A remote sensing-based tree mortality indicator for biodiversity monitoring using multispectral UAV and lidar in the Arctic region, Finland

Abstract

Deadwood is a recognised ecological indicator for forest biodiversity. Both the UN Convention on Biological Diversity and the EU Biodiversity Strategy 2030 mandates inventorying and reporting deadwood. However, monitoring tree mortality, the source of deadwood, remains challenging, especially in Arctic and sub-Arctic regions. Boreal forests in this region face increasing climate-driven stressors: droughts, pest outbreaks, and snow damage. These stressors accelerate tree mortality and threaten biodiversity. Deep learning-based semantic segmentation offers a scalable solution for monitoring standing tree mortality using remote sensing data such as multispectral UAV imagery and lidar-derived canopy height models. In this study, we (1) evaluated a U-Net model for detecting tree mortality across 208 ha of Picea abies, Pinus sylvestris, and deciduous dominated forests, and (2) produced a wall-to-wall mortality map to identify spatial patterns and potential drivers in Northern Lapland, Finland. The training dataset included 411 tiles containing 4380 manually delineated tree crowns. The model achieved an overall accuracy of 0.92, with a recall and F1-score for dead trees 0.82 and 0.57, for alive trees 0.93 and 0.96. The performance is sufficient for identifying spatial patterns despite false positive detections. Mortality hotspots were concentrated in Picea abies dominated areas at higher elevations where older trees were located, suggesting a link to snow damage besides natural mortality. These results demonstrate the potential of UAV based deep learning workflows for large-scale exploratory monitoring of tree mortality patterns in northern boreal forests in Arctic and sub-Arctic region and supports biodiversity reporting and restoration monitoring through spatial forest-health assessment.