Weed Mapping with Convolutional Neural Networks on High Resolution Whole-Field Images

Weed mapping is a technique used to identify and locate harmful weed plants in farm fields. Accurate weed mapping enables targeted herbicide application and helps plant scientists to estimate the effectiveness of field experiments. In this paper we discuss a highly practical and effective working pipeline to weed map a wheat field combining GIS and deep learning technology. This pipeline is an end-to-end process including using an unoccupied aerial vehicle (UAV) to collect ultra-high definition whole-field images, labelling and training deep learning models and an efficient evaluation process for the resulting weed map. We show that our method can generate accurate pixel-wise weed maps by only training on small regions of the field, and can generalize well when making predictions back on the larger wholefield orthomosaic image.