关键词: 无人机, 多光谱植被指数, 生物量, 机器学习, 精准化管理草甸草原

Abstract: Aboveground biomass (AGB) is a crucial indicator for measuring ecosystem health and ecological services of grasslands. Accurately assessing the impact of different grazing intensities on AGB is central to the precision management of grasslands. However, there are several limitations in traditional monitoring methods. First, field measurements are often inefficient and spatially unrepresentative. Second, the resolution of satellite remote sensing is insufficient to capture the fine-scale fragmented vegetation dynamics induced by grazing. Combining unmanned aerial vehicle (UAV) multispectral remote sensing data in conjunction with five machine learning models, we developed a high-precision estimation model capable of accurately quantifying AGB and its spatial heterogeneity in meadow steppes under different grazing management regimes. The results showed that: (1) The selected multispectral bands and vegetation indices exhibited significant correlation with field-measured biomass. (2) Among the five machine learning models employed, the gradient boosting decision tree (GBDT) model achieved the highest accuracy on both the training and validation sets, with R² values of 0.98 and 0.92, respectively, followed by the random forest (RF) model (R² values of 0.96 and 0.90). (3) The model predictions clearly demonstrated an AGB gradient corresponding to grazing management status. Mean values of AGB significantly increased from heavy grazing grasslands to light grazing grasslands and subsequently to fencing grasslands at both plot and site scales (from 98.70 g·m^-2 to 413.75 g·m^-2 at the site scale). Similarly, spatial heterogeneity of AGB, as indicated by the standard deviation, markedly rose with increasing grazing intensity (from 55.51 g·m^-2 to 154.00 g·m^-2). In summary, we conducted high-precision inversion of bands or vegetation indices and biomass of meadow steppe at a fine scale using UAV-based machine learning methods. By innovatively combining ground measurements with UAV remote sensing, this study provided crucial technical support for the precision quantification of management measures, which serves as a valuable reference for accurate, multi-scale (ground-air-space) biomass estimation.

Key words: unmanned aerial vehicle (UAV), multispectral vegetation index, biomass, machine learning, precision management of meadow steppe