Testing the ability of Unmanned Aerial Systems and machine learning to map weeds at subfield scales: a test with the weed Alopecurus myosuroides (Huds).

BACKGROUND: It is important to map agricultural weed populations in order to improve management and maintain future food security. Advances in data collection and statistical methodology have created new opportunities to aid in the mapping of weed populations. We set out to apply these new methodologies (Unmanned Aerial Systems - UAS) and statistical techniques (Convolutional Neural Networks - CNN) for the mapping of black-grass, a highly impactful weed in wheat fields in the UK. We tested this by undertaking an extensive UAS and field-based mapping over the course of two years, in total collecting multispectral image data from 102 fields, with 76 providing informative data. We used these data to construct a Vegetation Index (VI), that we used to train a custom CNN model from scratch. We undertook a suite of data engineering techniques, such as balancing and cleaning to optimize performance of our metrics. We also investigate the transferability of the models from one field to another. RESULTS: The results show that our data collection methodology and implementation of CNN outperform pervious approaches in the literature. We show that data engineering to account for "artefacts" in the image data increases our metrics significantly. We are not able to identify any traits that are shared between fields that result in high scores from our novel leave one field our cross validation (LOFO-CV) tests. CONCLUSION: We conclude that this evaluation procedure is a better estimation of real-world predictive value when compared to past studies. We conclude that by engineering the image data set into discrete classes of data quality we increase the prediction accuracy from the baseline model by 5% to an AUC of 0.825. We find that the temporal effects studied here have no effect on our ability to model weed densities.