Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2025
- Goals
- Record
- 6c588663-6d2a-4417-bd72-6a283e3f1354
- Published
- 2026-10-30
Universitas Dian Nuswantoro (UDINUS), with the participation of Dewanto, Bayu; Luthfiarta, Ardytha; Nugraha, Adhitya; Firdausillah, Fahri; Fitri, Maulatus Shaffira; and Marendratama, Radhitya, conducted research on plant disease detection using leaf images. The activity addressed the limited availability of labeled data, which poses a significant challenge for training deep learning models in precision agriculture. The research proposed AJISAI, a dual encoder self-supervised learning framework designed to learn meaningful features from unlabeled plant leaf images.
AJISAI processed global and local crops of the same image through separate encoders, ResNet50 and ResNet18. Each encoder was followed by an independent projection head, and the embeddings were aligned using a single InfoNCE loss. After training, the encoder was frozen and evaluated in a 5-way 5-shot classification setup using logistic regression. The research benchmarked AJISAI against SimCLR and AMDIM using the MangoLeafBD dataset.
The reported results showed that AJISAI outperformed both baselines, with an accuracy of 82.67% and an F1 score of 0.7687. The framework also achieved the shortest training time, 24.5 minutes, despite its larger model size. These findings highlighted AJISAI's ability to learn semantically rich and fine-grained features and to achieve robust performance under limited supervision. The research stated that combining multiscale representation learning with lightweight few-shot classifiers offered a practical and scalable solution for plant disease classification in data-scarce agricultural settings.
Through this documented activity, UDINUS presented a research-based approach to plant disease classification using unlabeled leaf images, multiscale representation learning, and few-shot evaluation. The activity was related to the stated goal through its agricultural research focus and its examination of a method for classification when labeled data are limited. The institutional evidence consists of the proposed framework, the described processing method, the benchmark comparison, and the reported evaluation findings.