Scholarlink

Plant leaf disease identification is a critical use of modern precision agriculture that helps in timely diagnosis and inhibitor of damage to crops. Though several deep learning architectures have been applied to this use, there has been a shortage of identifying the optimal combination of pre-trained model and optimizer for achieving maximum classification performance. Much previous work has focused on using individual models such as ResNet, VGG, and EfficientNet, typically with default optimizers, without comparative analysis. Unlike prior works that only evaluated individual CNN models with default optimizers, our approach systematically selects the best-performing optimizer for each architecture and combines them in an ensemble, resulting in a novel hybrid model with improved generalization. This paper extensively compares three pre-trained CNN models, ResNet152, VGG19, and EfficientNetB7, on a big leaf disease dataset. Three optimizers, SGD, RMSprop, and Adam, are employed to identify the optimal optimization method for every architecture. The performance of models is evaluated to determine the best-performing configuration. At last, we propose a hybrid deep learning model by mixing the top-performing architectures and optimizers and enhancing classification accuracy further. A comprehensive comparison of previous work identifies the research gap in determining the ideal optimizer for pre-trained models in leaf disease detection. Our findings reveal that even though Effi-cientNetB7 achieves the highest single accuracy with RMSProp (99%), the hybrid model introduced in this paper performs better than all single architectures at an accuracy rate of 99.2%. This optimizer-aware hybrid strategy, to our knowledge, has not been explored in prior work and provides a new direction for improving classification performance in plant disease detection. The experiments reveal that an optimally configured hybrid approach based on multiple architectures and optimizers can significantly improve leaf disease detection performance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leaf Disease Classification on Real Time Data: A Deep Learning Approach with Performance Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sanika Pendurkar,
  • Heemit Shah,
  • Smita Sankhe

摘要

Plant leaf disease identification is a critical use of modern precision agriculture that helps in timely diagnosis and inhibitor of damage to crops. Though several deep learning architectures have been applied to this use, there has been a shortage of identifying the optimal combination of pre-trained model and optimizer for achieving maximum classification performance. Much previous work has focused on using individual models such as ResNet, VGG, and EfficientNet, typically with default optimizers, without comparative analysis. Unlike prior works that only evaluated individual CNN models with default optimizers, our approach systematically selects the best-performing optimizer for each architecture and combines them in an ensemble, resulting in a novel hybrid model with improved generalization. This paper extensively compares three pre-trained CNN models, ResNet152, VGG19, and EfficientNetB7, on a big leaf disease dataset. Three optimizers, SGD, RMSprop, and Adam, are employed to identify the optimal optimization method for every architecture. The performance of models is evaluated to determine the best-performing configuration. At last, we propose a hybrid deep learning model by mixing the top-performing architectures and optimizers and enhancing classification accuracy further. A comprehensive comparison of previous work identifies the research gap in determining the ideal optimizer for pre-trained models in leaf disease detection. Our findings reveal that even though Effi-cientNetB7 achieves the highest single accuracy with RMSProp (99%), the hybrid model introduced in this paper performs better than all single architectures at an accuracy rate of 99.2%. This optimizer-aware hybrid strategy, to our knowledge, has not been explored in prior work and provides a new direction for improving classification performance in plant disease detection. The experiments reveal that an optimally configured hybrid approach based on multiple architectures and optimizers can significantly improve leaf disease detection performance.