Automating Agriculture: Using UAS and machine learning to monitor weed populations

Agriculture currently faces many challenges including a changing climate, the need to produce more nutritionally healthier foods, concerns about environmental impacts and a growing population to provide for. This means that new technologies and methodologies need to be applied to increase the efficiency of agriculture.
Alopecurus myosuroides presents a significant challenge to crop production owing to its competitive effects on yields of arable crops in the UK and beyond. Apart from effects on yields, weeds such as this are expensive and time-consuming to control. Precisely knowing where the weed is in a field potentially allows farmers to tailor management practices to the specific site, generating better control of the weed populations year to year.
Here I present an investigation into the use of Unmanned Aerial Systems (UAS) for mapping populations of A. myosuroides across multiple locations and years, whilst assessing the use of alternative machine learning techniques to model the data generated from the UAS. I undertook 3 seasons of field walking and UAS flights. This generated the largest known ground-truthed labelled dataset of remotely sensed images of an arable weed.
Using the first year’s data I developed a novel methodology for the combination of ground-based sampling in conjunction with aerial mapping to showcase the applicability of UAS for mapping A. myosuroides using linear models and random forests (published in Weed Research).
In the latter two field seasons I advanced the data collection methodology to create a more standardised process, and implemented a Convolution Neural Network, resulting in an improved ability to classify A. myosuroides in new previously unseen plots (published in Pest Management Sciences).
Finally, I assessed the use of alternative vegetation indices for the classification of A. myosuroides and compared the best performing CNN models to skilled human observers. This showed that Green Normalized Difference Vegetation Index (GNDVI) was the optimum index and that the models performed better than the observers across all vegetation indices.
Overall, this work established:
⦁ A methodology that can be expanded for the use of UAS to map plant populations.
⦁ The suitability of CNN for mapping A. myosuroides.
⦁ How data engineering can be used to increase the performance of CNN.
⦁ The need for further analysis into the factors driving the limited transferability of models.