Aboveground biomass estimation and multiscale spatial pattern analysis in desert rangelands using UAV hyperspectral imagery.

Accurate estimation of aboveground biomass (AGB) is essential for evaluating the productivity of desert rangelands and optimizing grazing management. However, existing remote sensing models based on unmanned aerial vehicle (UAV) data rely predominantly on spectral information and provide insufficient characterization of differences in plant community spatial distribution and their responses across spatial scales. To address these limitations, this study integrated UAV hyperspectral imagery with field quadrat data collected under four grazing intensities. Spectral, textural, and local statistical dispersion features were combined to develop and compare six machine learning models for AGB estimation. The responses of biomass and its spatial patterns to grazing disturbance across spatial scales were also investigated. The results showed that integrating multisource features effectively improved the accuracy of AGB estimation. In particular, textural and local statistical dispersion features provided information complementary to that contained in spectral features. The random forest (RF) model showed relatively high estimation accuracy (R2=0.914). As grazing intensity increased, AGB generally declined, while its spatial distribution shifted from a relatively homogeneous pattern to increasingly heterogeneous and patchy patterns. From moderate to heavy grazing, changes in mean biomass became less pronounced, whereas spatial variability metrics continued to differ, indicating that these metrics can provide information complementary to mean biomass. As the spatial aggregation scale increased, local variability in the estimated AGB maps was progressively smoothed, weakening their ability to represent spatial differences among grazing treatments. This study provides a technical reference for AGB estimation in desert rangelands and for characterizing spatial patterns under grazing disturbance.