Mapping large European aspens (Populus tremula L.) using national aerial imagery and a U-Net convolutional neural network

Biodiversity supports boreal forest stability, with European aspen ( Populus tremula L.) enhancing it by providing crucial habitats and resources for dependent species. Despite existing aspen remote sensing methods, there is no efficient method for mapping it over large areas to support forest management and conservation. In this study, we developed a deep learning-based semantic segmentation model to segment aspen from openly available national aerial imagery, evaluated its accuracy using field data, and compared the segmentation performance between leaf-on and leaf-off conditions. The study was conducted in three areas in southern Finland: Helsinki, Lohja, and Evo. We employed a U-Net encoder-decoder architecture using four-band aerial imagery (RGB and NIR) with a spatial resolution of 0.5 meters, captured during both seasonal conditions. Training data consisted of visually identified aspen locations from imagery between 2010 and 2023, covering 290 sites across Finland. We found notable differences in segmentation accuracy between leaf-on and leaf-off conditions and aspen size. The F1-score was higher in leaf-off (0.573) than leaf-on (0.494), with intersection over union (IoU) values of 0.280 and 0.256 for leaf-off and leaf-on, respectively. The diameter at breast height (DBH) of the segmented aspens was similar for both conditions. Moreover, segmentation accuracy improved for larger aspens, with F1-scores reaching 0.663 (leaf-off) and 0.551 (leaf-on) for aspens >20 cm DBH, and 0.710 (leaf-off) and 0.594 (leaf-on) for those >30 cm. The developed model reasonably locates aspen distribution and abundance, assisting forest managers to make informed management decisions. • •Aspens were segmented using both leaf-on and leaf-off aerial imagery. • •Segmentation accuracy was higher under leaf-off conditions compared to leaf-on. • •The model segmented larger aspens with higher probability than smaller ones. • •The mean DBH of segmented aspens was similar under both leaf-on and leaf-off conditions. • •The model mapped aspen distribution and abundance with reasonable accuracy.