Deep learning-based early weed segmentation using motion blurred UAV images of sorghum fields

Weeds are undesired plants in agricultural fields that affect crop yield and quality by competing for nutrients, water, sunlight and space. Site-specific weed management (SSWM) through variable rate herbicide application and mechanical weed control have long been recommended in order to reduce the amount of herbicide and impact caused by uniform spraying. Accurate detection and classification of weeds in crop fields is a crucial first step for implementing such precise strategies. Drones are commonly used for image capturing but high wind pressure and different drone settings have a severe effect on the image quality, which potentially results in degraded images, e.g. due to motion blur. We publish a manually annotated and expert curated drone image dataset for weed detection in sorghum fields under challenging conditions. Our results show that our trained models generalize well regarding the detection of weeds, even for degraded captures due to motion blur. An UNet-like architecture with ResNet-34 as feature extractor achieved an F1-score of over 89 % on a hold-out test-set. Further analysis indicate that the trained model performed well in predicting the general plant shape, while most mis-classifications appeared at borders of the plants. Beyond that, our approach can detect intra-row weeds without additional information as well as partly occluded plants in contrast to existing research.Github link: https://github.com/grimmlab/UAVWeedSegmentatio