Improved multiscale attention based deep learning approach for automated sugarcane leaf disease detection using BSRI data

Early and accurate detection of sugarcane leaf diseases is critical for improving crop productivity and reducing economic losses in the agricultural sector. Timely interventions enable sustainable crop management and better resource use. In this study, we propose a deep learning-based approach for sugarcane leaf disease classification that leverages a novel architecture, the Multi-scale Attention-based Dense Residual Network (MADRN). The MADRN model integrates dense residual learning and multi-scale attention mechanisms to effectively capture fine-grained, disease-specific features and address challenges related to domain variability and complex data patterns. Two datasets are used to evaluate the model: a Kaggle dataset and a blended dataset created by combining Kaggle images with those from the Bangladesh Sugarcrop Research Institute (BSRI), simulating real-world conditions. All images undergo preprocessing steps, including resizing, normalization, and data augmentation, before training. Additionally, several baseline models (CNN, VGG16, MobileNetV2, and XceptionNet) are fine-tuned and compared with the MADRN model. Experimental results demonstrate that MADRN consistently outperforms baseline models in accuracy, precision, recall, and F1-score across both datasets, achieving up to 94.78% accuracy on the Kaggle dataset and 92.25% on the blended dataset. These findings highlight MADRN's superior ability to learn discriminative features and generalize effectively across diverse data sources, making it a promising tool for precision agriculture and disease management. To facilitate practical implementation, a web-based application is developed, enabling real-time and user-friendly disease detection. This research lays a strong foundation for the development of accurate, scalable, and practical disease classification tools that can support sustainable agricultural practices.