Featured, Marijuana News

Using a fully convolutional neural network for detecting locations of weeds in images from cereal fields

Using a fully convolutional neural network for detecting locations of weeds in images from cereal fields

Information about the presence of weeds in fields is important in order to decide on a weed control strategy. This is especially crucial in precision weed management, where the position of each plant is essential for conducting mechanical weed control or patch spraying.
Although cereal fields make up a large portion of the world’s farmland, cereal fields are often ignored in relation to precision weed control. This gap in existing research may be explained by the narrow spacing in cereal crop rows, which is usually only a few centimeters, and thus often causes overlap between plants.
Deep-learning has proven capable of outperforming previous methods of classification within various machine vision tasks, including classification of plants and semantic segmentation of field images. In this study, we present a system capable of detecting weeds in images from cereal fields despite occlusion between the plants in the field. Due to this occlusion, most of the weeds are partially hidden behind or touching the crops or other weeds.
For detecting weeds, this study proposes a fully convolutional neural network, which assigns a bounding box to single weeds in images and classifies each one as a monocot or dicot. The network has been trained on over 10,000 weed annotations in high-resolution RGB images from Danish wheat fields.
In order to determine the weed locations, the network uses a set of default bounding boxes, for which it generates a score, determining whether monocot or dicot weeds are present at the location of each box. The network consists of a backbone convolutional network, which is connected to a set of feature extraction layers that determine the offset of the default bounding boxes and also how the boxes should be stretched to fit the plant shape in the images.
The weed detection network has been evaluated on an Nvidia Titan X, on which it is able to process a 5MPx image in 0.02s, making the method suitable for real-time field operation.
For mechanical weed control, this network is sufficient. However, for chemical weed control, we also need to know the weed species in order to choose which herbicides to use. Therefore, we combine the proposed convolutional neural network for detecting weeds with another convolutional neural network optimized for the classification of weed species, thereby allowing for the creation of weed distribution maps of individual weed species.

credit:pure.au.dk