4.3. Model performance and uncertainty analysis
Our study integrated field observation data, meteorological data, geospatial environmental predictors and multiple statistical models to study MAGT and ALT changes in the present and future QTP permafrost regions. Based on the CV analysis, the reliability of both predictions displayed relatively low uncertainty. For MAGT, the benefits of using the ensemble modeling approach were obvious, i.e., the average of the four methods yielded the best simulation result. For ALT, large errors still existed among the ensemble modeling approach after CV, which indicating that the method does not always produce the most reliable results. The simulation accuracy of ALT is lower than that of MAGT, and the result can only represent the general change trend of ALT. The main reason for this is that, the spatial heterogeneity of ALT on the QTP is large, with the change rate of ALT per unit (100 m2) reaching 80%, thus resulting in the relatively low R2values and large RMSEs (Cao et al., 2017). Additionally, our model predicts the equilibrium state of permafrost and does not consider the lag time associated with the formation and degradation of permafrost (Xu et al., 2017b). Compared with previous studies, although our results show great reliability, there are still some uncertainties embedded in the predictions, including the measurement accuracy of the data, the equilibrium assumption in the statistical modeling and the influence of other factors (Aalto et al., 2018).
Due to the limitations of the observation data, we had to use one-year or multi-year averages to represent the present state and to fit the model. MAGT and ALT changed during this period, however, in particular, ALT changed greatly at the inter-annual scale. We did our best to collect datasets with MAGT and ALT, but the number of sample points used for training was still limited, and the model was still highly sensitive to single observations. To some extent, this also indicates that the number of observation sites on the QTP is too sparse to represent the present large spatial heterogeneity of the plateau.
When calculating the input factors of the model, in the future warming scenarios, the TDD and FDD were calculated based on the monthly mean air temperature instead of the daily mean air temperature. This approximate calculation method will bring some unavoidable errors, especially when the temperature is close to 0 ℃ (Wu et al., 2011; Shi et al., 2019). Additionally, we simply take 0°C temperature as the critical temperature threshold between solid precipitation and liquid precipitation, while, in most cases, snowfall events even occur in some regions on the QTP when the air temperature is > 4°C, but not 0 ℃ (Wang et al., 2016).
In this study, some key soil parameters, including soil texture, soil moisture content and bulk density, were excluded from the analyses in the model due to missing data, which exerted strong influence on water and heat transfer in the active layer as well as the change in permafrost temperature (Wu et al., 2017b; Du et al., 2020). The PISR and SOC in permafrost region are not static. However, it was assumed to be the fixed value in our model. With the further research on the key predictors of the permafrost region, we will add more dynamic datasets to our model. In summary, we used statistical and ML models combined with easily accessible data to simulate the present and future dynamics of permafrost on the QTP. By comparison and verification, our model can obtain high precision results through a relatively simple calculation process.
5. Conclusions
In this study, the method combined of statistical and ML was used to obtain the key permafrost metrics in both the present and a half-century in the future (2061−2080) on the QTP. Based on the comparison with in situ observation data and previous researches, we found that this method was reliable for simulating the changes in MAGT and ALT. We demonstrated the permafrost degradation from a quantitative perspective. Our results can provide a scientific basis for the study of climate change in permafrost. The main conclusions are listed as follows:
  1. A combination method of statistical and ML models is efficient to capture the changes in the thermal state of the permafrost on the QTP.
  2. The present (2000−2015) permafrost area on the QTP is approximate to be 1.04 × 106 km2. The average MAGT and ALT of the permafrost region amount to -1.35 ± 0.42ºC and 2.3 ± 0.60 m, respectively.
  3. In the future (2061−2080), the maximum permafrost area may be reduced to 0.44 × 106 km2. The future changes of MAGT and ALT are forecast to be pronounced, but region-specific.
  4. The unstable permafrost mainly distributed at the edge of the permafrost region, and approximately half permafrost in the QTEC will be at risk of disappearing in the future.