Abstract: As an important carbon sink, the accurate estimation of aboveground biomass in bamboo forests is highly important for carbon cycle assessment, ecosystem carbon stock accounting, and the implementation of regional carbon neutrality goals. To resolve the problems of high cost in traditional field surveys and insufficient accuracy of single remote sensing models, we proposed a methodological framework for estimating aboveground biomass in bamboo forests based on Sentinel-2 remote sensing data. Firstly, by analyzing the time-series spectral characteristics of bamboo forests in infrared, near-infrared and other bands, optimal variables were selected to construct a layer-by-layer remote sensing classification method cascaded by multiple machine learning models such as random forest (RF) and XGBoost. This method achieved highprecision separation of bamboo forests from other land-cover types with an overall accuracy exceeding 0.95, which built a spatial foundation for biomass estimation. Secondly, within the bamboo distribution pixels, a model for estimating aboveground biomass was developed by integrating the random forest model with allometric equations. The biomass estimation achieved a coefficient of determination (R²) of 0.82, which outperformed single remote sensing models (with an average accuracy improvement of 28%). We applied this method to estimate the biomass of bamboo forests in Yanping District. The results showed that the aboveground biomass was 6.44×10⁴ tons, with high-value areas mainly being concentrated in the southwestern, northwestern, and eastern parts of the district. This approach provided a low-cost, high-efficiency, and replicable solution for estimating aboveground biomass in bamboo forests, offering critical data support for regional carbon sink inventory compilation, forest carbon sink trading project design, and precision management of bamboo forest ecosystems.

Key words: biomass estimation, Sentinel-2 remote sensing, machine learning, layer by layer classification, bamboo forest