Towards an Integrated AI Pipeline for Disease Diagnosis and Quality Assessment in Coffee Production

Coffee production faces increasing vulnerability to fungal diseases, inconsistent quality control, and growing sustainability pressures across the entire value chain. Although Artificial Intelligence and deep learning have advanced crop monitoring and defect detection, research on coffee remains fragmented, with most studies addressing isolated tasks such as leaf disease detection, berry severity grading, or bean quality assessment. This paper introduces a proof-of-concept framework for an Integrated Coffee AI Pipeline designed to unify these stages within a single, explainable, and scalable architecture. The framework integrates three analytical modules: (i) a YOLO-based field detection system for real-time identification of diseased leaves and berries under natural conditions; (ii) a hybrid Vision Transformer-CNN module for contextual severity estimation and explainable progression mapping; and (iii) a two-stream CNN coupled with a Hidden Markov Chain for post-harvest bean grading and defect analysis. An overarching Explainability and Decision Layer aggregates attention maps, uncertainty estimates, and economic indicators to generate traceable, data-driven insights from field to factory. By linking pre-harvest diagnosis with post-harvest quality evaluation, Integrated Coffee AI Pipeline establishes a continuous feedback loop that mirrors the biological and economic lifecycle of coffee. Although conceptual, the proposed architecture provides a replicable blueprint for developing integrated, transparent, and sustainability-oriented AI systems aligned with precision agriculture and the United Nations Sustainable Development Goals.