Deep learning and model transferability for standing dead tree mapping

Timely forest health monitoring depends on fast and accurate methods that identify tree mortality due to climate change driven phenomena such as bark beetle attacks. Aerial imagery coupled with deep learning is an efficient tool for detecting standing dead trees compared to field work but requires reliable and quickly accessible applications for forest owners and decision makers. Current challenges are related to laborious training data acquisition that models require in order to generalize for large areas. Few studies address how locally trained dead tree models can be transferred to new sites with minimal manual delineation of calibration data.

This presentation introduces three binary CNN segmentation models for detecting standing dead trees during the Finnish leaf-on season, with training and testing applied on aerial images of Koli 2017, Koli 2022 and Lapinjärvi 2022 study sites. These models were trained using 300–543 tiled aerial image samples and then transferred to images of Koli and Lapinjärvi taken in 2025 where only a small calibration set of n=12 samples are manually delineated for both images. To expand this calibration data, various geometric augmentations are applied to the samples. This dataset allows for transferability tests between eastern and southern Finland as well as across 8 years of aerial image data with varying imaging conditions.

Pixel-wise F1 scoring of all models ranged from 0.69 to 0.82 while the calibration improved transferred model F1 scores by 13–123% depending on site and year. This presentation will also provide a clear explanation of the used models, as well as the used aerial images with their inherent characteristics, for example spectral variations that affect calibration efficiency. Furthermore, standing dead tree mortality maps are shown to visualize the tree mortality extent in Koli and Lapinjärvi study areas.

Augmentation can efficiently generalize standing dead tree detection models as well as enable effortless calibration to new sites. Therefore, this approach can be extended to other tasks as well, such as forest fire mapping.