Artificial Intelligence-Driven Weed Segmentation for the Spot-Spraying in Blueberry Orchards

Excessive use of herbicides has detrimental effects on the environment and food quality. To mitigate this, autonomous systems are being developed for the purpose of precise spot-spraying of individual weed patches. Fast and reliable segmentation of weeds in the image is a crucial perception step in the process of autonomous weed spraying. Deep learning models have shown outstanding success with this task. In this paper, 120 RGB images with weed annotations from blueberry orchards are used to train the Unet++ model and compare multiple encoder depths and input image sizes. Along with this, the benefits of multiple frequently used augmentation approaches are evaluated. Results indicate that encoder depth and input size have minimal impact on the segmentation of both small and large weed patches. The gain from image augmentation is shown to be substantial, however a careful definition of augmentations is needed to achieve the greatest benefit from its application. With the mean IoU of 78.4% and good qualitative performance observed on a set of unseen image samples, described models show promising results for the task of weed segmentation, especially in early growing stages of weed.