Rapid On-Site Weed Identification with Machine Learning

Weed identification is a fundamental step in weed management. Traditional identification based on taxonomic features can be extremely challenging, especially at young seedling stage. It could also take days or months to confirm the identification through various channels, which would mean the loss of prime opportunity to control the weed. Recent advances in computer vision and machine learning have shown great success in various automatic visual detection tasks. It is therefore appropriate choice to capture visual field information and further process it to be able to realize autonomous weed identification promptly. This paper presents a convenient approach that applies image processing and machine learning for quick and accurate weeds identification on-site. Three deep models have been implemented to identify weeds via a smartphone. It is a proof-of-concept study targeting 16 selected most important agricultural weeds in Australia. We believe the proposed approach can help growers make a timely decision to spray the corresponding herbicide to reduce the financial loss annually.