDisease to RiceLeafBD direction, both Deep CORAL and FT10 improve performance, with FT10 achieving the strongest result. McNemar analysis across three seeds supports these findings. Overall, the study shows that cross- domain rice disease models should be evaluated with careful data separation, and that using a small labeled target sample can be a practical way to improve model transfer to new field conditions.
Keywords
Publication Details
- Type of Publication:
- Conference Name: 2026 IEEE International Conference on Signal Processing, Information, Communication and Systems (SPICSCON)
- Date of Conference: 13/08/2026 - 13/08/2026
- Venue: Bangladesh Army University of Engineering & Technology (BAUET), Qadirabad, Natore-6431, Bangladesh