Abstract Accurate estimation of aboveground biomass (AGB) in alpine ecosystems is challenging because of strong phenological variability, heterogeneous canopy structure and complex spectral–biomass relationships. Using 2 years (2023–2024) of monthly ground‐based hyperspectral observations collected during the growing season (May–September), this study examined alpine meadow, degraded alpine meadow and alpine shrub communities on the northeastern Qinghai–Tibetan Plateau. Four regression approaches—elastic net, random forest (RF), support vector regression with a radial basis function kernel and extreme gradient boosting (XGBoost)—were compared with an emphasis on temporal transferability. A two‐stage feature selection framework combining correlation screening and LASSO regression was applied to smoothed, first‐derivative and continuum‐removed spectra, reducing more than 1000 spectral predictors to approximately 20 physiologically meaningful features. Tree‐based ensemble models, particularly RF and XGBoost, consistently achieved the highest predictive performance and showed strong robustness across vegetation types. Model accuracy exhibited clear seasonal dependence, with lower performance early in the growing season and marked improvement during mid and late season periods. The most informative predictors were concentrated in the red‐edge, near‐infrared and shortwave‐infrared regions, and models based on these optimized features matched or exceeded full‐spectrum performance. The results demonstrate that combining targeted spectral features with ensemble learning provides a robust framework for seasonal AGB estimation in heterogeneous alpine grasslands.
Qian et al. (2026) studied this question.