Diversity of Remote Sensing-Based Variable Inputs Improves the Estimation of Seasonal Maximum Freezing Depth

Abstract

The maximum soil freezing depth (MSFD) is an important indicator of the thermal state of seasonally frozen ground. Its variation has important implications for the water cycle, ecological processes, climate and engineering stability. This study tested three aspects of data-driven predictions of MSFD in the Qinghai-Tibet Plateau (QTP), including comparison of three popular statistical/machine learning techniques, differences between remote sensing variables and reanalysis data as input conditions, and transportability of the model built by reanalysis data. The results show that support vector regression (SVR) performs better than random forest (RF), k-nearest neighbor (KNN) and the ensemble mean of the three models. Compared with the climate predictors, the remote sensing predictors are helpful for improving the simulation accuracy of the MSFD at both decadal and annual scales (at the annual and decadal scales, the root mean square error (RMSE) is reduced by 2.84 and 1.99 cm, respectively). The SVR model with climate predictor calibration using the in situ MSFD at the baseline period (2001–2010) can be used to simulate the MSFD over historical periods (1981–1990 and 1991–2000). This result indicates the good transferability of the well-trained machine learning model and its availability to simulate the MSFD of the past and the future when remote sensing predictors are not available.

Keywords

• seasonally frozen ground
• soil freezing depth
• statistical learning
• machine learning
• Qinghai-Tibet Plateau