1. Introduction
Permafrost is defined as ground (soil or rock and included ice and organic material) where temperature have remained at or below 0 °C for a period of least two consecutive years [
1,
2], which is an key component of cryosphere and assessing the impact of the climatic change. It is not only a crucial component in the soil biogeochemistry, local hydrology, ecosystems, heat and energy exchange, but also the sensitive indicator of global warming [
3]. The Qinghai-Tibet Plateau (QTP) known as “the Third Pole of the world” and “the Water Tower of Asia” represents 26.8% of the total area of China [
4]. The permafrost on the QTP is higher ground temperature, thin-thickness and poor-thermal stability, so it is extremely sensitive to the climate change and called as “driver and amplifier” of global climate change [
5]. Especially at high temperatures, in high-ice content regions, and human engineering activities (include the expansion of the QTP highway, the construction and maintenance of the QTP railway, oil pipeline and communication cable lying, etc.) destroy the topsoil structure of the permafrost and disturb the thermal balance between the surface and the atmosphere, which in turn affects the stability of engineering structures and increase environmental fragility [
6,
7,
8,
9,
10]. During recent decades, the permafrost of the QTP is experiencing significant degrading [
7,
11,
12] due to the combined effects of human activities and global climatic warming. Whereas, the QTP is an economically undeveloped area, how to plan the future infrastructures under such fragile ecological environment is a big challenge. Therefore, it is imperative to study the detailed permafrost distribution on the QTP in order to research the response and feedback of permafrost to human engineering activities and global climatic change.
In the past 50 years, many permafrost maps of the QTP are published in different scales. The previous permafrost distribution maps, such as the 1:600,000 permafrost map along the QTH [
13], 1:4,000,000 map of snow, ice, and frozen ground in China [
14] and 1:3,000,000 map of permafrost on the QTP [
15] were created using conventional or traditional classification method. The method is usually plotting the permafrost boundaries on the topographic maps by hand with aerial photographs, satellite images, air temperature, terrain analysis and empirical knowledge, so it is difficult to avoid artificial errors and some uncertainty in those maps [
16]. The most of recently published permafrost maps were compiled with three common permafrost distribution models, including: (a) Thermal offset model; (b) Empirical-statistical model; and (c) Process-oriented models [
17]. The process-oriented models and thermal offset model were based on finite-difference, finite-element and numerical heat flow [
18,
19,
20], which can reflect the changes of permafrost dynamics. The empirical-statistical model is based on establishing statistical relationships between altitude, latitude and permafrost distribution characteristics, such as elevation, mean annual ground temperature (MAGT), mean annual air temperature (MAAT) and low limit of permafrost. The mapping data were obtained through interpolation, extrapolation and statistical analysis based on field observation. However, the observation points and survey sites of the QTP were mostly located centrally in the Qinghai-Tibet Engineering Corridor (where the Qinghai-Tibet highway and Qinghai-Tibet railway crosses) and were relatively less in the western part of the QTP. Hence, due to the QTP’s huge extent, complex terrain, extreme climate, and inaccessibility of high altitude, it is difficult to accurately and comprehensively reflect the permafrost distribution over the entire QTP by the direct field investigations. In such a situation, the utilization of remote sensing data might be a fast and feasible manner.
Remote sensing cannot directly reflect the permafrost presence/absence, but it can help to determine the surface features of permafrost. In addition, the data acquired by remote sensing is more comprehensive compared to limited investigated data. Especially the use of higher spatial, spectral and temporal resolution remote sensing data does not just improve the accuracy of available data, but also prevents the insufficiency effect of field observations. During the last few decades, permafrost scientists used a portion remote sensing and developed several models to simulate the permafrost distribution on the QTP [
21,
22,
23,
24,
25,
26,
27,
28,
29,
30,
31], for example, Zou et al. [
16] generated a new permafrost map by the Temperature at the Top of Permafrost (TTOP) model combined with MODIS LST and ground observation datasets; Ran [
32] simulated the permafrost stability on the QTP by using a geographically weighted regression model based on remotely sensed land surface temperature, soil properties, leaf area index, and mean annual ground temperature of 142 boreholes. However, there is no map of permafrost distribution on the QTP that has been generated only by remote sensing data as yet. Therefore, the objective of this study is to make comprehensive use of remote sensing technology to obtain some surface environmental information (elevation, land surface temperature, vegetation distribution, and soil moisture) related to permafrost presence/absence, and fully use multi-source remote sensing data to study on permafrost mapping over the QTP. The mapping method is using decision tree algorithm, the classification thresholds of remotely sensed variables in decision tree analysis are defined by using the areas of agreement from two previous published permafrost maps (1:3,000,000 Map of Permafrost on the QTP Frozen soil and the permafrost map of the QTP derived from 1:4,000,000 Map of the Glaciers, Frozen Ground and Deserts in China), and remote sensing data from 2003 are used in order to reduce time gap with the two previous published permafrost maps. Finally, the result is verified by Western Kunlun, Wenquan and Gaize three areas. The consistency of permafrost distribution is evaluated by overall accuracy and Kappa coefficients among the different published permafrost maps over the QTP.
