Estimating Pinus taeda (L.) (Loblolly Pine) timber volume with multi-temporal SAR-optical fusion and supervised learning techniques

Accurate timber volume estimation is critical for the management of Pinus taeda (L.), yet traditional field-based inventories are resource-intensive. This study developed a remote sensing framework to predict plot-level timber volume using 24 months of Sentinel-1 and Sentinel-2 time-series data across 258 field plots (0.04 ha) in the southeastern United States. We evaluated seven machine learning and neural network algorithms. The Gated Recurrent Unit (GRU) model emerged as the most accurate, achieving a root mean squared error (RMSE) of 60.38 m3 ha−1, a mean absolute error (MAE) of 34.28 m3 ha−1, and a coefficient of determination (R2) of 0.49, marginally outperforming the Random Forest (RF) (RMSE = 60.50 m3 ha−1). In contrast, the Gradient Boosting Machine (GBM) model yielded the lowest accuracy (RMSE = 66.72 m3 ha−1). Model performance was highly sensitive to silvicultural conditions; the GRU model reduced RMSE from 67.01 m3 ha−1 in unthinned plots to 48.97 m3 ha−1 in thinned plots. Furthermore, a distinct U-shaped error trend was observed across stand densities: the RF model achieved its lowest error (RMSE = 36.66 m3 ha−1) in medium-density stands (618–1236 trees ha−1 (TPH)), while error rates increased to 88.44 m3 ha−1 in high-density stands (>1237 TPH). The feature importance analysis revealed that while RF relied on Sentinel-2 Short-Wave Infrared (SWIR) bands, the neural networks prioritized Sentinel-1 VH backscatter. Notably, RF and LSTM sustained comparable accuracy (R2 = 0.49) using only the top 10 of these variables. These results demonstrate the viability of open-source satellite imagery for monitoring yield in thinned, medium-density (618–1236 TPH) plantations, while quantifying specific saturation limits in high-density stands.