4.1. Comparisons with previous results
The most likely permafrost area on the QTP is 1.04 × 106 km2 (the region where MAGT < 0°C, Figure 4), or about 45.4% of the total QTP land surface area. Our results were compared with the permafrost distribution map of the QTP for the period 2003–2012 based on the TTOP model, which was basically consistent with the new permafrost area (1.06 × 106 km2, Zou et al., 2017). The two results showed substantial consistency, with a kappa coefficient of 0.63 (Table 3). However, there were still certain spatial differences (Figure 9). These differences mainly occurred at the southern margin of the continuous permafrost and the islands of permafrost in the south eastern QTP.
For the results of MAGT and ALT, a similar study showed relatively large deviations at the hemispheric scale (the RMSEs of MAGT and ALT were 1.6°C and 0.89 m, respectively; Aalto et al., 2018). In their study, an interesting discovery was mentioned, for both MAGT and ALT: after considering the area north of 60°N, the uncertainty was greatly reduced. This is primarily due to the fact that the permafrost on the QTP is quite different from that of the pan-Arctic region. The QTP is the dominant by the high-altitude permafrost, while the pan-Arctic is mainly the high-latitude permafrost. Compared with the pan-Arctic region, the active layer on the QTP is thicker, the ground temperature is higher, and the spatial heterogeneity is greater (Nicolsky et al., 2017; Cao et al., 2017; Qin et al., 2017). Therefore, combining the QTP permafrost and the pan-Arctic permafrost hemispherically will inevitably reduce the accuracy of the results.
We further compared the simulated results of MAGT and ALT with previous studies on the QTP. Table 4 summarizes the error statistics among different types of permafrost models (i.e., equilibrium model, transient model and statistical model). We can find that for the R-value, our method combined of the statistical and ML has the similar accuracy with the transient model. Although the RMSE of ALT in our study is the largest among all models, the differences are not significant. Moreover, the RMSE of MAGT in our study shows relatively smaller error. Meanwhile, from the overall spatial distribution of the ALT, although there are some differences in the spatial details, the distribution pattern of our result is comparable with the presented recently (Zhao and Wu, 2019; Wang et al., 2020b). In generally, our model can obtain a relatively higher simulation accuracy.
We qualitatively analyzed the main reasons for these differences as follows. Firstly, there are differences in accuracy among different types of models, such as the equilibrium models and mechanistic transient models. Secondly, there is a slight gap between the research period and the data used for verification. Permafrost is often viewed as a product of long-term climate change, which is slowly changing (Zhang et al., 2007); this may also lead to differences between the results. Finally, the 0.1° resolution of our model can’t capture all of regional information on climate change, which may limit the model’s ability to capture detailed changes in the permafrost to some extent, especially in the boundary of the permafrost region (Etzelmüller, 2013; Guo and Wang, 2016). Therefore, the ability to capture the permafrost edge information should be further improvement. Overall, by comparing with previous studies on the QTP, that our method is relatively simple and effective, and thus could be a useful tool to evaluate the permafrost conditions with a high accuracy on the QTP.