Abstract: Accurately monitoring the aboveground biomass of semi desertified grasslands is a necessary condition for evaluating the ecological status of grasslands and corresponding sustainable management of grasslands. Damao banner has abundant grassland resources and a simple vegetation community structure, which is a typical representative of semi desertification grassland. In order to improve the quality of information on semi desertification grassland resources, taking Damao banner as the research area, based on Landsat remote sensing images, 23 original features are constructed using ground measured sample data, combined with spectral, vegetation index, meteorological data, and digital terrain data. Random forest (RF), support vector machine (SVM), gradient boosting regression tree (GBRT) and decision tree(CART) regression algorithms are used for grassland aboveground biomass inversion, and feature importance score and recursive feature elimination (RFE) are used for feature optimization. Finally, the 2021 semi desertification grassland AGB inversion mapping in Damao banner was completed. The results show that the RF model had the highest accuracy in the AGB inversion of semi desertified grasslands. After recursive feature elimination, the optimal number of features is selected to 12, among which meteorological and topographic features contributed the most to the AGB inversion of grasslands. The final accuracy determination coefficient (R²) of the inversion model is 0.83, and the root mean square error (RMSE) is 20.31. This study estimates the biomass of semi desertified grasslands, providing a scientific basis for the management and protection of vulnerable grassland ecosystems and an effective methodology for biomass inversion research.

Key words: semi desertified grassland, aboveground biomass of grassland, Landsat imagery, machine learning, recursive feature elimination method