Advancements in maize leaf disease detection, segmentation and classification: A review

Abstract

Maize is one of the most widely produced and consumed crops in the world. Production and quality are directly dependent on crop health. Many of the machine-learning (ML) and deep-learning (DL) approaches for maize leaf disease detection, segmentation and classification (MLDDSC) have been implemented for crucial tasks in sustainable agriculture. A total of 82 papers between the years 2020 and 2024 were selected after applying preliminary selection criteria focusing on the review's major goal. In this review paper, the latest developments of MLDDSC in the context of dataset sources, image pre-processing, image augmentation, feature extraction, evaluation metrics, machine-learning architectures, deep-learning architectures, and customisation techniques. The paper also discusses the challenges and future directions of research in MLDDSC, such as severity measurement, hyperspectral imaging, and lightweight models. Finally, a systematic and in-depth analysis is provided of the state-of-the-art methods and techniques for MLDDSC to highlight the potential and limitations of each approach. Overall, from the comparative analysis among the selected papers for review, it was found that multimodal logistic regression outperformed all ML algorithms, whereas pre-trained GoogleNet was efficient among DL models. Likewise, a customised model with fusion of inception and residual structure and a transfer learning model with EfficientNet outperformed all others. Regarding severity measurement, diseased leaf area was the most significant, but the techniques for calculating area can differ. The review also provides a taxonomy and comparison of the existing methods and techniques and identifies the research gaps and opportunities for further improvement.