Classification of Shangri-La grasslands based on Google Earth Engine and remote sensing estimation of their biomass

The division of grassland types and the estimation of their biomass are of great importance for grassland management and conservation. The Shangri-La grasslands are rich in resources, diverse in types, and represent a typical example of high-altitude grasslands found at similar latitudes. To enhance the quality of grassland resource information in complex highland terrain, Shangri-La City was selected as the study area. Based on the Google Earth Engine cloud platform and Sentinel-2 remote sensing images, 33 original features were constructed combining spectral, vegetation index, texture, and terrain characteristics. Feature optimization was performed using recursive feature elimination and feature importance scores. Four algorithms: Random forest, support vector machine, classification and regression tree, and gradient boosting decision tree were employed to extract and classify the grasslands in the study area. Finally, the aboveground biomass (AGB) was inverted based on ground plot data. The RFE algorithm compressed the number of features to 21, and the altitude feature had the highest importance in the classification of grassland types. The RF algorithm had the highest classification accuracy, with an overall accuracy of 91.41% and a Kappa coefficient of 88.18%. Shangri-La grasslands can be divided into five types, with a total area of 3 265.77 km². The largest type was the subalpine meadow, covering an area of 2 230.03 km², accounting for 68.28% of the total grassland area, followed by alpine meadows, accounting for approximately 18.42% of the total grassland area. A quadratic polynomial model predicting aboveground biomass with the difference vegetation index was established, with an R² of 0.783 and a root mean square error of 154.72 g·m⁻². The total AGB of the Shangri-La grasslands was estimated at 1.5215 million tons, with subalpine meadows contributing 1.0241 million tons, accounting for 67.31% of the total AGB, followed by alpine meadows with 273 700 tons, accounting for 17.99% of the total AGB. This study successfully employed remote sensing technology and machine learning to classify grassland types and estimate biomass in the Shangri-La region. These findings not only provide a scientific basis for the management and conservation of plateau grasslands but also offer effective methodologies for research on similar ecosystems.