Next Article in Journal
Low-Cost Sensor Monitoring of Air Quality Indicators during Outdoor Renovation Activities around a Dwelling House
Previous Article in Journal
Varying Drivers of 2013–2017 Trends in PM2.5 Pollution over Different Regions in China
Previous Article in Special Issue
Prediction of Permafrost Subgrade Thawing Settlement in the Qinghai–Tibet Engineering Corridor under Climate Warming
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Atmosphere
Volume 15
Issue 7
10.3390/atmos15070788
atmosphere-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Susceptibility Modeling and Potential Risk Analysis of Thermokarst Hazard in Qinghai–Tibet Plateau Permafrost Landscapes Using a New Interpretable Ensemble Learning Method

by
Yuting Yang
1,2,
Jizhou Wang
2,
Xi Mao
2,
Wenjuan Lu
2,*,
Rui Wang
3 and
Hao Zheng
3
1
Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou 730070, China
2
Chinese Academy of Surveying and Mapping, Beijing 100830, China
3
Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China
*
Author to whom correspondence should be addressed.
Atmosphere 2024, 15(7), 788; https://doi.org/10.3390/atmos15070788
Submission received: 3 June 2024 / Revised: 26 June 2024 / Accepted: 28 June 2024 / Published: 29 June 2024
(This article belongs to the Special Issue Research about Permafrost–Atmosphere Interactions)
Browse Figures
Versions Notes

Abstract

:
Climate change is causing permafrost in the Qinghai–Tibet Plateau to degrade, triggering thermokarst hazards and impacting the environment. Despite their ecological importance, the distribution and risks of thermokarst lakes are not well understood due to complex influencing factors. In this study, we introduced a new interpretable ensemble learning method designed to improve the global and local interpretation of susceptibility assessments for thermokarst lakes. Our primary aim was to offer scientific support for precisely evaluating areas prone to thermokarst lake formation. In the thermokarst lake susceptibility assessment, we identified ten conditioning factors related to the formation and distribution of thermokarst lakes. In this highly accurate stacking model, the primary learning units were the random forest (RF), extremely randomized trees (EXTs), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost) algorithms. Meanwhile, gradient boosted decision trees (GBDTs) were employed as the secondary learning unit. Based on the stacking model, we assessed thermokarst lake susceptibility and validated accuracy through six evaluation indices. We examined the interpretability of the stacking model using three interpretation methods: accumulated local effects (ALE), local interpretable model-agnostic explanations (LIME), and Shapley additive explanations (SHAP). The results showed that the ensemble learning stacking model demonstrated superior performance and the highest prediction accuracy. Approximately 91.20% of the total thermokarst hazard points fell within the high and very high susceptible areas, encompassing 20.08% of the permafrost expanse in the QTP. The conclusive findings revealed that slope, elevation, the topographic wetness index (TWI), and precipitation were the primary factors influencing the assessment of thermokarst lake susceptibility. This comprehensive analysis extends to the broader impacts of thermokarst hazards, with the identified high and very high susceptibility zones affecting significant stretches of railway and highway infrastructure, substantial soil organic carbon reserves, and vast alpine grasslands. This interpretable ensemble learning model, which exhibits high accuracy, offers substantial practical significance for project route selection, construction, and operation in the QTP.

1. Introduction

The Qinghai–Tibet Plateau (QTP) is renowned as the primary region for high-altitude permafrost distribution within mid- and low-latitude zones worldwide. At the same time, the QTP is also known as the “Third Pole of the Earth” [1]. The QTP harbors an extensive expanse of permafrost spanning an estimated 1.06 × 106 km2 [2]. Characterized by elevated temperatures and substantial ice content, permafrost in the QTP is highly susceptible to fluctuations in climate and environmental disturbances [3], making it a sensitive indicator of environmental change [4]. In recent decades, global warming has instigated the ongoing degradation of permafrost across the Northern Hemisphere [5]. The warming rate in the QTP is twice the global average [6], with a significant rate of permafrost degradation [7,8,9], leading to escalating ground temperatures, increased active layer depths, and the thawing of ground ice [10,11].
Permafrost degradation typically triggers thermokarst landscapes [12,13,14], such as thaw slumps and thermokarst lakes. Thaw slumps tend to develop on slopes [15], while thermokarst lakes form in relatively flat terrain [16]. As typical thermokarst hazards, they are located in areas of geomorphic disturbance and are capable of affecting infrastructure, influencing biogeochemical cycles, and potentially altering environmental conditions when exhibiting high clustering characteristics [12,14,15]. In recent years, thermokarst lakes have garnered considerable attention [13,17]; they originate from the melting of ice-rich permafrost, leading to surface subsidence and the formation of depressions where water accumulates, resulting in the formation of lakes [18]. The seasonal fluctuations of thermokarst lakes can affect the surrounding soil moisture and heat properties, leading to degradation of alpine meadow ecosystems [19]. Moreover, thermokarst lakes influence the stability of nearby infrastructure by conveying heat to the permafrost beneath embankments [20,21]. In the context of climate warming, it is anticipated that the hazards arising from thermokarst processes and permafrost degradation will intensify [22]. However, the potential risks of thermokarst hazards to infrastructure and biogeochemical cycles are often overlooked, leading to less research in these fields [14,23,24].
To assess the risk of regional thermokarst hazards in complex terrain conditions, monitoring the spatiotemporal changes in susceptible areas can be achieved using Geographic Information Systems (GISs) and Remote Sensing (RS) [13,14,25]. The assessment of thermokarst hazard susceptibility has been widely recognized as a crucial strategy for mitigating thaw settlement disaster [15,23]. Specifically, it entails evaluating the likelihood of hazards occurring in a particular area by analyzing local geological and environmental factors [22]. Recent related research advancements have witnessed a growing application of machine learning in assessing the susceptibility of thermokarst hazards [26,27]. Previous studies typically employed single machine learning models to learn patterns from various environmental factors and obtain assessments of thermokarst hazard susceptibility. However, with algorithmic advancements, stacking models have achieved even more satisfactory results in susceptibility assessment [28,29]. To the best of our knowledge, stacking models have yet to be applied in assessing the susceptibility of thermokarst hazards. Moreover, despite the high accuracy of machine learning models, their limited interpretability as black-box models poses a challenge to their application in disaster prediction and early warning systems [30]. To address this issue, recent research has focused on interpretable machine learning methods to enhance model transparency and establish trust between users and models. Examples include predicting coronary heart disease mortality using Shapley additive explanations (SHAP), estimating crop yield with interpretable machine learning, and studying traffic safety with local interpretable model-agnostic explanations (LIME) [31,32,33]. To address the aforementioned limitations, we employed an interpretable stacking algorithm to generate reliable susceptibility maps for thermokarst hazards.
In this study, firstly, historical thermokarst hazard locations and conditioning factors were collected, and the relationships between them were analyzed. Secondly, we proposed and applied the stacking model, obtaining reliable susceptibility maps for thermokarst hazards. Thirdly, multiple interpretability methods were employed to elucidate the decision-making process of the stacking model and analyze the key factors influencing thermokarst hazards. Finally, the potential risks to main infrastructure, soil organic carbon (SOC), and vegetation were assessed based on the susceptibility map obtained from the stacking model. The findings can offer assistance to decision-makers in crafting suitable strategies for future management.

