Abstract: Grassland ecosystems can greatly contribute to the global carbon cycle and ecological balance. Among them, the above-ground biomass (AGB) can serve as one of the most important indicators to assess the grassland productivity and ecosystem health. Current AGB estimation can be hindered by the single data sources, redundant feature extraction, and inadequate capture of complex nonlinear relationships. It is often required for the accurate evaluation of the grassland ecological health response to climate change. In this study, a rapid and accurate AGB estimation was proposed to fuse the GF-1 and Sentinel-2 remote sensing images. The Gegentala desert grassland was taken as the research area. GF-1 images were captured with the high spatial resolution (up to 2m for panchromatic bands). The fine-scale surface details were obtained with the limited spectral resolution. In contrast, Sentinel-2 images maintained the spatial consistency at 10m or higher. The rich spectral information was then provided for the blue, green, red, near-infrared, and shortwave-infrared bands, in order to monitor the vegetation status and ecosystem. Both GF-1 and Sentinel-2 images were fused to balance the spatial resolution and spectral characteristics. The quality of the vegetation index was significantly improved to extract the texture feature. Data preprocessing involved the geometric correction, cropping, mosaicking, and band resampling using SNAP and ArcGIS, particularly for the spatial alignment between GF-1 and Sentinel-2 images. The GF-1 data were radiometrically calibrated using absolute calibration coefficients from the China Centre for Resources Satellite Applications. Atmospheric correction was followed via the FLAASH model. Cloud-contaminated pixels were then removed to combine the thresholding and visual interpretation. As such, high consistency and accuracy were achieved for the subsequent feature extraction. The vegetation indices were also calculated to extract the texture features. Specifically, the conventional indices, such as NDVI and SAVI, were derived from the GF-1 images, while the additional indices from Sentinel-2 images were derived using shortwave-infrared and red-edge bands. Texture features (e.g., mean, variance, and homogeneity) were extracted from blue, green, red, and near-infrared bands via the gray-level co-occurrence matrix (GLCM). The optimal parameters were determined using window sizes from 3×3 to 11×11. Random forest (RF) importance evaluation was combined with optimal subset regression in order to remove the redundant features. The high-quality input variables were provided for modeling. An RF regression and a multi-scale convolutional neural network (MCNN) model were developed after feature selection. The variables of the RF model were selected according to the feature importance scores. Some parameters, such as the number of trees and the maximum depth, were optimized via Bayesian optimization. A better performance was achieved in the coefficient of determination (R²) of 0.77 and a root mean square error (RMSE) of 29.59 g/m². The multi-scale convolutional neural network (MCNN) model was utilized with 3×3 to 9×9 multi-scale convolution kernels. A multi-head attention mechanism was used to capture the multi-scale and complex nonlinear features. Residual connections and global average pooling were integrated to enhance the feature fusion efficiency and robustness. Superior performance (R²=0.81, RMSE=28.49 g/m²) was achieved to capture the complex nonlinear relationships and multi-scale features, particularly in the high-AGB regions. The MCNN model was applied to the Gegentala grassland, Inner Mongolia, China. An AGB spatial distribution map was generated with an overall mean of 51.54 g/m² and a standard deviation of 23.64 g/m², indicating the mild desertification in the region. Spatial analysis revealed that the lower biomass was found in the northwest and southwest, whereas the higher biomass was found in the north and east. There were significant impacts of the topography, land use, and human activities on the grassland degradation. The GF-1 and Sentinel-2 images were utilized to extract the vegetation indices and texture features. The sensitivity of regression models was enhanced for grassland conditions. This finding can provide an accurate AGB inversion in the desert grasslands. The reliable data support can also be offered for the ecological health during desertification monitoring, particularly for the decision-making on the ecological protection.