Abstract: Grassland ecosystems are crucial to the global carbon cycle and ecological balance. Above-ground biomass (AGB) serves as a core indicator for assessing grassland productivity and ecosystem health. Current AGB estimation methods are plagued by reliance on single data sources, redundant feature extraction, and inadequate capture of complex nonlinear relationships, which hinder accurate evaluations of grassland ecological health and responses to climate change.This study focuses on the Gegentala desert grassland and proposes a novel AGB estimation method based on the fusion of GF-1 and Sentinel-2 remote sensing images. GF-1 images, with high spatial resolution (up to 2m for panchromatic bands), excel at capturing fine-scale surface details but have limited spectral resolution. In contrast, Sentinel-2 images maintain spatial consistency at 10m or higher and provide rich spectral information covering blue, green, red, near-infrared, and shortwave-infrared bands, which is critical for monitoring vegetation status and ecosystem changes. Their fusion achieves effective complementarity in spatial resolution and spectral characteristics, significantly improving the quality of vegetation index and texture feature extraction.Data preprocessing involved using SNAP and ArcGIS for geometric correction, cropping, mosaicking, and band resampling to ensure strict spatial alignment between GF-1 and Sentinel-2 images. GF-1 data were radiometrically calibrated using absolute calibration coefficients from the China Centre for Resources Satellite Applications, followed by atmospheric correction via the FLAASH model. Cloud-contaminated pixels were removed through a combination of thresholding and visual interpretation to guarantee consistency and accuracy in subsequent feature extraction. Feature extraction included calculating vegetation indices and texture features: traditional indices such as NDVI and SAVI were derived from GF-1 images, while additional indices from Sentinel-2 images were computed using shortwave-infrared and red-edge bands. Texture features (e.g., mean, variance, homogeneity) were extracted from blue, green, red, and near-infrared bands via the gray-level co-occurrence matrix (GLCM), with optimal parameters determined using window sizes ranging from 3×3 to 11×11. Feature selection employed random forest importance evaluation combined with optimal subset regression to eliminate redundant features, providing high-quality input variables for modeling.Two models were developed: a random forest (RF) regression model and a multi-scale convolutional neural network (MCNN) model. The RF model selected variables based on feature importance scores, with parameters such as the number of trees and maximum depth optimized via Bayesian optimization, achieving a coefficient of determination (R²) of 0.77 and a root mean square error (RMSE) of 29.59 g/m². The MCNN model, utilizing 3×3 to 9×9 multi-scale convolution kernels and a multi-head attention mechanism, captured multi-scale and complex nonlinear features. Integrated with residual connections and global average pooling, it enhanced feature fusion efficiency and model robustness, achieving superior performance (R²=0.81, RMSE=28.49 g/m²), particularly in high-AGB regions, demonstrating its strong capability to capture complex nonlinear relationships and multi-scale features.Application of the MCNN model to the Gegentala grassland generated an AGB spatial distribution map, with an overall mean of 51.54 g/m² and a standard deviation of 23.64 g/m², indicating mild desertification in the region. Spatial analysis revealed lower biomass in the northwest and southwest, and higher biomass in the north and east, reflecting significant impacts of topography, land use, and human activities on grassland degradation. The complementary advantages of GF-1 and Sentinel-2 images improved the extraction of vegetation indices and texture features, enhancing the sensitivity of regression models to grassland conditions. This study provides an accurate method for AGB inversion in desert grasslands, offering reliable data support for ecological health assessment, desertification monitoring, and the formulation of ecological protection strategies.