Resumen
Wheat stripe rust poses a serious threat to the quality and yield of wheat crops. Typically, the occurrence data of wheat stripe rust is characterized by small sample sizes, and the current research on severity identification lacks high-precision methods for small sample data. Additionally, the irregular edges of wheat stripe rust lesions make it challenging to draw samples. In this study, we propose a method for wheat stripe rust severity identification that combines SLIC superpixel segmentation and a random forest algorithm. This method first employs SLIC to segment subregions of wheat stripe rust, automatically constructs and augments a dataset of wheat stripe rust samples based on the segmented patches. Then, a random forest model is used to classify the segmented subregion images, achieving fine-grained extraction of wheat stripe rust lesions. By merging the extracted subregion images and using pixel statistics, the percentage of lesion area is calculated, ultimately enabling the identification of the severity of wheat stripe rust. The results show that our method outperforms unsupervised classification algorithms such as watershed segmentation and K-Means clustering in terms of lesion extraction when using the segmented subregion dataset of wheat stripe rust. Compared to the K-Means segmentation method, the mean squared error is reduced by 1.2815, and compared to the watershed segmentation method, it is reduced by 2.0421. When compared to human visual inspection as the ground truth, the perceptual loss for lesion area extraction is 0.064. This method provides a new approach for the intelligent extraction of wheat stripe rust lesion areas and fading green areas, offering important theoretical reference for the precise prevention and control of wheat stripe rust.