Core Concepts
Highly accurate deep learning models can effectively detect the presence of blackgrass, a problematic weed, in wheat and barley crops using multispectral imaging.
Abstract
This study presents a large dataset of over 15,000 multispectral images of wheat, barley, and blackgrass collected from 51 fields across 8 different soil types in the UK. The dataset is used to evaluate the performance of state-of-the-art deep learning models, including ResNet-50, EfficientNet B4, and Swin Transformer, in classifying images as containing blackgrass or not.
The key findings are:
The models achieve high accuracy, with Swin Transformer performing the best at 87.7% accuracy, followed by ResNet-50 at 87.3% and EfficientNet B4 at 83%.
The inclusion of near-infrared (NIR) spectral information is crucial, with NIR alone outperforming RGB and RGB+NIR combinations. Using all available spectral bands yields the highest accuracy.
The models perform better on late-season crops compared to mid-season, and on wheat compared to barley. This suggests the visual differences between the crop and weed become more pronounced as the plants mature.
Increasing the training data quantity improves performance up to around 6,000 images, beyond which there is no significant further improvement. This indicates the dataset provides sufficient diversity to train effective models.
The large-scale dataset and comprehensive evaluation of deep learning models for this challenging fine-grained visual classification task can accelerate the development of precision weed management technologies for major cereal crops.
Stats
The area of global cropland devoted to growing cereals, and in particular rice, wheat and maize, is orders of magnitude greater than that of many vegetable crops.
Wheat is second only to rice as a global staple, with a global consumption of 65.6 kg per person per year.
Quotes
"Efforts to reduce herbicide usage in staple cereal crops have the potential to deliver significant impact."
"Grass weeds are a particular problem in wheat production due to their biological similarities."
"Blackgrass is one of the most economically damaging weeds in Europe, so effective strategies to manage populations are a priority."