A study on the remote sensing estimation and spatiotemporal distribution patterns of aboveground biomass in savanna grasslands of the Yuanmou dry-hot valley

Savanna grasslands in dry-hot valleys play crucial ecological and productive roles, yet accurate estimation of their aboveground biomass (AGB) remains challenging due to rugged terrain, climatic variability, and intensive human disturbances. To address this, we investigated the Yuanmou dry-hot valley in Yunnan, China, by constructing a multi-source dataset that integrated remote sensing indices, topographic factors, meteorological variables, and biodiversity metrics. Three feature selection techniques were applied to identify key predictors, and the selected variables were used to train ensemble machine learning models. Of all tested model combinations, the Random Forest model with Recursive Feature Elimination achieved the highest predictive accuracy, with a coefficient of determination of 0.6975, a root mean square error of 89.3436 g/m², and a mean relative error of 0.7282. The most influential predictors included temperature, latitude, biodiversity indices, and specific spectral bands and vegetation indices. From 2019 to 2024, AGB in the study area exhibited an overall increasing trend and showed increased spatial homogeneity, although low-altitude areas consistently maintained lower biomass due to stronger grazing and land-use pressures. These findings demonstrate the advantages of integrating multi-source variables with machine learning in ecologically heterogeneous regions. The proposed framework effectively reduced redundancy, enhanced sensitivity to ecological drivers, and showed strong adaptability to complex environments. The observed biomass dynamics further highlight the positive effects of ecological restoration policies, while revealing persistent trade-offs between conservation and land use in lowland zones. Overall, this study provides a practical methodological framework for improving the accuracy and applicability of AGB estimation in savanna ecosystems, offering valuable insights for ecological monitoring, policy implementation, and sustainable grassland management.