5. Discussion
According to the verification result over investigated regions and permafrost area comparison with other permafrost maps of the QTP, permafrost map over the QTP based on Multi-Source remote sensing data is better than two published permafrost maps of QTP. It also demonstrates the macroscopic, dynamic and convenient advantages of remote sensing data in permafrost research.
So far, a number of models (elevation model, Mean Annual Ground Temperature Model (MAGT) [
54], Emissivity Model [
61], and Annual Average Ground Surface Temperature Model [
62]) have been developed and they can basically reflect the permafrost distribution on the QTP. However, they still have several uncertainties, such as:
- (1)
The input parameter of elevation model is not enough and the resolution is coarser, which affect the accuracy of the result;
- (2)
The MAGT is short of consecutive ground temperature observations, so there is much error in Southeast Tibet permafrost regions, North Gangdise and near Himalaya Range;
- (3)
The inversion accuracy of the emissivity ratio in Emissivity Model is limited because of the influence of elevation and MAST Model cannot reflect the distribution of frozen soil in south and southeast Tibet due to scarcity of ground surface temperature observation data.
Besides, although the Frost Number [
63] and Temperature at the Top of Permafrost (TTOP) [
64] have been transferred to the permafrost distribution on the QTP [
65], the simulation results were not satisfactory. On the one hand, high-altitude features of the QTP was not considered because frost number was more suitable for high latitude permafrost distribution, on the other hand, topographic conditions of QTP are so complex that it affected accuracy. However, compared with the previous permafrost distribution models of the QTP, permafrost map based on multi-source remote sensing data does not use single factor like the Elevation Model or the Surface Frost Number Model, but comprehensively considers the governing factors (latitude, elevation, LST, NDVI and soil moisture) related to permafrost distribution. Then, on the local scale, the gradual classification is achieved by the approach of decision tree based on the importance of each factor. Furthermore, the permafrost map of the QTP derived from 1:4,000,000 map of the Glaciers, Frozen Ground and Deserts in China compiled in 2002 and the Map of Permafrost on the Qinghai–Tibetan Plateau compiled in 1996, but Aqua satellite was launched on 4 May 2002. In order to reduce time gap with the prior information (two previous published maps), and keep data integrity and accuracy, this study selects multi-source remote sensing data of 2003 year.
When it comes to a series of published permafrost maps of the QTP simulated using different methods, the results of 1:3,000,000 map of permafrost distribution on the QTP by Li and Cheng (129.8 × 10
4 km
2 of permafrost, 122.4 × 10
4 km
2 of seasonally frozen ground) has been internationally recognized. In fact, the permafrost area of this map and the permafrost map over the QTP derived from the 1:4,000,000 map of snow, ice, and frozen ground in China (hereafter referred to as the “QTP88_map”) are overestimated [
36,
60,
66]. The QTP06_map (the frozen soil map of QTP based on MAGT) derived from 1:4,000,000 Map of the Glaciers, Frozen Ground and Deserts in China also was used benchmark (Wang, Rinke, et al., 2016), which permafrost area was 111.8 × 10
4 km
2. However, the permafrost map over the QTP based on multi-source remote sensing data (hereafter referred to as the “MRSD_QTP map”) in this study shows that the permafrost area and seasonally frozen ground area are 111.3 × 10
4 km
2 and 140.9 × 10
4 km
2 respectively, its permafrost area is most similar to the QTP06_map and the new map of the permafrost distribution on the Tibetan Plateau (hereafter referred to as the “QTP16_map”) provided by Zou et al. [
16]; the overall accuracy are 84.82%, 87.52% respectively, and kappa coefficients are 0.70 and 0.74 (
Table 5). In addition, the permafrost area from the permafrost stability type distribution map over the QTP in the 2000s provided by Ran et al. (2017) is approximately 107.9 × 10
4 km
2 when the extremely unstable type of permafrost is neglected because this kind of permafrost mainly belongs to frozen gravel and cave ice that are distributed below the lower limit of permafrost and usually is not added to the total area of permafrost [
55], this result is also closer to the MRSD_QTP map. The consistency between the two maps is 87.52%, and the Kappa coefficient is approximately 0.74, which illustrates this study generates a better result. However, this study has rather much difference with other frozen soil maps over the QTP (
Table 6), such as Regional statistical survey, HY/CLM4, elevation model and TTOP, which may be caused by permafrost degradation due to climate warming in different times, method variability and historic climate data uncertainties.
