Lightweight Self-Supervised and Uncertainty-Aware Framework for Efficient Plant Disease Classification

Lightweight Self-Supervised and Uncertainty-Aware Framework for Efficient Plant Disease Classification

Accurate and timely identification of plant diseases is vital for food security, yet deep learning models like LeafDisDiff suffer from critical drawbacks in real-world deployment: they require large amounts of labeled data, lack uncertainty awareness, and are computationally intensive. We propose a novel plant disease classification framework designed to overcome these limitations. The proposed system integrates self-supervised representation learning, continual fine-tuning, and uncertainty-aware prediction into a lightweight architecture optimized for edge deployment. The approach leverages self-supervised pretraining to learn robust visual representations from unlabeled leaf images, mitigating the need for extensive annotation. Through continual learning strategies, the model incrementally adapts to new diseases and seasonal shifts without catastrophic forgetting. By incorporating Bayesian approximations, it quantifies predictive uncertainty, enabling selective predictions or expert referral under ambiguity. Finally, a knowledge distillation process transfers generalization capabilities from computationally expensive diffusion models into a compact, real-time model suitable for deployment on resource-constrained agricultural devices.

Candidates are expected to carry out their research full-time within the structures of EUROMED University.

• Develop and implement self-supervised learning algorithms to learn robust and transferable plant disease representations from large collections of unlabeled leaf images.
• Design and evaluate continual learning strategies (domain-incremental and class-incremental) to ensure adaptive performance across evolving agricultural environments and new disease classes.
• Integrate Bayesian uncertainty estimation methods such as Monte Carlo Dropout, SWAG, and deep ensembles to quantify epistemic and aleatoric uncertainty for reliable disease diagnosis and expert referral.
• Implement knowledge distillation pipelines to transfer robustness and generative expressiveness from high-capacity diffusion models into lightweight models optimized for edge deployment.
• Conduct comprehensive experimental validation on benchmark and real-field datasets, analyzing trade-offs among accuracy, uncertainty calibration, and computational efficiency.
• Publish research results in top-tier AI and agricultural vision journals and contribute to the development of deployable, uncertainty-aware disease monitoring tools for precision agriculture.

• MSc (or BAC+5) in Computer Science, Artificial Intelligence, or a closely related field.
• Solid understanding of deep learning architectures (CNNs, Transformers), optimization techniques, and modern training paradigms.
• Proficiency in Python and experience with deep learning frameworks such as PyTorch or TensorFlow.
• Experience or strong interest in self-supervised learning, uncertainty modeling, or continual learning.
• Knowledge of probabilistic machine learning, Bayesian deep learning, or generative models (e.g., diffusion models, VAEs) is an asset.
• Publications in machine learning, computer vision, or AI for sustainability is a plus.
• Capacity for independent research and team collaboration in interdisciplinary and international environments.
• Excellent written and oral communication skills in English.

CV, a cover letter, the PhD project, diplomas, and academic transcripts.

The application file must be sent to the Doctoral Studies Center (CEDoc) of the Euro-Mediterranean University of Fes by email no later than October 24, 2025, to the following contacts:

Administrative Affairs Officer of the CEDoc: Mrs. Boutaina Jai Mansouri : : b.jai-mansouri@emadu.ueuromed.org)

Director of Research and of the CEDoc: Prof. Abdelghafour Marfak : : a.marfak@euromed.org)

• Pr. Abebaw Degu WORKNEH (a.degu-workneh@ueuromed.org)