Optimizing wheat scab in remote sensing monitoring accuracy using interridge background elimination

Wheat scab has been one of the most severe diseases for grains in recent years. Unmanned aerial vehicles (UAV) can be used to monitor the wheat scab using multispectral remote sensing. In this study, a series of operations were carried out to eliminate the inter-ridge background elements (such as soil and shadow intercage background elements) under natural conditions. The precision segmentation of inter-ridge background was then realized to improve the identification accuracy of wheat scab. A high-precision monitoring model was also constructed for the wheat scab. Specifically, the research object was selected as the wheat at the filling stage in the Zhenjiang Academy of Agricultural Sciences in Jiangsu Province, China. Rededge MX multispectral camera (five channels) was installed on a DJI M600 Pro six-rotor UAV to capture the multispectral images of the wheat canopy. Multiple vegetation indices were then calculated using the spectral features of the wheat canopy. A correlation analysis was implemented to determine the highest correlation vegetation indices (VIs) with the wheat scab, and the gray scale co-occurrence matrix (gray-level co-occurrence matrix, GLCM and texture feature (TFs). Otus, threshold segmentation and support vector machine (SVM) were used for the semantic segmentation of wheat scab multispectral images. There was a reduced influence of field edge shadow and soil background elements on the identification accuracy of wheat scab. Partial least squares regression (PLSR) was used after eliminating the ridge background. The regression model was established for the VIs, TFs, Vis&TFs and wheat scab disease index (DI). The regression equation of the monitoring model was then constructed to draw the change chart of disease severity. The experimental results show that the best performance was achieved in the visual interpretation threshold segmentation (overall accuracy: 92.06%, Kappa coefficient: 0.84), and the OTSU threshold segmentation (overall accuracy: 90.52%, Kappa coefficient: 0.81) in the inter-row background elimination model. In the wheat scab monitoring model, the accuracy of the Vis&TFs model was higher than that of the VIs and TFs model, and the accuracy of the refined semantic segmentation model was better than that of the unsegmentation model. Taking the model on May 8 as an example, the training set and validation set R² of the VIs-PLSR undivided model were 0.71 and 0.73, RMSE were 5.61 and 6.72, and RPD were 1.83 and 1.89, respectively. The training set and validation set R² of the TFs-PLSR undivided model were 0.64 and 0.62, RMSE were 6.03 and 6.92, and RPD were 1.67 and 1.69, respectively. Vis&TFs undivided model training set and validation set R² were 0.73 and 0.61, RMSE were 6.31 and 6.68, RPD were 1.93 and 1.89, respectively. The training set and validation set R² of VIs&TFs fine segmentation model were 0.78 and 0.81, RMSE were 4.73 and 4.31, and RPD were 2.26 and 2.05. In summary, the refined semantic segmentation with the inter-ridge background feature effectively improved the accuracy of the wheat scab monitoring model. The monitoring model was realized to classify the diseases, and then to generate the spatial distribution map of wheat scab grade. The disease distribution of wheat scab was obtained to predict the trend of wheat scab diseases. The finding can also provide effective guidance for the variable application in the late stage of wheat cultivation.