A novel approach for crop disease detection using transfer-learning-based deep convolutional neural networks

Crop production quality can be enhanced through effective early stage monitoring of crops. In contemporary agriculture, precision farming systems that incorporate advanced technologies, such as drones, are employed for monitoring. Nevertheless, current models lack the ability to be applied broadly across various agricultural settings, leading to the suboptimal detection of crop diseases. Therefore, this study proposes a novel method for monitoring crop diseases using the Transfer-Learning-based Deep Convolutional Neural Network technique. The process commenced with the collection of a multi-crop multi-disease dataset, followed by data preprocessing through the application of Bilateral Filtering and Partial Dynamic Range Adjustment with Histogram Equalization. Subsequently, the dimensionality of the images was reduced using the Multi-Weighted Covariance Matrix Component Analysis method, and the data were organized using the K-Means Clustering algorithm. In addition, various features were extracted, with the most relevant features selected using the Mountain Gazelle Pareto Front Optimizer algorithm. To identify distinct disease characteristics, a Multidimensional Feature Distribution plot was created using the chosen features. Ultimately, the assessed spectral index of the clustered color bands, optimal features, and Multidimensional Feature Distribution plot were used to detect multi-crop diseases using the proposed Transfer Learning based Dynamic Contextual Feature Pooled Convolutional Neural Network model. The effectiveness of the proposed method was evaluated through an experimental analysis using a publicly accessible dataset. The proposed method demonstrated accuracy, precision, recall, and F1 score values of 98.09%, 97.24%, 97.58%, and 97.99%, respectively, which is higher than those of existing methods.