Development and validation of a weakly-supervised learning-based crop detection system for automated weeding in soybean cultivation

In this study, a Crop Recognition System for Automated Weeding of Soybean Fields was developed. A vision camera was employed for real-time image acquisition, and considering the field of view of 110°(D) x 86°(H) x 64°(V) and the ground distance, the camera was positioned at a height of 1 meter above the ground. The dataset was constructed by designating the soybean area as the region of interest (RoI). The deep learning model was trained using a weakly supervised learning method, leveraging labeled data. Central points of crops were detected through visualization, employing a Class Activation Map (CAM). The system's performance was evaluated using a linear regression approach, yielding a mean squared error (MSE) of 6.64 cm along the X-axis and 5.09 cm along the Y-axis, with root mean square error (RMSE) values of 1.24 cm on the X-axis and 2.25 cm on the Y-axis, respectively. These results demonstrate the high detection accuracy of the proposed system. An evasion success rate test was conducted in a controlled testbed environment to assess the system's practical applicability. 300 samples were tested over five repeated trials, achieving an average success rate of 98.7%. These results validate the system’s high success rate and operational stability.