2. Materials and Methods

2.1. Study Area

The QTP is renowned for its towering altitude and extensive permafrost extent, stretching across an area of approximately 1.06 × 106 km2 [2]. From 1980 to 2018, significant increases in temperature, precipitation, and soil moisture content have been observed in permafrost regions [34,35]. This region is characterized by abundant permafrost and ground ice, with a mean annual ground temperature (MAGT) of about 1.56 °C and a regional average active layer thickness (ALT) of 2.32 m [36]. The total area of the QTP is about 2.60 × 106 km2, with a longitude range of 73–104° E and a latitude range of 26–40° N. The QTP is known as the highest plateau in the world, with an average altitude of about 4000 m (Figure 1). Lakes are prevalent throughout the QTP, with lakes larger than 1 km2 covering an area of 5 × 104 km2, representing 1.67% of the total area of the QTP [37,38].

2.2. Data Sources and Processing

2.2.1. Data Preparation

Lake thermokarst hazards were utilized for susceptibility assessment and potential risk analysis in this study. Our catalog of lake thermokarst hazards systematically collected spatial distribution information, sourced from the existing dataset [40]. We screened a total of 136,187 thermokarst hazard sites within the study area for subsequent analysis, from which we randomly selected 10,000 thermokarst hazard points for machine learning modeling. A total of 1000 sample points were randomly selected and are marked in Figure 1. Additionally, an equal number of non-thermokarst hazard points were randomly generated in areas devoid of thermokarst hazards to ensure the robustness of our susceptibility evaluation. Afterwards, we partitioned the complete dataset, consisting of positive samples (thermokarst hazards) and negative samples (non-thermokarst hazards), into training and test sets with a ratio of 7:3 for subsequent machine learning modeling. Additionally, a detailed explanation of the modeling factors for the machine learning model can be found in Section 2.2.2.
In addition to the data involved in the machine learning modeling process, data on major infrastructures, the SOC density, and vegetation type were utilized for potential analysis. The infrastructure was sourced from the National Platform for Common Geospatial Information Services (https://www.tianditu.gov.cn/, accessed on 24 June 2024). The SOC density and vegetation types were obtained from papers [41,42].

2.2.2. Conditioning Factors

There are many factors that influence the formation of lake thermokarst hazards. According to data availability and hazard mechanisms [23,24,27], 10 conditioning factors were selected for susceptibility assessment. These conditioning factors include elevation, aspect, slope, topographic wetness index (TWI), MAGT, ALT, solar radiation, normalized difference vegetation index (NDVI), fine soil content (FSC), and precipitation (Figure 2). All thematic layer data were clipped to the study area and resampled to a spatial resolution of 1000 m.
Thermokarst is a geomorphic phenomenon initiated by the melting of permafrost containing high ice content [43]. As for permafrost in the QTP, elevation is the key factor determining the presence or absence of permafrost [44]; the elevation was sourced from the global digital elevation model (DEM) [45]. Aspect and slope were calculated using the elevation factor, which, respectively, affect the melting process of permafrost and the hydraulics within the slope [23]. The TWI was derived from public data [46], which serve to quantify the influence of terrain on hydrological processes [47]. Regions characterized by high TWI values typically exhibit abundant soil moisture, facilitating the formation of ground ice [48]. Permafrost is a prerequisite for the generation of thermokarst hazards. MAGT and ALT, as indicators of permafrost status, were obtained from publicly available data, with a spatial resolution of 1 km for the time range of 2000 to 2016 [36]. High solar radiation can lead to ground ice melting [49]. The solar radiation factor was calculated as an average from publicly available data from 2000 to 2016 [50]. In permafrost areas, vegetation plays a crucial role in moderating solar radiation and reducing soil temperature, thereby maintaining the stability of the permafrost layer [51]. The NDVI was calculated as the average of the annual mean values from 2000 to 2016 based on three Landsat satellite datasets (Landsat 5, 7, and 8) using Google Earth Engine (GEE) [52]. The FSC (e.g., clay and silt) helps accommodate excess ice [53], and was obtained from publicly available data [54]. Climatic factors, such as precipitation, wield considerable influence over the thermal regime of permafrost [24]. Precipitation was calculated as the average annual rainfall accumulation spanning from 2000 to 2016 from publicly available data [55].

2.2.3. Multicollinearity Test

The detection of multicollinearity is an essential prelude to susceptibility modeling due to its potential to introduce predictive errors [56,57]. In this study, we utilized the Pearson correlation coefficient, variance inflation factor (VIF), and tolerance (TOL) to assess the correlation among the selected conditioning factors. For the Pearson correlation coefficient, each factor’s correlation coefficient should adhere to the following criterion, |R| ≤ 0.75, indicating the independence for each factor [58]. As for the VIF and TOL, the VIF is a metric that evaluates the relationship between a predictor variable and other variables. The higher the VIF value, the stronger the collinearity between the variable and others. TOL is the reciprocal of the VIF. The lower the TOL value, the stronger the collinearity between that variable and others. VIF > 10 or TOL < 0.1 signifies multicollinearity issues necessitating variable elimination [59].
As illustrated in Figure 3, the correlation among conditioning factors in this study was below 0.7, confirming that the 10 factors employed in this analysis satisfy the prerequisite of mutual independence. Moreover, the findings presented in Table 1 also affirmed the suitability of the conditioning factors for this study.

