Segmenting and grading the blast disease of rice leaves using VCDM-UNet

Rice is one of the main food crops in China. An efficient and accurate assessment of leaf blast disease severity is crucial to the 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 the rice leaf blast segmentation and disease severity grading. Subsequently, the training set was enhanced through some operations, such as contrast enhancement, flipping and rotation, regional erasure, and deformation. An improved UNet model (vgg16 coordinate dropout focal-dice mixed loss UNet, VCDM-UNet) was proposed to segment the leaf blast. The high efficiency and strong subjectivity were achieved in disease grading, compared with the traditional. Firstly, VGG16 was utilized as the backbone network of UNet to extract features of leaves and lesions, particularly for irregular lesion shapes. Secondly, the CA (coordinate attention) mechanism and Dropout were introduced in the upsampling module, in order to enhance the generalization of the model on the leaf and lesion pixels. Thirdly, LFD (focal-dice mixed loss) was adopted to improve the loss function. Sample imbalance was optimized to mitigate the impact of a large proportion of background pixels on model learning. Small proportions of leaves and lesions were then identified. The experimental results were as follows: 1) The segmentation effectiveness of the VCDM-UNet model was validated on rice leaf lesions. UNet, PSPNet, and DeepLabV3+ were employed on a test dataset to compare their segmentation. PSPNet exhibited significant deficiencies in producing noticeably incomplete edges when segmenting leaf images and rendering them largely unusable. Although DeepLabV3+ and UNet were fully capable of segmenting the leaf edges, the insufficient lesion feature was also captured, leading to suboptimal segmentation. In contrast to the rest three models, VCDM-UNet demonstrated greater accuracy in lesion segmentation, and closely matching manual annotation, thus showing a distinct advantage. 2) The VCDM-UNet was achieved a mIoU (mean intersection over union) of 82.93%, mPA (mean pixel accuracy) of 88.87%, and F1-Macro of 89.96%, indicating the 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 for leaves and lesions. In terms of model segmentation efficiency, the number of parameters was reduced by 82.8, 113.8, and 22.8 M, compared with PSPNet, DeepLabV3, and UNet, respectively. Nevertheless, the prediction time also increased by 1.00, 0.38, and 0.21 s, respectively. Considering that segmentation accuracy was more important than segmentation efficiency in the tasks of disease image segmentation, indicating that the slight increase was reasonable in prediction time. 3) The severity of rice leaf blast disease was further assessed, according to the GB/T15790-2009 standard and the leaf blast disease severity calculation formula. In this case, the average grading accuracy of VCDM-UNet reached 83.95%, thus outperforming the comparative models. It was verified that the improved model can provide technical support to the grading of rice leaf blast disease severity. Overall, the improved model demonstrated the better performance on individual and whole rice leaves, even with a relatively small sample size. The finding can lay the foundation for the practical application of large-scale leaf blast disease grading in the next steps. Additionally, since leaf recognition in natural environments was 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.