High-Throughput Molecular Labelling to Train Image Classifiers

Integrated Pest Management (IPM) aims to reduce the risks and impacts of pesticide use and promote ecologically-sound agricultural practices. This relies on an intimate understanding of the ecology of pest and beneficial insects and how their population dynamics can be influenced to maintain crop productivity and profitability. IPM is increasingly prominent in national and international regulations because it aligns with a broader drive to sustainably transition the farming system; but it is also a very knowledge-intensive approach, and the evidence to support and accelerate that transition is still lacking. Monitoring is a core principle of all IPM frameworks to identify the connection between interventions, pests, and natural enemies, and continually adapt practice accordingly. A significant barrier is that due to the complexity of rapid and accurate identification, traditional surveillance methods focus on just a small number of insect pests and have a limited capacity to scale up. The quality and quantity of information generated by traditional approaches is therefore insufficient to address dynamic pest behaviours at scale, and non-target insect biodiversity samples are often discarded. As a result, automated identification using Deep Learning (DL) tools to process imagery has the potential to transform how we achieve early detection, run predictive models, and enhance decision-making for pest control. However, the challenge is that developing effective DL tools require large volumes of training data. Previous studies have successfully demonstrated the potential automated imaging for insect identification, but tools that make use of easily accessible data simply reinforce existing biases, so ensuring reference databases are accurate and representative of target agroecosystems is key. The greatest gaps in our understanding are among diverse taxonomic groups of natural predators and parasites/parasitoids that are key to understanding how to sustainably suppress pest populations and minimise crop losses. Such groups underpin, without the need for human intervention, the biological control services we gain form ecosystems. Reducing the taxonomic bottleneck during sample processing would transform the rigour by which we can test different agroecosystems. Our project integrates the strengths of molecular and DL methods to produce a novel generic pipeline for labelling large volumes of images. Molecular identification is the gold standard by which we can consistently identify taxa, and there are increasingly large-scale efforts to catalogue the diversity of life to facilitate confident assignment of DNA sequences to species. Integration of sequencing with techniques like imaging has been constrained by the fact that it remains costly and time consuming to apply at the individual level. Instead, we will employ a phased approach to train algorithms to build associations between thousands of specimens and generated sequence-labels that are then trained to discriminate taxa consistently across samples. Our aim is to develop the first image classifier trained using high-throughput sequencing (HTS), and that has the capacity accurately process large volumes of field data. The techniques are all separately proven methods within their fields, but this is the first time multiple DL tools will have been combined in this way. Our real-world understanding of processing constraints are informed by our industry partner Fera, and our goal is to alleviate the longstanding constraints in their workflow. In future the same approach can be developed in countries where the availability of taxonomic expertise is low and will accelerate new avenues of research that support the global transition to sustainable agricultural management.