Improving aboveground biomass estimation in desert steppe using hyperspectral semantic segmentation.

Introduction: Accurate estimation of aboveground biomass (AGB) in desert steppes is essential for evaluating ecosystem productivity, optimizing grazing management, and monitoring ecological degradation. However, AGB retrieval in these ecosystems remains challenging because sparse vegetation cover and exposed soil backgrounds often cause severe spectral mixing, thereby reducing the accuracy and robustness of biomass estimation models. Methods: We developed a hyperspectral AGB estimation framework that improves vegetation spectral purity through semantic segmentation and enhances model performance through spectral feature optimization. Using hyperspectral imagery from the desert steppe of Otog Banner, Ordos, Inner Mongolia, China, we first extracted grass vegetation to reduce background interference using U-Net, SegNet, DeepLabV3+, and random forest (RF). The extracted vegetation spectra were then optimized through spectral preprocessing, vegetation index integration, and feature selection, and subsequently subjected to partial least squares regression (PLSR) for AGB estimation. Model performance was evaluated using leave-one-out cross-validation (LOO-CV). Results: The U-Net had the best performance for vegetation extraction, with an overall accuracy (OA) of 0.923, effectively reducing interference from bare soil, shadows, and other non-vegetation backgrounds. The modeling combination (SNV+VI+CARS+PLSR) yielded optimal performance, with a coefficient of determination (R²) of 0.83, a root mean square error (RMSE) of 18.93 g/m², and a ratio of performance to deviation (RPD) of 2.39. Discussion: These findings demonstrate that hyperspectral semantic segmentation can effectively improve AGB estimation in desert steppes by enhancing vegetation spectral purity and reducing background contamination. The proposed framework provides an effective approach to biomass estimation in sparsely vegetated grassland ecosystems.