Improved weed segmentation in UAV imagery of sorghum fields with a combined deblurring segmentation model

Abstract

Abstract Background Efficient and site-specific weed management is a critical step in many agricultural tasks. Image captures from drones and modern machine learning based computer vision methods can be used to assess weed infestation in agricultural fields more efficiently. However, the image quality of the captures can be affected by several factors, including motion blur. Image captures can be blurred because the drone moves during the image capturing process, e.g. due to wind pressure or camera settings. These influences complicate the annotation of training and test samples and can also lead to reduced predictive power in segmentation and classification tasks. Results In this study, we propose , a combined deblurring and segmentation model for weed and crop segmentation in motion blurred images. For this purpose, we first collected a new dataset of matching sharp and naturally blurred image pairs of real sorghum and weed plants from drone images of the same agricultural field. The data was used to train and evaluate the performance of on both sharp and blurred images of a hold-out test-set. We show that outperforms a standard segmentation model that does not include an integrated deblurring step, with a relative improvement of $$13.4 \%$$ 13.4 % in terms of the Sørensen-Dice coefficient. Conclusion Our combined deblurring and segmentation model is able to accurately segment weeds from sorghum and background, in both sharp as well as motion blurred drone captures. This has high practical implications, as lower error rates in weed and crop segmentation could lead to better weed control, e.g. when using robots for mechanical weed removal.