Local data matters: Improving biodiversity risk and impact assessment through a data quality focus.

Widespread degradation of nature has increased pressure on corporations and financial institutions to assess and mitigate their biodiversity impact, however, collecting relevant local data can be costly. The increasing availability of biodiversity data and Earth observation (EO) data provides a route to cost effective impact assessment via extrapolation using existing data and statistical modeling. Through a review of the datasets and tools currently used by corporations and financial institutions we show that extrapolation to local sites from global datasets, or using only proxies, is the dominant approach. We test the reliability of such assessments by combining high resolution earth observation time series data with extensive biodiversity data from recent environmental DNA (eDNA) surveys of two countries with widely varying conditions, Sweden and Madagascar. We use machine learning in combination with high-quality biodiversity data to predict five essential biodiversity variables (EBVs) for local sites, using cross-validation to test prediction accuracy. The results show that reasonably accurate EBV predictions can be obtained for sites with some local data, but performance declines considerably when modelling summary measures at new sites. Moreover, the quality of predictions, both within sites and at new sites, is dependent on the EBV and local context. To address the concerns over the reliability of model-based EBV assessments, we propose a biodiversity data hierarchy framework, which can be used by organisations to track stepwise improvements in the data sources underpinning their biodiversity impact assessments.