AI-Based Crop Disease Detection from Leaf Images: A Focused Review and Robust Evaluation Framework

Plant disease detection from images has become a major application area for computer vision in agriculture. Deep learning studies have demonstrated very high classification performance on curated leaf-image datasets, but the meaning of these results changes when photographs are captured under natural field conditions. This paper focuses on one task only: classifying the disease or healthy status of a crop from a leaf image. It reviews representative image-based studies, identifies the controlled-to-field generalization problem as the central research gap, and proposes a focused evaluation framework based on image preprocessing, transfer learning, leakage-resistant dataset partitioning, and independent external testing. PlantVillage is treated as a controlled benchmark and PlantDoc as a field-like external dataset, with only semantically matched disease categories used for cross-dataset evaluation. The framework reports accuracy, precision, recall, macro-F1, and confusion-matrix analysis rather than relying on accuracy alone. Published results show why this design is necessary: early PlantVillage studies reported approximately 99% accuracy, while recent cross-domain analyses have found large performance reductions when models are transferred to field-like images. The paper does not claim new experimental numbers; instead, it defines a reproducible research design and a clear evidence base that can later be completed with original experiments. The central conclusion is that reliable crop disease detection should be evaluated not only by in-domain performance but also by how consistently the same disease classes are recognized under realistic image variation.