Very-high resolution aerial imagery and deep learning uncover the fine-scale spatial patterns of elevational treelines

Abstract. Treelines are sensitive indicators of global change, as their position, composition and pattern directly respond to numerous ecological and anthropogenic factors. Most studies are case-specific and treeline features vary greatly worldwide making it very difficult to model an overall pattern. Therefore, the further development of methods to accurately map fine-scale treeline spatial patterns, especially through innovative approaches such as remote sensing with unmanned aerial vehicles (UAV) and deep learning models, is of scientific importance for the conservation of forest ecosystems in the face of ongoing and future ecological challenges.

In this study, we aimed to fill this gap by combining field and UAV-based data with a deep learning model to retrieve single tree-scale information over 90 ha distributed on 10 study sites in the Italian Alps. Using the proposed methodology, we were able to correctly detect individual tree crowns of conifers taller than 50 cm with a detection rate of 70 % and an F1 score of 0.76. The detection rates of individual tree crowns improved with increasing tree height, reaching a peak value of 86 % when only tall trees (>2 metres) were considered. Canopy delineation was good when all trees were considered (Intersection over Union (IoU) = 0.76) and excellent when only tall trees were considered (IoU = 0.85). The estimates of tree position and height achieved an RMSE of 59 cm and 92 cm, respectively. Our univariate and bivariate heterogeneous Poisson Point Pattern Analysis (PPA) revealed a clustered pattern for spatial scales < 20 m, and a strong repulsion between small and tall trees at all the tested spatial scales, respectively. PPA results suggest that in the Alps, seedlings tend to progressively occupy safe sites and colonise non-competitive sites, resulting in the evenly sized clusters found. We demonstrated that the proposed methodology effectively detects, delineates, georeferences and, measures tree height of most trees across diverse Alpine treeline ecotones. This enables the analysis of fine-scale spatial patterns and underlying ecological processes. The inclusion of heterogeneous study areas facilitates the transferability of the segmentation model to other mountain regions and makes the present study a benchmark for creating a global network of fine-scale mapped treeline spatial patterns to monitor the effects of global change on ecotone dynamics.