Region-wide Prediction of Boreal Understorey Vegetation using Spaceborne Remote Sensing

Understorey vegetation (USV) plays a vital role in forest ecosystems by influencing biodiversity, nutrient cycling, and disturbance dynamics. Accurate mapping of USV is essential for understanding ecosystem functioning and its relationship with environmental variables across landscapes. The use of remote sensing (RS) for its application to USV prediction has been attempted a handful of times. Open-access Sentinel-1 C-band Synthetic Aperture Radar (SAR) and Sentinel-2 multispectral imagery have been used in this study for USV cover modelling, where USV cover is defined as the surface area (in m²) covered by the USV on the forest floor. Two sources of field reference data used here for USV cover measurements were from: (A) the Swedish National Forest Inventory (NFI) (for year 2019-2024) covering 55,000 km² area in the Västerbotten County, northern Sweden, and (B) detailed field measurements from a smaller test site – the Krycklan catchment research area covering 70 km² from 2024. This study aimed to 1) Develop two separate regression-based generalized additive models (GAM), Model A and Model B using an area-based approach integrating Senitnel-1 and 2 metrics and trained using the two field reference datasets A and B, respectively, and 2) Further extend Model A to account for interaction of USV with additional environmental covariate rasters of, e.g., soil moisture, elevation, land-use/land-cover classes, bedrock type, soil type, and bioclimatic metrics such as seasonal and annual temperature and precipitation, acquired over the entire Västerbotten county. 

Both baseline models A and B were developed using explanatory variables from Sentinel-1 namely, difference between the vertical transmit, vertical receive (VV) backscatter and vertical transmit, horizontal receive (VH) backscatter and total backscatter power (∣VV∣²+∣VH∣²), and from Sentinel-2 namely, Normalized difference vegetation index (NDVI), Visible atmospherically resistant index (VARI), and the difference between surface reflectance of the red-edge band in summer and autumn seasons. Both baseline models A and B - demonstrated comparable performance with similar magnitude of root mean square error (RMSE) and coefficient of determination (R²) values when validated against a common test subset derived from the NFI field reference data. With including the environmental covariates in model A, the USV cover showed correlation with soil moisture, elevation, land-use/land-cover classes, and seasonal and annual temperature and precipitation variables. The inclusion of these variables improved the extended model A performance compared to the baseline model A, with 15% increase of R² and 8% decrease of RMSE values. These results highlight the importance of integrating climate and topographic covariates along with RS data for improved USV prediction and mapping. 

This study demonstrates the feasibility of large-scale USV prediction using open access Sentinel-1 and 2 data combined with field reference data, and environmental covariates. While SAR and multispectral data provide valuable information, incorporating biophysical and climatic variables substantially enhances model performance. This approach offers a cost-effective and scalable workflow for monitoring USV in boreal forests, benefiting sustainable forest management and biodiversity studies.