2.3. Modeling Methods

2.3.1. Ensemble Learning Framework

Susceptibility evaluation often involves multiple environmental factors due to the complex causes of thermokarst hazards [15,23,26]. Facing diverse environmental factors, the stacking algorithm has achieved satisfactory results in susceptibility assessment in recent years [28,29]. Therefore, we adopted a stacking model to evaluate the susceptibility to thermokarst hazards. The flowchart in this study is shown in Figure 4. The core of the stacking algorithm consists of a primary learner and a secondary learner. The primary learner is responsible for extracting features. These features are used by the secondary learner to continue learning and obtain the final evaluation results. In this study, the stacking model consisted of two layers. The first layer (i.e., the primary learner) selected ensemble learning models such as random forest (RF), extremely randomized trees (EXTs), categorical boosting (CatBoost), and extreme gradient boosting (XGBoost) due to their powerful performance and complex principles. The second layer (i.e., the secondary learner) used gradient boosted decision trees (GBDTs) to learn the results of the primary learner.
The flowchart for thermokarst hazard susceptibility assessment consists of four parts (Figure 5). First, we constructed the thermokarst hazard list and conditioning factors of the dataset, and then performed collinearity tests on the dataset using the Pearson correlation coefficient, VIF, and TOL. Second, we obtained the best-performing stacking model composed of multiple machine learning algorithms, based on random partitioning of the above dataset into training data (70%) and testing data (30%). We also compared the accuracy of thermokarst hazard susceptibility prediction results from the stacking model and other machine learning models. Third, three interpretability methods, SHAP, LIME, and accumulated local effects (ALE), were used to explore the reasoning process of the best-performing machine learning model and then quantitatively illustrate the contribution of conditioning factors to the formation of thermokarst hazards. Fourth, the potential risks of major infrastructure, SOC, and vegetation were analyzed based on this thermokarst hazard susceptibility map, and the Qinghai–Tibet Highway (QTH) and Qinghai–Tibet Railway (QTR) in high-risk areas were further discussed.

2.3.2. Model Performance

To compare the performance of the six models, training data were used in the modeling process, while testing data were used to obtain evaluation metrics. Thermokarst hazard susceptibility evaluation is a classification task, so we chose accuracy, precision, recall, F1-score, and area under the curve (AUC) as evaluation criteria, aligning with similar studies on machine performance [27,60,61]. These formulas are shown in Equations (1)–(4) [29]:
Accuracy = T P + T N T P + F P + T N + F N
Precision = T P T P + F P
Recall = T P T P + F N
F 1 - score = 2 × Precision × Recall Precision + Recall
where TP (True Positive) and TN (True Negative) represent the number of correctly classified positive and negative samples, respectively, while FP (False Positive) and FN (False Negative) represent their respective opposite numbers.
The receiver operating characteristic curve (ROC) and AUC values serve as reliable measures for accurately evaluating various disaster susceptibility models [62]. The AUC value, ranging from 0 to 1, reflects the area under the ROC curve.
Moreover, the AUC value is chosen as an additional metric to quantitatively assess model performance. It categorizes model performance as excellent (0.9–1), very good (0.8–0.9), good (0.7–0.8), moderate (0.6–0.7), and poor (0.5–0.6) [63].

2.4. Model Interpretability Techniques

The lack of interpretability in the thermokarst hazard susceptibility model poses potential risks due to the inherent nature of machine learning models resembling black boxes. Consequently, users are unable to fully comprehend how the model performs inference predictions. To address this issue, ALE, LIME, and SHAP were incorporated. These techniques aim to enhance the transparency of the model and provide insights into the contribution of individual variables to the model’s outputs.

2.4.1. Accumulated Local Effects

Daniel W. Apley and Jingyu Zhu introduced the accumulated local effects (ALE) method to mitigate the impact of feature correlation and isolate the pure effect of features on prediction outcomes [64]. ALE achieves this by computing local effects, which involves sorting a feature of interest and partitioning it into intervals to determine the boundary value of each interval along with the corresponding samples. Subsequently, the feature values within each interval are replaced with their respective boundary values, and the predicted value for each sample at the boundary value is computed. Since the intervals are very small, the values of other relevant features within each interval remain largely unchanged. Consequently, any difference in predicted values between the two boundary points can be attributed solely to variations in the values of the feature of interest. This approach effectively isolates the pure influence of individual feature variables. Moreover, the cumulative local effect method aggregates individual local effects, providing insights into the overall impact of a single feature variable on prediction outcomes.

2.4.2. Local Interpretable Model-Agnostic Explanations

Local Interpretable Model-Agnostic Explanations (LIME) is a model-agnostic algorithm proposed by Marco et al. [65]. LIME aims to provide interpretable explanations for individual predictions made by classifiers. It operates under the assumption that the behavior of the classifier can be approximated by a simple linear model at a local level for a given sample. LIME achieves this by introducing small perturbations to the input data around the specific data point of interest. It then observes how the model’s predictions change in response to these perturbations and assigns weights based on the proximity between the perturbed points and the original data. This process allows LIME to generate an interpretable model and prediction outcomes tailored to the local context.

2.4.3. Shapley Additive Explanations

Shapley Additive Explanations (SHAP) is a method widely employed in machine learning, encompassing deep learning, aimed at interpreting and comprehending model predictions. Rooted in cooperative game theory, SHAP utilizes Shapley values to attribute a value to each feature in a prediction, thereby revealing their individual contributions to the final outcome [66]. This approach provides valuable insights into feature importance and furnishes both global and local interpretability, facilitating a comprehensive understanding of model behavior. In this study, we performed interpretability analysis of a stacking model based on the “KernelExplainer” function from the “shap” package in Python 3.9.7.

3. Results

3.1. Relationship between Thermokarst Hazard and Conditioning Factor

Histograms were used to analyze the relationship between samples and ten conditioning factors (Figure 6). We obtained the percentages of thermokarst hazard and non-thermokarst hazard samples for each factor and plotted kernel density curves, along with the corresponding mean values (dashed lines). For elevation, thermokarst hazards exhibited clustering characteristics, with 86.76% of the hazards distributed in the range of 4381 to 5084 m, which was lower than the average elevation of non-thermokarst hazards. From the perspective of aspect, thermokarst hazards and non-thermokarst hazards exhibited similar characteristics, mainly distributed on the north-, northeast-, and south-facing slopes. Regarding slope, thermokarst hazards primarily occurred in areas with gentle slopes, with 92.64% of the hazards located in slopes ranging from 0 to 1.32°, which was significantly lower than the average slope of non-thermokarst hazards. The TWI indicates the tendency of water accumulation at any point within a catchment area. Thermokarst hazards were located in regions with higher TWI values. Regarding the MAGT, as the MAGT increased, the frequency of thermokarst hazards initially rose and then declined, with its mean value higher than that of non-thermokarst hazards. For ALT, thermokarst hazards were primarily located within the range of 216.92 to 270.12 cm, with a mean value of 228.78 cm. Intense solar radiation can lead to the melting of ground ice, resulting in the formation of thermokarst landforms. More than 81% of thermokarst hazards occurred in areas with solar radiation density ranging from 181.87 to 190.90 W/m2. Regarding the NDVI, over 57% of thermokarst hazards appeared in areas with NDVI values ranging from 0.03 to 0.13. From the perspective of the FSC, 76.41% of thermokarst hazards were distributed in areas where FSC values ranged from 51.12% to 67.56%. With increasing precipitation, the frequency of thermokarst hazards generally rose first, then fell, rose again, and then fell once more, and its average precipitation was higher than that of non-thermokarst hazards.

