Spatially Constrained Grassland Aboveground Biomass Estimation by Identifying and Masking Achnatherum splendens: Integrating UAV Remote Sensing and Deep Learning

Accurate estimation of forage aboveground biomass (AGB) from unmanned aerial vehicle (UAV) imagery is essential for monitoring alpine grassland productivity and supporting sustainable grassland management. However, the biomass of non-palatable species may be included in remote sensing-based estimates when their spatial distribution is not explicitly considered, potentially affecting spatial assessments of forage biomass. This study developed a UAV RGB-based framework for spatially constrained grassland AGB assessment by integrating the extraction of the growing-season non-palatable species Achnatherum splendens, mask-based spatial exclusion, feature optimization, and AGB inversion using field measurements from the northwestern shore of Qinghai Lake. Among tested semantic segmentation models, the Attention U-Net achieved the highest segmentation accuracy and was selected to identify A. splendens, while vegetation indices (VIs) and gray-level co-occurrence matrix (GLCM) texture features were optimized using the minimum redundancy maximum relevance (mRMR) algorithm. Random forest regression (RFR), support vector regression (SVR), and partial least squares regression (PLSR) models were subsequently evaluated. The Attention U-Net achieved high segmentation accuracy (PA = 89.98%, mIoU = 88.89%, and F1 = 93.88%). Feature optimization improved the performance of all regression models, with SVR providing the highest estimation accuracy (R2 = 0.72; RMSE = 23.73 g m−2). The estimated AGB ranged from 67.88 to 205.85 g m−2, revealing pronounced spatial heterogeneity. In a typical sample area with high-density A. splendens, excluding pixels classified as A. splendens reduced the estimated AGB by 15.45%, demonstrating the influence of dense A. splendens patches on spatial AGB assessment. These results demonstrate that integrating deep learning-based species masking with feature optimization provides a spatially explicit approach for grassland AGB assessment after excluding areas classified as A. splendens and provides useful spatial information for fine-scale grassland monitoring and management.