Kazi Sanam Salman, Palkar Bhakti, Mishra Dhirendra
Department of Computer Engineering, K. J. Somaiya School of Engineering (KJSSE), Somaiya Vidyavihar University, Mumbai, Maharashtra 400077, India.
Department of Engineering and Technology, Bharati Vidyapeeth Deemed to be University, Kharghar, Navi Mumbai, Maharashtra 410210, India.
MethodsX. 2025 Aug 5;15:103551. doi: 10.1016/j.mex.2025.103551. eCollection 2025 Dec.
Convolution Neural Networks (CNN) are best in their ability to detect rice diseases but still face challenges in generalizing equally well for all classes of disease in multiclass classification. Detecting rice crop disease like sheath rot is still challenging due to unavailability of dataset and intraclass variations in symptoms. Transfer learning models take more resources for execution due to its deep architecture. To conquer these challenges, VCNet, an optimized, novel and efficient multiclass rice crop disease detection framework is proposed. The study focuses on developing a shallow model with deep feature extraction to bring down the computational load with reduced time for training without compromising on any performance parameters. Further the model goes through two level optimization process where optimal hyperparameters identified through experimentation is given as parameters to genetic algorithm for optimization of VCNet during training. Novel dataset containing field images is generated with the help of plant pathologist to improve model capability to identify diseases. Rigorous empirical comparison and evaluation with state-of-the-art models for each class of disease is done to validate proposed technique. VCNet outperforms the existing transfer learning models with training accuracy 99.72 % and testing accuracy 97.71 %. It also requires fewer parameters and takes minimum training time.•The major contribution of this study is the design of an optimized, efficient and enhanced deep learning technique for multiclass rice crop disease detection embracing with batch normalization, dropout and genetic optimization algorithm to improve generalization power and restrict the overlearning capability for seen and unseen data.•Proposed VCNet, a shallow model with deep feature extraction, employs VGG16 layers for initial extraction fused with custom CNN architecture to correctly detect the challenging classes of diseases like sheath rot in multiclass classification.•The most significant observation is that VCNet accurately predicts the rice disease for each class of diseases under study whereas the existing powerful models largely misclassified for some classes of diseases in multiclass classification.
卷积神经网络（CNN）在检测水稻病害方面能力卓越，但在多类分类中对所有病害类别进行同等良好的泛化仍面临挑战。由于缺乏数据集以及症状的类内变化，检测诸如鞘腐病等水稻作物病害仍然具有挑战性。迁移学习模型因其深度架构而需要更多资源来执行。为了克服这些挑战，提出了VCNet，这是一个优化、新颖且高效的多类水稻作物病害检测框架。该研究专注于开发一种具有深度特征提取的浅层模型，以降低计算负载并减少训练时间，同时不影响任何性能参数。此外，该模型经过两级优化过程，通过实验确定的最优超参数作为参数输入遗传算法，在训练期间对VCNet进行优化。在植物病理学家的帮助下生成了包含田间图像的新数据集，以提高模型识别病害的能力。针对每种病害类别，与现有先进模型进行了严格的实证比较和评估，以验证所提出的技术。VCNet的训练准确率为99.72%，测试准确率为97.71%，优于现有的迁移学习模型。它还需要更少的参数，并花费最少的训练时间。
•本研究的主要贡献在于设计了一种优化、高效且增强的深度学习技术，用于多类水稻作物病害检测，该技术采用批量归一化、随机失活和遗传优化算法，以提高泛化能力，并限制对可见和不可见数据的过学习能力。
•所提出的VCNet是一个具有深度特征提取的浅层模型，采用VGG16层进行初始提取，并与自定义CNN架构融合，以在多类分类中正确检测诸如鞘腐病等具有挑战性的病害类别。
•最显著的观察结果是，VCNet能够准确预测所研究的每种病害类别的水稻病害，而现有的强大模型在多类分类中对某些病害类别存在大量误分类。