3.2. Assessment of Model Predictions

Table 2 shows the performance of each model in different evaluation indicators. Overall, the stacking model had the best performance, followed by CatBoost, RF, XGBoost, EXT, and GBDT. The AUC of the stacking model was 0.9322, the accuracy was 0.8627, the recall was 0.9040, the precision was 0.8334, and the F1-score was 0.8673. CatBoost also had a relatively good performance with an AUC of 0.9316, an accuracy of 0.8605, a recall of 0.8932, a precision of 0.8367, and an F1-score of 0.8641. RF exhibited an AUC of 0.9286, an accuracy of 0.8568, a recall of 0.8781, a precision of 0.8406, and an F1-score of 0.8589. XGBoost demonstrated an AUC of 0.9276, an accuracy of 0.8563, a recall of 0.8845, a precision of 0.8357, and an F1-score of 0.8594. EXT showed an AUC of 0.9275, an accuracy of 0.8600, a recall of 0.8872, a precision of 0.8398, and an F1-score of 0.8628. Compared to other models, GBDT obtained the lowest performance in susceptibility evaluation with an AUC of 0.9195, an accuracy of 0.8448, a recall of 0.8751, a precision of 0.8234, and an F1-score of 0.8484. The overall performance (AUC and ACC) of the stacking model reached the highest level, proving its excellent predictive ability in susceptibility evaluation.

3.3. Spatial Distribution of Thermokarst Hazard Susceptibility Maps

Upon completing model validation, the subsequent step entails generating the thermokarst hazard susceptibility map, esteemed for its utility in hazard risk management [67].
Following the integration of training outcomes from the stacking model and individual models such as CatBoost, RF, EXT, XGBoost, and GBDT into ArcGIS software, the likelihood of thermokarst hazard formation within each grid range was determined. Subsequently, the natural breaks (Jenks) method was employed to objectively categorize these results into five susceptibility classes: very low, low, moderate, high, and very high. In general, the susceptibility results of different models showed similar spatial distribution results, with high and very high susceptibility areas concentrated in the central QTP (refer to Figure 7).
To enhance the evaluation of the model’s generalization ability, we utilized a comprehensive dataset comprising 136,187 thermokarst hazard samples rather than the modeled dataset (i.e., 10,000 thermokarst hazard samples) to verify the accuracy. Table 3, Table 4 and Table 5 present characteristics of different susceptibility classes. The proportions of high and very high susceptibility marked by the six models were as follows: 20.08% for the stacking model, 19.76% for the CatBoost model, 18.50% for the EXT model, 18.50% for the RF model, 17.70% for the XGBoost model, and 19.80% for the GBDT model. The characteristics of the remaining three models are shown in Tables S1–S3. In particular, under the high and very high susceptibility classes, the proportions of thermokarst hazard points for the six models were as follows: 91.20% for the stacking model, 89.24% for EXT, 88.98% for RF, 88.96% for CatBoost, 87.68% for XGBoost, and 87.74% for GBDT. Notably, all values exceeded 87%, indicating satisfactory performance across all six models, with the stacking model accounting for the highest proportion.

3.4. Interpretability of the Ensemble Learning Model

3.4.1. Shapley Additive Explanations

Figure 8a provides insight into how individual factors impact model predictions across the entire dataset. The vertical axis represents each factor, while the horizontal axis represents the corresponding SHAP value, representing each factor’s influence on the model’s prediction. Factors are ordered by their total impact on the model, with the most influential factors at the top of the plot and the proportion of each factor’s importance shown. Among them, slope, elevation, TWI, and precipitation emerged as the four most influential characteristic factors. Meanwhile, aspect was the least important during training for predicting thermokarst hazards.
Figure 8b depicts the relationship between different factors and SHAP values. When the slope was low, the corresponding SHAP value was positive, indicating that the low slope value (i.e., flat terrain area) had a positive contribution to the development of the thermokarst hazard. Additionally, when elevation, TWI, and precipitation were relatively high, the corresponding SHAP value was also positive, indicating that these factors within this range facilitated the formation of thermokarst hazards.
Figure 9 provides a comprehensive insight into the impact of factor of the dataset. Specifically, it presents the relationship diagram among key factors such as slope, TWI, elevation, and precipitation. In Figure 9a, a positive SHAP value was observed when the slope was low, corresponding to a relatively high TWI value, indicating that flat terrain and a high soil moisture content were conducive to the development of thermokarst hazards. In Figure 9b, when the elevation value fell between approximately 4500 and 5000 m, the corresponding SHAP value was positive, suggesting a contributory role in the development of thermokarst hazards. Moreover, within this range, the slope values corresponding to most points were lower. In Figure 9c, similar to the analysis above, high TWI and small slope values jointly favored the development of thermokarst hazards. In Figure 9d, when precipitation was relatively high, the corresponding SHAP value was positive, suggesting a contributory role in developing thermokarst hazards. At the same time, the TWI values corresponding to most points were also relatively high, indicating a nonlinear relationship between higher precipitation and higher TWI values.
The SHAP summary plot offers explanations for the model from a global interpretability perspective, while also providing local explanations for individual samples. Figure 10 depicts the interpretation of single samples for predicted thermokarst hazards and non-thermokarst hazards, respectively, illustrating the impact of various sample characteristics on the model’s prediction outcomes. In these diagrams, red indicates features that increase the predicted probability of non-thermokarst hazards, while blue indicates features that elevate the likelihood of predicted thermokarst hazards. The length of the color denotes the degree of influence, and the black value on the graph represents the probability predicted by the model. Figure 10a elucidates the contribution rates of each characteristic factor when predicting thermokarst hazards. Notably, slope, precipitation, elevation, TWI, and MAGT all exhibit positive contribution rates, with slope having the highest contribution value of 0.17. As for Figure 10b, factors such as slope, TWI, and elevation had greater negative contributions to the negative sample. Incorporating SHAP values into our machine learning workflows enabled us to delve deeper into our models, ensuring transparency and extracting valuable insights from our data.

