Ullah, Naeem (2026) Explainable Artificial Intelligence Across Agrifood and Environmental Domains: From Plant Disease Imaging to Time Series Forecasting. [Tesi di dottorato]

Documento PDF Ullah_Naeem_unina_thesis_final.pdf Download (18MB)

Collection description

The increasing use of artificial intelligence approaches in agriculture and environmental domains raises an urgent need for models that are not just accurate but explainable as well to ensure trustworthiness and transparency. This thesis aims to enhance trustworthiness and transparency by developing novel deep learning frameworks that consider explainability as an integral part of the methodology. The thesis considered two high-impact domains: agrifood systems and environmental monitoring, where transparent and reliable AI models are crucial for sustainable and ethical decision-making. Specifically, the contributions of the thesis to the agrifood domain include adaptive PSO-LemonNetX for smart classification of lemon diseases and Deep-CCNet optimized with meta-heuristic feature selection algorithms for cotton leaf disease detection. These approaches combine deep learning with feature selection and explainability approaches such as LIME , which ensure more interpretability and efficiency. Experimental results show state-of-the-art classification performance while ensuring lightweight architectures suitable for IoT and edge deployments. Furthermore, in the environmental domain, this thesis proposes an attention-based LSTM framework for multi-horizon NO2 and O3 air pollutant forecasting. The proposed approach combines attention mechanisms, hyperparameter optimization using Optuna, and explainability through integrated gradients and attention visualization. The approach is validated on the data from five urban monitoring stations, and the results confirm high predictive accuracy with meaningful, interpretable details at both the temporal and feature levels. The thesis outlines common challenges across the studies, including model complexity versus explainability trade-offs, domain-specific generalization, and the need for human-centered evaluation metrics. Building on this, a cross-cutting synthesis is made to capture conceptual and methodological lessons from both domains. The emphasis is that explainability not only enhances transparency but also model robustness, user trust, and real-world applicability. In conclusion, this thesis demonstrates how deep learning models can be made both interpretable and efficient for complex applications in agrifood and environmental systems. The frameworks developed here offer a pathway for building more transparent, sustainable, and human-centered AI solutions that are viable for real-world deployment.

Downloads

Actions (login required)

View Item