Predicting wheat scab levels based on rotation detector and Swin classifier

Abstract

Wheat scab is a highly destructive disease that adversely impact wheat crops throughout their growth cycle. It is crucial to promptly evaluate the levels of wheat scab in the field to prevent its spread. Manual observation, however, is inefficient and time-consuming. Recent research has indicated that computer vision-based methods can enhance efficiency in this regard. This study proposed a method for predicting wheat scab levels using a rotation detector and Swin classifier.

To minimise background interference, the study incorporated the rotation wheat detector (RWD) network for detecting wheat heads. The RWD network employed the Kalman filter Intersection over Union (KFIoU) to predict the angle, thereby improving accuracy. The Swin wheat classifier (SWC) network was employed to classify healthy and diseased wheat heads. The SWC network benefited from the shifted window self-attention module (SW-MSA), which enhanced feature extraction by establishing connections with other windows. The proposed method was evaluated using wheat field images collected over 3 years. The results demonstrate promising performance, achieving a 96% accuracy in predicting wheat scab levels. Furthermore, the R² and RMSE values for diseased wheat count were 97.62% and 3.61, respectively. This method offers an accurate means of predicting wheat scab levels through the analysis of wheat field images. Additionally, the introduction of the rotation detector presents a novel contribution to research on wheat scab detection.