English Abstract
Low vegetation cover and patchy distribution in arid regions make vegetation spectral signals susceptible to interference from bare-soil backgrounds, while differences in canopy structure among herbaceous vegetation, shrubs, and trees further increase the difficulty of accurate aboveground biomass (AGB) estimation. This study investigated heterogeneous vegetation in the middle and lower plains of the Yarkand River Basin, China, using field-measured AGB, UAV multispectral imagery, and LiDAR point clouds. Spectral (S), canopy height (H), and texture (T) features were derived, and four combinations (S, S + H, S + T, and S + H + T) were evaluated with SVR, GPR, XGBoost, and RF under repeated nested five-fold cross-validation. The results showed that the optimal texture window size varied among models, with the 7 × 7 pixel window providing a reasonable balance between prediction accuracy and stability. Adding canopy height to spectral-only models increased R2 by 0.282–0.455 and reduced RMSE by 28.2–41.9%, exceeding the incremental contribution of texture. S + H + T achieved the best overall performance (R2 = 0.643–0.806; RMSE = 1.189–1.629 kg·m−2). XGBoost yielded the highest mean accuracy, whereas RF was more stable. SHAP analysis attributed 63.6%, 30.2%, and 6.2% of the overall contribution to canopy height, spectral, and texture features, respectively. Overall, multisource feature integration improved AGB estimation, with canopy height providing the main incremental information; however, accurately retrieving canopy height for low-stature vegetation in arid regions remains a key challenge for further improving estimation reliability.