Inversion of Forest Volume by Combining Sentinel-1 and Sentinel-2 Data

In order to clarify the influence of remote sensing data sources and machine learning models on the estimation of forest volume, and improve the estimation accuracy of regional forest volume. Based on the field survey data of 38 Larix gmelinii plots and 43 Pinus tabuliformis plots in Wangyedian forest farm, Inner Mongolia, the remote sensing characteristic information of Senitnel-1 and Sentinel-2 images such as spectrum and polarization were extracted. According to different feature combinations, four volume inversion models of Support Vector Regression (SVR), k-Nearest Neighbor (kNN), Multi-Layer Neural Network (MLP) and Multiple Linear Regression (MLR) models were established, and the assessment and comparison of these models were verified and compared. The results showed that: (1) Compared with a single data source, combining Sentinel-1 and Sentinel-2 data were benefit to improve the accuracy of forest volume estimation (an increase of 0.08 in R² and 10.28 m³·hm⁻² in RMSE for Larix gmelinii, and 0.05 in R² and 4.51 m³·hm⁻² in RMSE for Pinus tabuliformis); (2) Compared with MLP and MLR models, the SVR and kNN models had better performance on the estimation of forest volume. Among them, the SVR model achieved the highest accuracy in the estimation of Larix gmelinii volume (R²=0.84, RMSE=44.58 m³·hm⁻²), and the kNN model obtained the highest accuracy in the estimation of Pinus tabuliformis volume (R²=0.74, RMSE=41.41 m³·hm⁻²). The machine learning method combining Sentinel-1 and Sentinel-2 multi-source data can achieve a high precision of volume inversion, which can provide theoretical support and feasible solutions for remote sensing inversion of forest volume at regional scale.