Northern maize leaf blight disease detection and segmentation using deep convolution neural networks

Maize is ranked as the third most important food crop in India after rice and wheat. The cultivation of this particular crop and its associated agricultural practices have faced many challenges over the centuries. Maize is susceptible to Northern Leaf Blight (NLB), a highly contagious fungal foliar disease. The primary reason for the reduction in yield resulting from NLB is the loss of photosynthetic leaf area. Therefore, identifying diseases at an early stage is critical to ensure that the treatment process is carried out correctly and that the quality of results is maintained. This research uses a deep learning-based Attention U-Net model to explore a real-time, effective approach for segmentation and detecting NLB diseases in maize crops. Data augmentation is performed after image annotation to increase the model's effectiveness, and the model is trained from scratch. The model was trained for a maximum of 50 epochs using an initial learning rate of 0.0001 with Adam as an optimizer, and its performance was tested on the test dataset. This study shows that the Attention U-Net model outperformed other image segmentation methods, such as Res U-Net and Plain U-Net, and showed better results with an Intersection over Union (IoU) of 72.41%, 70.91% and 51.95%, respectively. The proposed model achieves an average pixel-wise F1 score of 85.23%. The diseased segmentation accuracy clenched to 98.97%, and the Dice coefficient (DC) of disease spot segmentation is 81.39%. Adding an attention mechanism to the U-Net architecture improves its ability to express local features, resulting in improved NLB disease image segmentation performance.