摘要
Maize, which is the primarycrop in many sub-Saharan countries, including Zambia, is susceptible to a wide range of diseases that have a significant impact on food production. To tackle this challenge and improve disease detection efficiency, deep learning methods have been employed to accurately classify and identify plant diseases. In recent times, manual inspection of maize fields for disease detection has been the standard practice in many parts of Zambia. However, this approach is not only time-consuming but also impractical for large-scale agricultural operations. Hence, the development of precise and automated classification models has become crucial in modern agriculture. In this study, we propose a novel deep-learning model called DenseSwin, specifically designed for maize disease classification in both the early visible stage and late indisputable stage of the disease. DenseSwin combines the strengths of densely connected convolution blocks with a shifted windows-based multi-head self-attention mechanism. This unique fusion of techniques enables the model to effectively capture intricate patterns and features in maize plant images, thereby enhancing disease classification performance. Through extensive experimentation and evaluation, DenseSwin achieves an impressive accuracy of 97.18%. These results highlight the model's remarkable ability to accurately detect and classify maize diseases, offering promising potential for real-world applications in agricultural settings, particularly in Zambia.