Although the permafrost map over the QTP based on multi-source remote sensing data performed better than the two published maps (QTP96_map and QTP06_map) and had substantial agreement with the investigated regions, it still had some uncertainties as follows: (1) The prior information was relatively imperative, although this study selected the two most representative permafrost maps at different stages as the reference basis, there are few uncertainties due to the own errors of those two maps; (2) The final uncertain pixels after the classification with decision tree method were reclassified by a regional growth search algorithm; although this had certain theoretical basis, it was still necessary to find a more scientific and reasonable approach to classify these pixels as permafrost or seasonally frozen ground; (3) AMSR-E soil moisture products used in this study is 25 km in spatial resolution; the resolution had not been substantially improved though it was resampled to 1 km, thereby soil moisture was used as ancillary data. With the development of microwave remote sensing, if the accuracy and resolution of soil moisture are improved, the accuracy of permafrost map will also be improved accordingly; (4) Compared with other prediction models, especially the physical model, the result of this study cannot reflect the temporal dynamic changes of permafrost distribution.
The accuracy verification results of permafrost maps show that the area of seasonally frozen ground of the investigated regions is more than that shown by the other permafrost maps in general and discrepancies exist in transition zone between permafrost and seasonally frozen soil. On the one hand, snow (thickness, duration, time and cover) solar radiation, topographic factor, climate change and other factors, except governing factors, will also affect permafrost spatial distribution. For example, the temperature is extremely low in the eastern, southern and abdomen mountains of the QTP where snow exists all year round, thick and long snow cover on the permafrost. It has the effect of heat preservation because the thermal conductivity of snow is very poor so that solar radiation during the daytime is hard to get into the soil and the heat is not easy to lose in the nighttime. However, in the high plains, valleys and basins of high-altitude areas on the QTP, snow thickness is thin and duration is short, so snow might plays a cooling role [
71]; On the other hand, time difference among the different permafrost maps of the QTP in which climate warming caused permafrost degradation, will lead to the area of permafrost reducing and the area of seasonally frozen ground increasing gradually. Therefore, it is possible to generate a more accurate and reasonable permafrost map of the QTP integrating with passive microwave snow products.
6. Conclusions
This study indirectly obtained the permafrost map on the QTP by exploiting the relationship between environmental factors and frozen ground rather than limited investigated data, it is a pragmatic approach for regional scale permafrost mapping. Compared to traditional mapping method, spatial data derived from the satellite sensors has the macro, dynamic and comprehensive advantages, and remedy the lack of investigated data.
Statistical result of this study shows that permafrost area is 42.5 percent of the QTP area (111.3 × 104 km2), and the seasonally frozen ground area is 53.8 percent of the QTP area (140.9 × 104 km2), which has the highly consistent with benchmark map (The Map of Permafrost on the Qinghai–Tibetan Plateau). Moreover, the validation result shows that permafrost map of QTP based on multi-source spatial data has high overall accuracy and kappa coefficient with three regions (West Kunlun, Wenquan and Gaize) compared other two existing permafrost maps.
Remote sensing techniques have considerable potential for use as a tool in regional scale permafrost mapping, which might provide a simple and promising approach for further permafrost research. Furthermore, we believe that combined with passive microwave snow products, soil types, energy exchange process model between ground and atmosphere, terrain factors and physical mechanism of permafrost, the mapping model would be improved and might yield still more accurate results.