Zsofia Koma
Research Scientist
Abstract
Spatially explicit information on forest resources and structure is essential for sustainable forest management and evidence-based policy-making. In the Nordic region, large-scale forest mapping often relies on integrating National Forest Inventory (NFI) field plots with airborne laser scanning (ALS) data. However, infrequent nationwide ALS campaign coverage limits their use for continuous monitoring. Satellite imagery, with its high temporal and spatial resolution, provides a promising alternative. We evaluate UNet-based deep learning models trained on wall-to-wall ALS-derived forest resource maps for predicting volume and Lorey’s height in Norway using optical (Sentinel-2) and SAR (Sentinel-1, PALSAR-2) data. The UNet models, trained on both Finnish and Norwegian ALS maps, are benchmarked against extreme gradient boosting (XGB) models. Transfer learning is further explored by finetuning models using Norwegian NFI plots. Model accuracies are assessed using 541 reserved test NFI plots and 44 independent forest stands, representing high‑volume mature boreal forests (>200 m3 ha−1). The UNet model trained on Norwegian ALS‑based data achieved R2 values of 0.59 for both volume and Lorey’s height when evaluated on NFI plots, and 0.70 and 0.59 for forest stands, respectively, outperforming the XGB models. Finetuning improved model transferability, yielding gains of up to 0.13 in R2 for volume and 0.46 for Lorey’s height when adapting the Finnish model to Norwegian conditions. Utilizing SAR data alongside optical data enhanced model accuracy. Overall, our findings demonstrate the potential of UNet models trained on wall-to-wall ALS maps for forest resource mapping across Nordic countries.
Abstract
Fine-scale, spatially explicit forest attribute maps are essential for guiding forest management and policy decisions. Such maps, based on the combination of National Forest Inventory (NFI) and remote sensing datasets, have a long tradition in the Nordic countries. Harmonizing the pixel size among national forest attribute maps would considerably improve the utility of the maps for users. However, the maps are often aligned with the NFI plot size, and the influence of creating these maps at different spatial resolutions (i.e. pixel sizes) is little studied. We assess the stand-level uncertainty (RMSE) of biomass, volume, basal area, and Lorey’s height estimates resulting from the aggregation of maps across varying spatial resolutions. Models fit at 16 m native resolution using more than 14 000 NFI plots were applied for predictions at pixels sizes (side lengths) of 1, 5, 10, 16, and 30 m. For independent validation, we used more than 600 field plots – that cover a total area of 24 ha and were clustered within 65 stands across Norway. For all attributes, the lowest RMSEs, ranging from 6.86% for Lorey’s height to 13.86% for volume, were observed for predictions at pixel sizes of 5 m to 16 m. The RMSE changes across resolutions were generally small (< 5%) for biomass, volume, and basal area. For Lorey’s height, changing the spatial resolution resulted in large RMSEs of up to 25%. Overall, our findings suggest that the main forest attributes can be mapped at a finer resolutions without complex adjustments.
Abstract
No abstract has been registered
