Segmentation and grading of rice leaf blast disease based on VCDM-UNet

Rice is one of the main food crops in China. Efficient and accurate assessment of leaf blast disease severity is crucial for early prevention, precision application of pesticides, and yield loss prediction in rice. In this study, the indica rice variety CO39 at the three-leaf stage was selected as the research sample. The samples were collected and placed in a laboratory environment for image acquisition using a Nikon D3200 camera. A high-quality dataset was constructed for rice leaf blast segmentation and disease severity grading. Subsequently, the training set was enhanced through methods such as contrast enhancement, flipping and rotation, regional erasure, and deformation. To address the issues of low efficiency and strong subjectivity in traditional disease grading methods, an improved UNet model (VGG16 Coordinate Dropout Focal-Dice Mixed Loss UNet, VCDM-UNet) for leaf blast segmentation had been proposed. Firstly, to tackle the problem of irregular lesion shapes and difficulty in identification, VGG16 was utilized as the backbone network of UNet to enhance the model's ability to extract features of leaves and lesions. Secondly, to increase the model's focus on leaf and lesion pixels and to enhance its generalization capability, CA (coordinate attention) mechanism and Dropout were introduced in the upsampling module. Then, to address the issue of small proportions of leaves and lesions, L_FD(focal-dice mixed loss) was adopted to improve the model's loss function, optimizing sample imbalance and mitigating the impact of a large proportion of background pixels on model learning. The experimental conclusions were as follows:1) To validate the segmentation effectiveness of the VCDM-UNet model on rice leaf lesions, this study employed UNet, PSPNet, and DeepLabV3+ on a test dataset to compare their segmentation outcomes. PSPNet exhibited significant deficiencies, producing noticeably incomplete edges when segmenting leaf images, rendering it largely unusable. Although DeepLabV3+ and UNet were capable of segmenting the leaf edges fully, they captured insufficient lesion feature information, leading to suboptimal segmentation results. In contrast to the other three models, VCDM-UNet demonstrated greater accuracy in lesion segmentation, closely matching manual annotation results, thus showing a distinct advantage. 2) The VCDM-UNet achieved an mIoU(mean intersection over union) of 82.93%, mPA(mean pixel accuracy) of 88.87%, and F1-Macro of 89.96%, representing improvements of 12.43, 11.43, and 9.64 percentage points over PSPNet, 6.97, 7.32, and 5.52 percentage points over DeepLabV3+, and 3.37, 4.60, and 2.59 percentage points over UNet, respectively, with the best feature extraction effects for leaves and lesions. In terms of model segmentation efficiency, the number of parameters was reduced by 82.8M, 113.8M, and 22.8M compared to PSPNet, DeepLabV3, and UNet, respectively. However, the prediction time increased by 1, 0.38, and 0.21 seconds, respectively. Considering that segmentation accuracy is more important than segmentation efficiency in disease image segmentation tasks, the slight increase in prediction time is reasonable. 3) Based on the segmentation results of the model, the severity of rice leaf blast disease could be further assessed according to the GB/T15790-2009 standard and the leaf blast disease severity calculation formula. In this process, the average grading accuracy of VCDM-UNet reached 83.95%, outperforming the comparative models. It had been verified that the model can provide technical support for the grading of rice leaf blast disease severity. Overall, the model presented in this paper had demonstrated good performance on individual and whole rice leaves, even with a relatively small sample size, which can lay the foundation for the practical application of large-scale leaf blast disease grading detection in the next steps. Additionally, since leaf recognition in natural environments is influenced by more complex factors, such as the diversity of leaf postures, mutual occlusion between leaves, and multiple diseases on a single leaf, it is necessary to expand the sample size and diversity of rice leaf data in the future. Further improvements to the model are also required to enhance its adaptability and robustness in complex environments.