3.4.2. Accumulated Local Effects

ALE serves a dual purpose: it not only estimates the trend in prediction results but also eliminates the effects of correlation between factors, thereby emphasizing their impact on prediction outcomes. Here, we focused on analyzing the four most significant factors: slope, elevation, TWI, and precipitation. Figure 11a presents the ALE of the slope factor in the stacking model. When the slope value was less than approximately 5°, the predicted occurrence probability of thermokarst hazards tended to decrease as the slope value increased. As for Figure 11b, the ALE of the elevation in the stacking model revealed that within the range of 4500 to 5000 m, the predicted occurrence probability of thermokarst hazards was higher. Additionally, below 4500 m, the average predicted value increased with elevation, while above 5000 m, the average predicted value decreased with elevation. Moving to Figure 11c, the ALE indicated that when the TWI was greater than 0, it had a positive effect on the occurrence of thermokarst hazards, but it was not monotonic. Lastly, Figure 11d presents the ALE of the precipitation in the stacking model. When precipitation was within the range of 100 to 500 mm, it favored the formation of thermokarst hazards.

3.4.3. Local Interpretable Model-Agnostic Explanations

The LIME method endeavors to align the prediction outcomes of a simple model closely with those of the target prediction model by training a simplified model. Primarily, this method is employed to analyze the interpretability of prediction results for individual samples and furnish single-sample explanations for the target prediction model.
The samples depicted in the LIME model in Figure 12 correspond to the samples utilized in local interpretability within the SHAP model, ensuring that the results obtained from both interpretable models are comparable. Figure 12 illustrates the application of the LIME method to elucidate the prediction outcomes of the stacking machine learning model. The abscissa delineates the weights assigned by LIME to each factor in the model, wherein a higher weight signifies greater importance of the factor. Positive numbers indicate a positive impact of the factor on the result, while negative numbers denote the opposite influence. The analysis of Figure 12a revealed that the weight assigned to the slope was 0.12, signifying its utmost importance, followed by TWI and elevation, among others. Similarly, in Figure 12b, the weight attributed to the slope was −0.3, with the largest absolute value, indicating its paramount importance, followed by TWI, elevation, and so on. In practical scenarios, thermokarst hazards are more likely to occur in permafrost-rich areas characterized by lower slopes and higher soil moisture content.
The outcomes of interpretation employing the LIME technique were close to those of SHAP.

3.5. The Potential Risk Caused by Thermokarst Hazards

The importance of scientific research lies in guiding practice [23]. It is essential to assess the potential risks triggered by thermokarst hazards in order to better understand which areas of infrastructure are susceptible to such disasters and subsequently adopt appropriate mitigation measures. In this study, the main infrastructures (buildings, railway stations, railways, and highways) in the QTP were considered for potential risk analysis. Figure 13a illustrates the spatial distribution of the main infrastructure, and Table 6 presents the statistical results. Buildings and railway stations were less affected, while the road network was threatened by thermokarst hazards. In terms of railways, approximately 388.12 km was situated in high and very high susceptibility zones, with the majority located within the Qinghai–Tibet Engineering Corridor (QTEC). As for highways, about 1024.22 km was distributed in high and very high susceptibility zones.
The disturbance of permafrost affects the development of thermokarst landscapes, which are often associated with the emission of soil organic carbon (SOC) [14,17,68]. Moreover, variations exist in greenhouse gas fluxes within thermokarst landscapes situated under diverse vegetation types [69]. In this study, we conducted a statistical analysis of the distribution characteristics of SOC and vegetation types under high and very high susceptibility classes. Figure 13b,c illustrate their spatial distribution, while Table 7 and Table 8 present the statistical results. The estimated average SOC density for the high and very high classes was 12.69 kg/m2 and 12.64 kg/m2, respectively. The SOC storage in high and very high susceptibility areas was estimated to be 1.32 Pg C and 1.21 Pg C, respectively, accounting for 19.06% of the total. The dominant vegetation types were alpine grassland and alpine meadow, with areas of 115,457 km2 and 65,397 km2, respectively.

4. Discussion

4.1. Comparison of Typical Region

The Qinghai–Tibet Engineering Corridor (QTEC) is a key corridor connecting China’s inland cities to Tibet and contains important infrastructure, such as the QTH and QTR [27]. Permafrost in the region is predicted to be degrading, threatening the stability of infrastructure [16,24]. Consequently, we conducted a systematic examination of the results obtained from the stacking model concerning selected QTRs and QTHs within the QTEC, and infrastructure data were obtained from the Resource and Environmental Science Data Platform (https://www.resdc.cn/, accessed on 24 June 2024). Analyzing the potential risks associated with thermokarst hazards is imperative for discerning vulnerable segments of infrastructure and their spatial distribution and devising suitable management strategies.
According to the results in Figure 14, thermokarst hazards were mainly distributed in the central QTP. In the QTEC, this study pointed out that the high-risk areas for QTRs and QTHs were mainly concentrated from Budongquan to Beiluhe. According to Table 9, the length of QTH identified as the very low susceptibility class was 70.71 km, while that classified as the low susceptibility class covered 59.95 km. The moderate susceptibility area extended over 39.25 km, with the high susceptibility region including 113.70 km and the very high susceptibility area encompassing 222.45 km. The potential high-risk areas for QTHs are shown in Figure 14c. As for QTRs, the lengths attributed to each susceptibility level were as follows: the very low region reached 40.93 km, the low region covered 67.05 km, the moderate region extended over 52.15 km, the high region included 122.83 km, and the very high region encompassed 222.16 km. The potential high-risk areas for QTRs are shown in Figure 14d. The lengths corresponding to the high and very high susceptibility areas of QTHs and QTRs amounted to 336.15 km and 344.99 km, respectively, collectively representing 66.42% and 68.17% of the total length of the QTEC. This indicates that enhanced monitoring and management of infrastructure are needed in the relatively high susceptibility areas of the QTEC region.
The susceptibility assessment of thermokarst hazards provides insight into the spatial distribution of occurrence probabilities. This is achieved by integrating multiple conditioning factors and sample datasets into the model and determining the weight of each factor through expert knowledge or statistical methods, as discussed in previous studies. Traditionally, past research has predominantly relied on training samples to assess the reliability of results. However, in this study, we obtained additional thermokarst hazard datasets from the QTEC [70], which were not utilized in model training, including information on thermokarst hazard locations, to validate the results from the Tibetan Plateau Data Center (TPDC). Evaluating accuracy with additional datasets not only assesses the ability of machine learning models to generalize in addressing nonlinear problems but also tests the reliability of susceptibility predictions.
Table 10 presents statistics on the training data for various susceptibility classes used in this study and an additional dataset from Niu and Luo (2022). Notably, over 90% of the thermokarst hazard points in both datasets were situated in high susceptibility and very high susceptibility areas, suggesting the reliability of the results. Moreover, the distribution of thermokarst hazards in our study was largely consistent with the findings of Wang et al. [24]. Additionally, the proportion of hazard points in the high and very high susceptibility areas was 91.85%, exceeding Wang et al.’s reported outcome of 90.61% [24]. In summary, comparisons with other datasets and similar studies affirm the good reliability of our results.

