A Review of Machine Learning and Deep Learning Techniques for Automated Tomato Leaf Disease Detection

Abstract

Tomato is an economically important crop, and disease infection can reduce plant health, productivity, and fruit quality. Image-based machine learning (ML) and deep learning (DL) methods have therefore been investigated for automated recognition of tomato leaf diseases. This narrative review examines the progression from handcrafted feature-based ML methods to convolutional neural networks (CNNs), transfer learning, lightweight architectures, attention mechanisms, and recent model-compression and multi-task approaches. The review considers the data characteristics, preprocessing strategies, learning paradigms, evaluation conditions, computational requirements, and evidence of generalization reported by representative studies. The literature shows that conventional ML can perform well on small, carefully prepared datasets but depends on the quality of handcrafted representations. CNN-based methods learn task-specific visual representations and generally provide stronger performance on multi-class tomato leaf image datasets. However, very high reported accuracies are frequently obtained on controlled or curated datasets, particularly PlantVillage-derived data, and therefore should not be interpreted as direct evidence of field performance. Studies using field images or explicitly measuring model size, FLOPs, inference speed, or mobile/edge execution provide stronger evidence concerning practical deployment. The principal research gaps are limited field diversity, illumination and background variation, class imbalance, weak cross-dataset validation, inconsistent evaluation protocols, and insufficient reporting of deployment metrics. Future research should therefore evaluate not only classification accuracy but also robustness, independent validation, computational cost, and reproducible deployment conditions.