Measuring the impact of rewilding on pollinator biodiversity: what can machine learning tell us?

点击翻译按钮获取中文摘要

Rewilding is an approach to conservation which seeks to regenerate degraded ecosystems in a self-sustaining way, with relatively little ongoing management. The approach commonly involves the planned retreat from intensive agriculture, and the reintroduction of large herbivores and other keystone species. Rewilding projects are proliferating in the UK, especially following the high-profile success of the Knepp Wildland project in West Sussex, and are potentially transformative in attempts to protect and enhance biodiversity nationally. There is an urgent need to monitor and evaluate the success of rewilding projects, to facilitate evidence-based design and management of sites for maximal conservation value. Central to evaluating the impact of conservation projects is accurate and cost-effective surveying of biodiversity. However, the challenges of species identification often lead to patchy or ineffective monitoring of all but the most easily identifiable taxa. Recent technological developments, especially using machine learning (ML) in species recognition, have the potential to reduce dramatically the costs of biodiversity monitoring. Thus far, ML image (and sound) recognition has shown great promise in citizen science projects, where publicly contributed records can be mined by ecologists interested in spatial and temporal biodiversity patterns. But these tools also have huge potential to increase the efficiency of directed field surveys by professional ecologists. This project will evaluate ML species recognition tools for monitoring biodiversity in the context of a major new rewilding project (Boothby Wildland) being implemented by Nattergal Ltd, founded by the creators of the Knepp Wildland. We will survey key pollinator taxa over the first three years of the Boothby project, assessing the impact of the retreat from arable farming on spatial and temporal patterns in biodiversity. Simultaneously, we will validate an ensemble of species recognition apps with conventional expert-led insect identification, enabling us to assess the long-term feasibility of rapid, low-cost monitoring of biodiversity.