4.2. Rationality of Model Selection

In this study, the process of model selection assumed paramount importance. However, the lack of established guidelines presented a significant challenge in the realm of machine learning. To tackle this concern, we chose a stacking model to evaluate thermokarst hazard susceptibility for the first time, comparing it with five alternative models. This decision stemmed from the acknowledged capability of stacking models to enhance predictive accuracy through the integration of diverse base models.
Initially, we conducted learning exercises across eleven prevalent machine learning models using a ten-fold cross-validation technique (Table 11). These models encompass distinct categories: CatBoost, EXT, RF, XGBoost, XGBoost, GBDT, AdaBoost, logistic regression (LR), Bayes, classification and regression trees (CARTs), k-nearest neighbors (KNN), and support vector machine (SVM). The test scores presented in Table 11 showed that strong learners had higher scores, and weak learners such as the KNN model had lower scores. While the SVM model is a strong learner, its score was relatively lower, possibly due to its higher data requirements. In contrast, tree-based models may have less stringent data requirements and therefore achieve higher accuracy.
Subsequently, we identified five models—RF, EXT, CatBoost, XGBoost, and GBDT—with superior performance for further analysis. After iterative experimentation, we identified the best combination of stacking models, employing RF, EXT, CatBoost, and XGBoost as primary learners, with GBDT serving as the secondary learner. This amalgamation effectively harnesses the strengths of individual models to yield superior predictive outcomes (Table 2).
While our study achieved commendable accuracy, it underscores the utilization of traditional models. Future research endeavors will pivot towards leveraging advanced methodologies, including novel hybrid deep learning models, forecasting the spatial distribution of thermokarst hazards under prospective climate scenarios, and employing optimization algorithms to fine-tune model parameters.

4.3. Model-Agnostic Interpretability

In the past, susceptibility studies based on machine learning typically relied on the average Gini index to gauge the significance of individual features [23,24,29]. However, due to its direct association with machine learning algorithms, this approach lacks universality. Conversely, utilizing algorithm-independent interpretability tools like SHAP, ALE, and LIME algorithms can provide a more objective assessment of evaluation factors, thereby enhancing the evaluation of thermokarst hazard susceptibility.
Previous research primarily concentrated on factor importance and the spatial distribution of susceptible areas, neglecting the influence of each factor on prediction probability. This study employed SHAP and ALE algorithms to investigate how evaluation factors affect the prediction probability of thermokarst hazards across various value ranges. Empirical findings demonstrate consistent trends in feature curves across different models, with the SHAP algorithm uncovering inter-factor interactions. Notably, the top four conditioning factors (i.e., slope, elevation, TWI, and precipitation) exhibited a positive correlation with thermokarst hazard prediction probability, with slope exerting the most significant influence. Positive SHAP values were observed when the slope was lower, indicating that the lower slope positively contributed to the development of thermokarst hazards, thereby promoting their formation. At the same time, in the case of lower slope values, most data points corresponded to relatively higher TWI values, indicating that lower slope values were generally associated with higher TWI values (Figure 9).
To address the lack of local explanation in prior disaster susceptibility models, this paper employs LIME and SHAP algorithms for localized interpretation, revealing spatial variations in thermokarst hazard probability. The congruence between the results of both interpretability algorithms further substantiates the reliability of the stacking model. Based on the interpretability analysis, slope, elevation, TWI, and precipitation emerged as key factors in the thermokarst hazard susceptibility evaluation, affirming their positive impact on prediction accuracy and underscoring the model’s effectiveness.

5. Conclusions

This study concentrated on the QTP and collected data pertaining to various factors, including elevation, aspect, slope, TWI, MAGT, ALT, solar radiation, NDVI, FSC, and precipitation. These 10 factors were utilized as susceptibility evaluation criteria. Through the implementation of the stacking model, the susceptibility of thermokarst hazards was thoroughly assessed, resulting in high evaluation accuracy. The primary findings are outlined as follows:
(1)
The stacking model emerged as the most appropriate method for evaluating the susceptibility of thermokarst hazards across the QTP. The stacking model demonstrated impressive performance with an AUC of 0.9332, an accuracy of 0.8627, a precision of 0.8334, a recall of 0.9040, and an F1-score of 0.8673, surpassing those of five other machine learning models overall. Remarkably, the results based on the stacking model indicate that 20.08% of permafrost regions in the QTP were located in high and very high susceptibility areas, encompassing 91.20% of all thermokarst hazard points.
(2)
From the global interpretation perspective, slope, elevation, TWI, and precipitation exerted the most significant influence on the susceptibility of thermokarst hazards within the QTP. Regions characterized by slope (<1.32°), elevation (4381–5084 m), TWI ((−0.49)–1.36), and precipitation (60.81–524.39 mm) were the main distribution areas of thermokarst hazards.
(3)
Based on the results from the stacking model, it is evident that areas prone to thermokarst hazards within the QTP were primarily concentrated in the central region. Overall, 388.12 km of railway, 1024.22 km of highway, 2.53 Pg of the SOC, and 115,457 km2 of alpine grassland were located in high and very high susceptibility zones. The QTEC is an area that merits special attention. About 336 km of QTH and 345 km of QTR in the QTEC were identified as high and very high susceptibility areas, and the potential risk was most obvious in the section from Budongquan to Beiluhe.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/atmos15070788/s1, Table S1: Characteristics of different susceptibility classes based on the EXT model. Table S2: Characteristics of different susceptibility classes based on the GBDT model. Table S3: Characteristics of different susceptibility classes based on the XGBoost model.

