Enhancing Domain Generalization in Crop and Weed Segmentation via Entropy Minimization

Weeds pose a significant threat to crops and must be effectively managed. Manual weed management is labour intensive and time-consuming, underscoring the need for efficient solutions. Utilising robotic systems presents a promising alternative. These systems require advanced vision-based capabilities to accurately distinguish between crops and weeds. To enhance the robot's vision performance for automatic differentiation, this study proposes a method that jointly trains a segmentation model using both labelled and unlabelled data. The labelled data is trained in a supervised manner, while the unlabelled data is incorporated through an entropy minimisation mechanism, promoting learning from diverse data sources. To further enhance generalisation across different fields, we employ instance selective whitening, which enables the network to become style-agnostic and focus on contextual features. Our proposed method achieved a mean Intersection over Union (mIoU) of 0.938 in in-domain experiments, and 0.644 and 0.565 in out-domain experiments on two distinct datasets. These results demonstrate the robust performance of our method across various agricultural datasets, underscoring its potential for practical applications in precision agriculture. This approach not only improves the accuracy of robotic weed detection but also offers a scalable solution for diverse agricultural environments.