In the context of global warming, permafrost degrades gradually. To cope with the instability of the cryosphere, it is very important to strengthen the monitoring of the seasonal freeze-thaw cycle. At present, active and passive microwave remote sensing data are widely used in freeze/thaw (F/T) onset detection. There is some potential to improve accuracy through the combination of active and passive microwave data. Compared with the traditional method for combination, the machine learning algorithm has a stronger nonlinear expression ability. Therefore, it is advisable to use machine learning to combine multi-source data for freeze/thaw onset detection. In this study, the temporal change detection method is applied to SMAP data and ASCAT data respectively for preliminary detection. Then the Random Forest algorithm (RF) is used to combine the preliminary results of active and passive microwave data with site observation to estimate the freeze/thaw onsets more accurately. The method is validated with data obtained in Alaska from 2015 to 2019. The accuracy evaluation shows that the proposed method can effectively improve the accuracy of freeze/thaw onset detection. The predicted distribution of the freeze/thaw cycle indicates that the variation of the freeze-thaw cycle is closely related to latitude. In general, the proposed method based on machine learning is promising in the research of freeze-thaw state monitoring.

Long-term and high-quality surface soil moisture (SSM) and root-zone soil moisture (RZSM) data is crucial for understanding the land-atmosphere interactions of the Qinghai-Tibet Plateau (QTP). More than 40% of QTP is covered by permafrost, yet few studies have evaluated the accuracy of SSM and RZSM products derived from microwave satellite, land surface models (LSMs) and reanalysis over that region. This study tries to address this gap by evaluating a range of satellite and reanalysis estimates of SSM and RZSM in the thawed soil overlaying permafrost in the QTP, using in-situ measurements from sixteen stations. Here, seven SSM products were evaluated: Soil Moisture Active Passive L3 (SMAP L3) and L4 (SMAP-L4), Soil Moisture and Ocean Salinity in version IC (SMOS IC), Land Parameter Retrieval Model (LPRM) Advanced Microwave Scanning Radiometer 2 (AMSR2), European Space Agency Climate Change Initiative (ESA CCI), Advanced Scatterometer (ASCAT), and the fifth generation of the land component of the European Centre for Medium-Range Weather Forecasts atmospheric reanalysis (ERAS-Land). We also evaluated three RZSM products from SMAP-L4, ERA5-Land, and the Noah land surface model driven by Global Land Data Assimilation System (GLDAS-Noah). The assessment was conducted using five statistical metrics, i. e. Pearson correlation coefficient (R), bias, slope, Root Mean Square Error (RMSE), and unbiased RMSE (ubRMSE) between SSM or RZSM products and in-situ measurements. Our results showed that the ESA CCI, SMAP-L4 and SMOS-IC SSM products outperformed the other SSM products, indicated by higher correlation coefficients (R) (with a median R value of 0.63, 0.44 and 0.57, respectively) and lower ubRMSE (with a median ubRMSE value of 0.05, 0.04 and 0.07 m(3)/m(3), respectively). Yet, SSM overestimation was found for all SSM products. This could be partly attributed to ancillary data used in the retrieval (e.g. overestimation of land surface temperature for SMAP-L3) and to the fact that the products (e.g. LPRM) more easily overestimate the in-situ SSM when the soil is very dry. As expected, SMAP-L3 SSM performed better in areas with sparse vegetation than with dense vegetation covers. For RZSM products, SMAP-L4 and GLDAS-Noah (R = 0.66 and 0.44, ubRMSE = 0.03 and 0.02 m(3)/m(3), respectively) performed better than ERAS-Land (R = 0.46; ubRMSE = 0.03 m(3)/m(3)). It is also found that all RZSM products were unable to capture the variations of in-situ RZSM during the freezing/thawing period over the permafrost regions of QTP, due to large deviation for the ice-water phase change simulation and the lack of consideration for unfrozen-water migration during freezing processes in the LSMs.

Long-term and high-quality surface soil moisture (SSM) and root-zone soil moisture (RZSM) data is crucial for understanding the land-atmosphere interactions of the Qinghai-Tibet Plateau (QTP). More than 40% of QTP is covered by permafrost, yet few studies have evaluated the accuracy of SSM and RZSM products derived from microwave satellite, land surface models (LSMs) and reanalysis over that region. This study tries to address this gap by evaluating a range of satellite and reanalysis estimates of SSM and RZSM in the thawed soil overlaying permafrost in the QTP, using in-situ measurements from sixteen stations. Here, seven SSM products were evaluated: Soil Moisture Active Passive L3 (SMAP L3) and L4 (SMAP-L4), Soil Moisture and Ocean Salinity in version IC (SMOS IC), Land Parameter Retrieval Model (LPRM) Advanced Microwave Scanning Radiometer 2 (AMSR2), European Space Agency Climate Change Initiative (ESA CCI), Advanced Scatterometer (ASCAT), and the fifth generation of the land component of the European Centre for Medium-Range Weather Forecasts atmospheric reanalysis (ERAS-Land). We also evaluated three RZSM products from SMAP-L4, ERA5-Land, and the Noah land surface model driven by Global Land Data Assimilation System (GLDAS-Noah). The assessment was conducted using five statistical metrics, i. e. Pearson correlation coefficient (R), bias, slope, Root Mean Square Error (RMSE), and unbiased RMSE (ubRMSE) between SSM or RZSM products and in-situ measurements. Our results showed that the ESA CCI, SMAP-L4 and SMOS-IC SSM products outperformed the other SSM products, indicated by higher correlation coefficients (R) (with a median R value of 0.63, 0.44 and 0.57, respectively) and lower ubRMSE (with a median ubRMSE value of 0.05, 0.04 and 0.07 m(3)/m(3), respectively). Yet, SSM overestimation was found for all SSM products. This could be partly attributed to ancillary data used in the retrieval (e.g. overestimation of land surface temperature for SMAP-L3) and to the fact that the products (e.g. LPRM) more easily overestimate the in-situ SSM when the soil is very dry. As expected, SMAP-L3 SSM performed better in areas with sparse vegetation than with dense vegetation covers. For RZSM products, SMAP-L4 and GLDAS-Noah (R = 0.66 and 0.44, ubRMSE = 0.03 and 0.02 m(3)/m(3), respectively) performed better than ERAS-Land (R = 0.46; ubRMSE = 0.03 m(3)/m(3)). It is also found that all RZSM products were unable to capture the variations of in-situ RZSM during the freezing/thawing period over the permafrost regions of QTP, due to large deviation for the ice-water phase change simulation and the lack of consideration for unfrozen-water migration during freezing processes in the LSMs.