Deep learning segmentation model produces accurate biodiversity predictions at 10 m spatial resolution

Abstract Biodiversity loss is an increasingly intensifying global crisis with significant societal, political, and financial implications. However, our understanding of biodiversity remains limited due to vast data gaps, and the existing data products fail to meet the needs of decision-makers because of their limited accuracy, spatial scale, and resolution. As we approach the short- to medium-term implementation deadlines of major international biodiversity agreements, such as the United Nation's Kunming-Montreal Global Biodiversity Framework and the European biodiversity strategy, there is a pressing need for high-quality data products to guide large-scale conservation prioritization decisions. In this study, we present a novel deep learning segmentation approach for detecting high biodiversity-conservation-value forests using a set of remote sensing data products. This approach has enabled us to produce a national data product of continuous coverage for Sweden, with an unprecedented accuracy of 91% and a high-detail spatial resolution of 10-meter pixel size, significantly advancing on existing approaches. Sweden serves as a good test case for the here developed approach, as it is home to the largest portion of Europe's remaining old-growth forests and is also characterized by robust biodiversity and environmental data availability. Our approach allowed us to map over 50,000 km² of high conservation value forest (HCVF) at high confidence, saving more than 750 years of manual inventory effort and significant financial and personnel resources. With its high accuracy and spatial resolution, our data product offers substantial utility for decision-makers at different administrative scales, and directly addresses the goals set by large international biodiversity conservation plans. While this study focuses on detecting high conservation value forests in Sweden, the presented model serves as a proof-of-concept implementation that can be adapted and applied for modeling other regions and biodiversity metrics.