Ada-Xgboost model-based fractional vegetation cover inversion from synthetic aperture radar images

Optical remote sensing satellites cannot provide valid fractional vegetation cover (FVC) data under cloudy, rainy, and foggy weather conditions. To address this, this study proposed a method for FVC inversion based on synthetic aperture radar (SAR) images. First, this study extracted a total of 25 SAR features from Sentinel-1 single look complex (SLC) images, including polarization entropy, anisotropy, alpha angle, covariance matrix elements, and radar vegetation index. Among them, key features sensitive to FVC were selected using Pearson correlation analysis and then utilized for FVC inversion in the study area. Moreover, to optimize overall image quality, this study excluded non-vegetated areas using the European Space Agency (ESA) WorldCover 10m land cover dataset. Finally, it conducted FVC inversion in Kaifeng City by constructing an Ada-Xgboost model using the voting method. The results indicate that the Ada-Xgboost model exhibited significant advantages in capturing the spatial distribution patterns of FVC and improving inversion accuracy. These advantages are evinced by a coefficient of determination (R²) of 0.781 0, a root mean square error (RMSE) of 0.179 3, a mean squared error (MSE) of 0.032 1, a mean absolute error (MAE) of 0.126 4, and a feature number (FN) of seven. Compared to individual models such as AdaBoost and XGBoost, the Ada-XGBoost model effectively reduced the redundancy of SAR features and achieved higher spatial consistency with the reference FVC data. This study further demonstrates that SAR data can effectively compensate for the limitations of optical remote sensing, offering robust data support for large-scale, time-series dynamic FVC monitoring.