Assessing Annual Change in Forest Canopy Mortality in Southern Finland, Using Very High-Resolution Satellite Imagery and Deep Learning

Abstract

Globally, tree mortality has increased in recent decades due to factors such as severe droughts, extreme heat, and pest outbreaks. However, the limited spatial coverage of existing observations hampers a comprehensive understanding of mortality patterns and the development of predictive models. This article employed very high-resolution (0.5 m) satellite imagery and a deep-learning (DL) model to detect forest canopy mortality (FCM) and monitor annual changes. Conducted across seven study areas in Southeastern Finland, each approximately 2500 ha in size, the research utilized a DL model based on the U-Net architecture, trained on 21 144 manually labeled segments of standing dead trees, primarily caused by spruce bark beetle damage in Norway spruce. Validation was done using the 0.5-m satellite image for 2023 data and a drone image (2.5 cm) for 2024. Forest inventory data were used to analyze the influence of tree species on detection accuracy, while a canopy height model was used to assess the impact of tree height. DL model accuracy, F1 score, ranged from 0.74 to 0.93, averaging higher (0.84–0.89) when the dominant single-species proportion was >70%, and lower (0.37–0.81) with greater species diversity. Tree height ranged from 15 to 24 m for correctly detected trees. We observed an increase in FCM from 15 (2022) to 32 ha (2023) in the study area. The methodology in this artile advances large-area tree mortality mapping and underscores the urgent need for accurate monitoring and model development to support proactive forestry risk management.