Reliability of EO time-series-based assessments of forest biomass dynamics driven by management practices

Understanding how, and to what extent, management practices affect forest biomass dynamics is essential for optimizing management to achieve long-term economic and ecological benefits. However, in-situ forest inventories are spatially and temporally limited due to labor and time costs; thus, post-management forest development and long-term biomass trajectories are typically under-observed or poorly characterized. Earth observation (EO) imagery offers dense, multi-decadal archives with broad spatial coverage, but most studies focus on natural disturbances rather than management interventions because of constraints in spatial coverage and temporal resolution. Consequently, the extent to which the impacts of management measures can be detected and quantitatively assessed using EO time series remains unclear.

This study compared the reliability of EO-based assessments of forest biomass dynamics using conventional optical vegetation indices (VIs) and deep learning–derived canopy height as proxies. VIs such as NDVI and NBR are derived from harmonized Landsat-5/8 and Sentinel-2. Management-specific event-aligned trajectories were used to characterize the interventions following different cutting practices.  The ability of VIs and DL–derived canopy heights to depict biomass dynamics is assessed through alignment with management trajectories.

The study focused on managed boreal forests at the Hyytiälä research site in southern Finland, dominated by Scots pine, using plot-level measurements and management records spanning 1909–2024. EO time series were compiled from Landsat-5/8 (1984–present), Sentinel-2 (10 m optical), and Sentinel-1 (SAR). ALS canopy height data (2019, 2021) were used to evaluate and augment field-measured as calibration. Moreover, meteorological records were included to support interpretation of seasonal variability.

In principle, canopy height is understood as a more reliable predictor for biomass. However, EO-derived VIs showed only moderate correlations with canopy height, and the correlation strength varied across stratification schemes (e.g., stand stage, species, and sensor), due to the saturation and increasing structural heterogeneity in mature stands.

Nevertheless, historical management events since 1985 showed consistent VI patterns, indicating that VIs capture immediate post- management dynamics within 5 years. NBR was most sensitive to abrupt canopy removal, whereas NDVI better reflected gradual recovery. Intensive removals (e.g., clearcutting, shelterwood cutting) produced larger VI responses and longer return times than partial removals (e.g., first thinning, thinning). NBR increased in both broadleaf (Birch) and conifer stands (i.e. Scots pine, Spruce) but recovered more slowly in conifers. NDVI recovery time was similar across species, yet conifer responses were insignificant relative to broadleaf stands. Finally, NBR showed stronger responses and slower recovery in taller stands, whereas NDVI varied little across stand height classes.

U-net DL models produced canopy heights from EO imagery with moderate accuracy (R² = 0.67–0.88; MAE = 1.66–2.98 m), strongly depended on dense harmonized multi-sensor inputs and reliable structural reference data. Ongoing work is evaluating whether the dynamic of such canopy heights aligned with historical management events; detailed results will follow.

Overall, VIs support characterization of managed disturbance and condition-dependent post-management trajectories but are limited as reliable proxies for biomass assessment. DL-based approach offers a potential pathway toward canopy height as proxy for biomass through multi-sensor and high-quality data.