Author Contributions

Conceptualization, Y.Y. and R.W.; Methodology, Y.Y. and R.W.; Software, Y.Y.; Validation, Y.Y.; Writing—original draft, Y.Y.; Writing—review & editing, Y.Y., J.W., X.M., W.L., R.W. and H.Z.; Visualization, H.Z.; Supervision, J.W., X.M., W.L. and R.W.; Funding acquisition, J.W., X.M. and W.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Key Research and Development Program of China (Grant No. 2023YFF0611901) and the Exploration of Global Intelligent Positioning of Multilingual Place Names (Grant No. AR2412).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The original contributions presented in the study are included in the article. The simulation data are available upon request to [email protected].

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Abatzoglou, J.T.; Dobrowski, S.Z.; Parks, S.A.; Hegewisch, K.C. TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015. Sci. Data 2018, 5, 170191. [Google Scholar] [CrossRef]
  2. Amatulli, G.; McInerney, D.; Sethi, T.; Strobl, P.; Domisch, S. Geomorpho90m, empirical evaluation and accuracy assessment of global high-resolution geomorphometric layers. Sci. Data 2020, 7, 162. [Google Scholar] [CrossRef]
  3. Apley, D.W.; Zhu, J. Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models. J. R. Stat. Soc. Ser. B Stat. Methodol. 2020, 82, 1059–1086. [Google Scholar] [CrossRef]
  4. Biskaborn, B.K.; Smith, S.L.; Noetzli, J.; Matthes, H.; Vieira, G.; Streletskiy, D.A.; Schoeneich, P.; Romanovsky, V.E.; Lewkowicz, A.G.; Abramov, A.; et al. Permafrost is warming at a global scale. Nat. Commun. 2019, 10, 264. [Google Scholar] [CrossRef] [PubMed]
  5. Bowden, W.B. Climate Change in the Arctic—Permafrost, Thermokarst, and Why They Matter to the Non-Arctic World. Geogr. Compass 2010, 4, 1553–1566. [Google Scholar] [CrossRef]
  6. Chadburn, S.E.; Burke, E.J.; Cox, P.M.; Friedlingstein, P.; Hugelius, G.; Westermann, S. An observation-based constraint on permafrost loss as a function of global warming. Nat. Clim. Chang. 2017, 7, 340–344. [Google Scholar] [CrossRef]
  7. Chen, W.; Li, Y. GIS-based evaluation of landslide susceptibility using hybrid computational intelligence models. CATENA 2020, 195, 104777. [Google Scholar] [CrossRef]
  8. Chen, X.; Chen, W. GIS-based landslide susceptibility assessment using optimized hybrid machine learning methods. CATENA 2021, 196, 104833. [Google Scholar] [CrossRef]
  9. Chen, Z.; Chang, R.; Guo, H.; Pei, X.; Zhao, W.; Yu, Z.; Zou, L. Prediction of Potential Geothermal Disaster Areas along the Yunnan–Tibet Railway Project. Remote Sens. 2022, 14, 3036. [Google Scholar] [CrossRef]
  10. Chen, Z.; Chang, R.; Pei, X.; Yu, Z.; Guo, H.; He, Z.; Zhao, W.; Zhang, Q.; Chen, Y. Tunnel geothermal disaster susceptibility evaluation based on interpretable ensemble learning: A case study in Ya’an–Changdu section of the Sichuan–Tibet traffic corridor. Eng. Geol. 2023, 313, 106985. [Google Scholar] [CrossRef]
  11. Chou, Y.; Sheng, Y.; Zhu, Y. Study on the relationship between the shallow ground temperature of embankment and solar radiation in permafrost regions on Qinghai–Tibet Plateau. Cold Reg. Sci. Technol. 2012, 78, 122–130. [Google Scholar] [CrossRef]
  12. Das, S.; Datta, S.; Zubaidi, H.A.; Obaid, I.A. Applying interpretable machine learning to classify tree and utility pole related crash injury types. IATSS Res. 2021, 45, 310–316. [Google Scholar] [CrossRef]
  13. Ding, Y.; Mu, C.; Wu, T.; Hu, G.; Zou, D.; Wang, D.; Li, W.; Wu, X. Increasing cryospheric hazards in a warming climate. Earth-Sci. Rev. 2021, 213, 103500. [Google Scholar] [CrossRef]
  14. Fan, X.; Yunus, A.P.; Scaringi, G.; Catani, F.; Siva Subramanian, S.; Xu, Q.; Huang, R. Rapidly Evolving Controls of Landslides After a Strong Earthquake and Implications for Hazard Assessments. Geophys. Res. Lett. 2021, 48, e2020GL090509. [Google Scholar] [CrossRef]
  15. Gao, Z.; Niu, F.; Wang, Y.; Luo, J.; Lin, Z. Impact of a thermokarst lake on the soil hydrological properties in permafrost regions of the Qinghai-Tibet Plateau, China. Sci. Total Environ. 2017, 574, 751–759. [Google Scholar] [CrossRef]
  16. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sens. Environ. 2017, 202, 18–27. [Google Scholar] [CrossRef]
  17. Grosse, G.; Jones, B.M.; Arp, C.D. Thermokarst lakes drainage drained basins. In Treatise on Geomorphology; Shroder, J.F., Ed.; Elsevier: Amsterdam, The Netherlands, 2013; pp. 325–353. [Google Scholar] [CrossRef]
  18. Guodong, C. The mechanism of repeated-segregation for the formation of thick layered ground ice. Cold Reg. Sci. Technol. 1983, 8, 57–66. [Google Scholar] [CrossRef]
  19. Ibrahim, M.; Louie, M.; Modarres, C.; Paisley, J. Global Explanations of Neural Networks: Mapping the Landscape of Predictions [WWW Document]. arXiv 2019, arXiv:1902.02384v1. Available online: https://arxiv.org/abs/1902.02384v1 (accessed on 11 February 2024).
  20. in ’t Zandt, M.H.; Liebner, S.; Welte, C.U. Roles of Thermokarst Lakes in a Warming World. Trends Microbiol. 2020, 28, 769–779. [Google Scholar] [CrossRef]
