A hybrid deep learning neural network for early plant disease diagnosis using a real-world Wheat-Barley vision dataset: challenges and solutions

Approximately 35% of India's annual crop yield is lost due to plant diseases. Due to a lack of lab equipment and infrastructure, early diagnosis of plant diseases remains challenging. Categorization and detection of foliar diseases is a rapidly evolving research subject in which machine learning and neural computing concepts are employed to assist agricultural businesses. The lack of adequate large-scale data sets remains a significant barrier to enable vision-based plant disease diagnosis. A possible approach to solving such a problem is to use a publicly available dataset. Using a publicly available dataset from the internet raises a slew of concerns. Significant challenges include employing such datasets from various geographical regions deployed in another location, model overfitting owing to small dataset size, etc. In this research study, we release a novel "Wheat and Barley dataset," which features wheat and barley grain images categorized into three major disease types (yellow rust, brown rust, and loose smut). Additionally, the research introduces an innovative hybrid deep learning neural network that leverages transfer learning and feature concatenation from MobileNet and DenseNet architectures. Extracted features undergo dimensionality reduction using particle swarm optimization before being integrated into a conventional learning algorithm. Empirical findings validate the effectiveness of concatenated features in improving classification performance. The study assesses the performance of three traditional machine learning classifiers: support vector machine, decision tree, and random forests, with the latter exhibiting superior accuracy at an average of 98.89%. This investigation provides valuable insights into plant disease diagnosis and overcoming challenges through an impactful hybrid deep learning approach.