Spatio-Temporal Characteristics of Permafrost Evolution Based on Ground Surface Temperature Simulated by Random...

Changes in permafrost in cold regions directly affect groundwater circulation and biogeochemical processes. To accurately capture these complex permafrost-groundwater interactions, precise identification of permafrost thermal regimes is of paramount importance. In large-scale permafrost modeling, the model for the temperature at the top of permafrost (TTOP) has been widely applied due to its clear physical mechanisms and computational efficiency. However, most studies directly employ satellite-based land surface temperature (LST) products as forcing data, which often leads to biases in simulating permafrost distribution. In this study, we used a Random Forest (RF) regression model to reconstruct spatially continuous ground surface temperature (GST) data as the driving input to simulate permafrost characteristics in the Source Area of the Yellow River (SAYR) from 2000 to 2020. The results indicate that over the past two decades, the permafrost area has decreased by 2252 km², while the active layer thickness (ALT) has increased by 1.6 cm/10a, and the maximum depth of seasonally frozen ground (MFDSFG) has decreased by 7.4 cm/10a. Extensive ground ice thaw directly reshapes local aquifer frameworks, subsequently impacting groundwater routing and cycling processes. Compared with measured data from 132 boreholes, the accuracy of our permafrost mapping reached 79.5%. Validation against field observations reveals that incorporating GST as the forcing data for SAYR permafrost identification enhances the TTOP model accuracy by 6.4%. However, the result indicates that directly using LST products to drive the TTOP model in SAYR tends to overestimate the permafrost area by 5583 km², particularly in the marginal zones of permafrost degradation and around large lakes.