The role of synthetic imagery in training AI to detect blackgrass in wheat

Introduction

In modern agriculture, one of the most persistent challenges is the management of weeds like blackgrass in wheat. A highly competitive weed, blackgrass can severely impact crop yields if not effectively controlled. In the UK, in the face of blackgrass resistance to herbicides in recent years, the response is to drill the wheat later to minimise blackgrass germination and density. We know this is costing the industry £400 million a year!

To address this, AgriSynth received a government grant for a 453,000 project to use synthetic imagery to train an AI solution.

Why Synthetic Imagery?

AI models, particularly those based on deep learning, require vast amounts of labelled data to distinguish between crops and weeds accurately. However, collecting real-world images of blackgrass in various stages of growth and under different environmental conditions can be time-consuming and expensive. Labelling those images is even more difficult and impossible to do accurately for any volume. This is where synthetic imagery comes into play. 

Synthetic imagery refers to computer-generated images that simulate real-world conditions. These images can depict blackgrass in different stages of growth, lighting conditions, and angles, which might be challenging to capture comprehensively in the field. By creating synthetic images of wheat with varying densities and distributions of blackgrass, in various lighting conditions, we can generate large and diverse datasets where the scope and variance of those datasets in determined and known.

Benefits of Synthetic Imagery

1. Diversity and Scalability - Synthetic imagery allows for the creation of a virtually unlimited variety of scenarios, ensuring that the AI model can generalize well to real-world situations. This diversity helps the model learn to identify blackgrass even in uncommon or complex conditions.

2. Cost-Effectiveness - Generating synthetic images is much more cost-effective than collecting and annotating real-world data, especially when dealing with large-scale agricultural environments

3. Controlled Environment - We can control the variables in synthetic images, such as the presence of other plants, soil types, and weather conditions. This control helps in fine-tuning the AI model to detect blackgrass with higher accuracy.

4. Enhanced Model Robustness - By training on synthetic imagery, the AI model becomes more robust and capable of handling variations it might encounter in real-world fields, reducing the risk of false positives or negatives.

Challenges and Considerations

While synthetic imagery offers significant advantages, it is essential to ensure that the synthetic data closely resembles real-world conditions. Any discrepancy between the synthetic and real images can lead to a performance gap. We test extensively to ensure that this is not the case. Additionally, the large numbers of images we can generate together with the 100% pixel-perfect labelling, overcomes minor differences between synthetic images and those of the real world. During the project, we collected 300,000 images from real-world trials with differing but known densities of wheat and blackgrass in differing combinations. This gave a robust test of our trained AI models.

Outcome

AgriSynth completed the government Smart grant project successfully to full claim by the monitoring team.