Peatland is a type of wetland that can store substantial carbon and retain groundwater. Knowledge of its spatial distribution in high-altitude areas of the world remains limited. My research in this field focuses on two entirely different plateaus: the Qinghai-Tibet Plateau (QTP) in western China and the Andes Plateau in South America. In both regions, I combine GIS-based spatial analysis with remote sensing and machine learning to map peatland extent and to link the extent changes to a variety of physical processes and human activities.
Since peatlands in the QTP are concentrated in the Zoige Basin and the Source Upper Yellow River watershed, my studies have focused on these areas. Using a Random Forest machine learning model, GIS-based terrain and hydrological variables, and training samples of peat depths measured directly in the field, we successfully identified the spatially variable peat depths across a sub-watershed within the Zoige Basin (the Black River watershed) at a spatial resolution of 13 m. Using an expanded set of training samples obtained from the Zoige Basin, we are now developing a two-stage modified Random Forest model to map the spatial distribution of peat depths across the entire Source Upper Yellow River watershed, an area exceeding 130,000 km². The core of this modified model is a method for overcoming the limitations posed by spatially clustered training samples—a challenge stemming from the naturally clustered distribution of peatlands themselves. The resulting spatially distributed peat depths will support subsequent investigations into reconstructing paleo-geomorphology and river networks, as well as the impact of peatlands on watershed hydrological regimes.
In the Andes Plateau, the study focuses on the Titicaca-Desaguadero-Poopó-Salares (TDPS) watershed, which spans Bolivia, Peru, and Chile. Peatlands, locally termed bofedales, are natural resources that provide various ecological services to local indigenous residents. Using a set of field samples of bofedales and other land use types (e.g., grassland and agricultural lands) collected manually in the summer of 2025, together with GIS-based spatial data integration and remote sensing imagery, we developed a machine learning method to identify the spatial distribution of bofedales across the TDPS watershed. Building on this work, we are examining spatial changes in bofedales driven by natural factors and human disturbance, such as glacier retreat, precipitation variability, population growth, and grazing management.