Estimating Forest Volume and Forest Volume Change Using Random Forests and Sentinel-2 Data

Accurate and precise estimation of forest volume and its changes are fundamental for sustainable forest management, resource assessment, and long-term inventory monitoring. Multi-temporal remotely sensed data provide spatial auxiliary information that can substantially increase the precision of estimates of volume and volume change obtained from probability-based field samples. However, methodological differences between indirect and direct change estimation approaches may lead to differences in variance estimation and uncertainty propagation. This study, therefore, compares forest volume and forest volume change estimation under design-based inference by integrating probability-sampled National Forest Inventory (NFI) data with multi-temporal Sentinel-2 auxiliary variables at 10 m and 20 m spatial resolutions, using the random forest (RF) prediction algorithm and both indirect and direct estimation approaches. Forest volume means are estimated for two inventory years, 2013 and 2023. For the indirect approach, simple expansion and model-assisted regression estimators are formulated separately for each year, and change is estimated as the difference between time-specific estimates. For the direct approach, plot-level change is used as the response variable, and model-assisted regression estimators are constructed using changes in auxiliary variables, with residual-based variance estimators. Random forest models using Sentinel-2 auxiliary data explained approximately 51–62% of the variation in forest volume, with RMSE values ranging from about 79 to 97 m3/ha. The 10 m resolution data produced slightly more precise predictions than the 20 m data, though the gain was small relative to the greater processing effort required. Model-assisted estimators using both 10 m and 20 m Sentinel-2 data produced substantially smaller standard errors than simple expansion estimators, with relative efficiency analysis indicating an approximately fivefold gain in efficiency. The direct and indirect model-assisted approaches produced similar estimates of volume change, although the direct approach resulted in smaller standard errors. Overall, remotely sensed auxiliary data primarily improved the precision of forest volume change estimates but not the magnitude of the estimated change.