  21. Jones, B.M.; Grosse, G.; Farquharson, L.M.; Roy-Léveillée, P.; Veremeeva, A.; Kanevskiy, M.Z.; Gaglioti, B.V.; Breen, A.L.; Parsekian, A.D.; Ulrich, M.; et al. Lake and drained lake basin systems in lowland permafrost regions. Nat. Rev. Earth Environ. 2022, 3, 85–98. [Google Scholar] [CrossRef]
  22. Lei, S.; Chen, H.; Bian, Z.; Liu, Z. Evaluation of integrating topographic wetness index with backscattering coefficient of TerraSAR-X image for soil moisture estimation in a mountainous region. Ecol. Indic. 2016, 61, 624–633. [Google Scholar] [CrossRef]
  23. Li, R.; Zhang, M.; Konstantinov, P.; Pei, W.; Tregubov, O.; Li, G. Permafrost degradation induced thaw settlement susceptibility research and potential risk analysis in the Qinghai-Tibet Plateau. CATENA 2022, 214, 106239. [Google Scholar] [CrossRef]
  24. Li, R.; Zhang, M.; Pei, W.; Melnikov, A.; Zhang, Z.; Li, G. Risk evaluation of thaw settlement using machine learning models for the Wudaoliang-Tuotuohe region, Qinghai-Tibet Plateau. CATENA 2023, 220, 106700. [Google Scholar] [CrossRef]
  25. Li, X.; Cheng, G. A GIS-aided response model of high-altitude permafrost to global change. Sci. China D Earth Sci. 1999, 42, 72–79. [Google Scholar] [CrossRef]
  26. Li, Y.; Wang, G.; Sun, S.; Lin, S.; Huang, P.; Xiao, J.; Guo, L.; Li, J.; Song, C. Methane Emissions From the Qinghai-Tibet Plateau Ponds and Lakes: Roles of Ice Thaw and Vegetation Zone. Glob. Biogeochem. Cycles 2024, 38, e2024GB008106. [Google Scholar] [CrossRef]
  27. Lin, Z.; Hu, G.; Zou, D.; Wu, X.; Ma, L.; Sun, Z.; Yuan, L.; Zhou, H.; Liu, S. Permafrost Changes and Its Effects on Hydrological Processes on Qinghai-Tibet Plateau. Bull. Chin. Acad. Sci. 2019, 34, 1233–1246. [Google Scholar] [CrossRef]
  28. Lin, Z.J.; Niu, F.J.; Fang, J.H.; Luo, J.; Yin, G.A. Interannual variations in the hydrothermal regime around a thermokarst lake in Beiluhe, Qinghai-Tibet Plateau. Geomorphology 2017, 276, 16–26. [Google Scholar] [CrossRef]
  29. Lundberg, S.; Lee, S.-I. An unexpected unity among methods for interpreting model predictions. arXiv 2016, arXiv:1611.07478. [Google Scholar] [CrossRef]
  30. Luo, D.; Wu, Q.; Jin, H.; Marchenko, S.S.; Lü, L.; Gao, S. Recent changes in the active layer thickness across the northern hemisphere. Environ. Earth Sci. 2016, 75, 555. [Google Scholar] [CrossRef]
  31. Lv, L.; Chen, T.; Dou, J.; Plaza, A. A hybrid ensemble-based deep-learning framework for landslide susceptibility mapping. Int. J. Appl. Earth Obs. Geoinf. 2022, 108, 102713. [Google Scholar] [CrossRef]
  32. Mateo-Sanchis, A.; Piles, M.; Amorós-López, J.; Muñoz-Marí, J.; Adsuara, J.E.; Moreno-Martínez, Á.; Camps-Valls, G. Learning main drivers of crop progress and failure in Europe with interpretable machine learning. Int. J. Appl. Earth Obs. Geoinf. 2021, 104, 102574. [Google Scholar] [CrossRef]
  33. Merghadi, A.; Yunus, A.P.; Dou, J.; Whiteley, J.; ThaiPham, B.; Bui, D.T.; Avtar, R.; Abderrahmane, B. Machine learning methods for landslide susceptibility studies: A comparative overview of algorithm performance. Earth-Sci. Rev. 2020, 207, 103225. [Google Scholar] [CrossRef]
  34. Mu, C.; Shang, J.; Zhang, T.; Fan, C.; Wang, S.; Peng, X.; Zhong, W.; Zhang, F.; Mu, M.; Jia, L. Acceleration of thaw slump during 1997–2017 in the Qilian Mountains of the northern Qinghai-Tibetan plateau. Landslides 2020, 17, 1051–1062. [Google Scholar] [CrossRef]
  35. Ni, J.; Wu, T.; Zhu, X.; Hu, G.; Zou, D.; Wu, X.; Li, R.; Xie, C.; Qiao, Y.; Pang, Q.; et al. Simulation of the Present Future Projection of Permafrost on the Qinghai-Tibet Plateau with Statistical Machine Learning Models. J. Geophys. Res. Atmos. 2021, 126, e2020JD033402. [Google Scholar] [CrossRef]
  36. Nitze, I.; Grosse, G.; Jones, B.M.; Romanovsky, V.E.; Boike, J. Remote sensing quantifies widespread abundance of permafrost region disturbances across the Arctic and Subarctic. Nat. Commun. 2018, 9, 5423. [Google Scholar] [CrossRef]
  37. Niu, F.; Lin, Z.; Lu, J.; Luo, J.; Wang, H. Assessment of terrain susceptibility to thermokarst lake development along the Qinghai–Tibet engineering corridor, China. Environ. Earth Sci. 2015, 73, 5631–5642. [Google Scholar] [CrossRef]
  38. Niu, F.; Luo, J. Distribution Data of Freezing-Thawing Hazards in Qinghai Tibet Engineering Corridor (2019–2020); National Tibetan Plateau Data Center: Beijing, China, 2022. [Google Scholar] [CrossRef]
  39. Obu, J.; Westermann, S.; Bartsch, A.; Berdnikov, N.; Christiansen, H.H.; Dashtseren, A.; Delaloye, R.; Elberling, B.; Etzelmüller, B.; Kholodov, A.; et al. Northern Hemisphere permafrost map based on TTOP modelling for 2000–2016 at 1 km2 scale. Earth-Sci. Rev. 2019, 193, 299–316. [Google Scholar] [CrossRef]
  40. Olefeldt, D.; Goswami, S.; Grosse, G.; Hayes, D.; Hugelius, G.; Kuhry, P.; McGuire, A.D.; Romanovsky, V.E.; Sannel, A.B.K.; Schuur, E.A.G.; et al. Circumpolar distribution and carbon storage of thermokarst landscapes. Nat. Commun. 2016, 7, 13043. [Google Scholar] [CrossRef]
  41. Peng, E.-X.; Sheng, Y.; Hu, X.-Y.; Wu, J.-C.; Cao, W. Thermal effect of thermokarst lake on the permafrost under embankment. Advances in Climate Change Research, Including special topic on degrading permafrost and its impacts. Adv. Clim. Chang. Res. 2021, 12, 76–82. [Google Scholar] [CrossRef]
  42. Peng, S.; Ding, Y.; Liu, W.; Li, Z. 1km monthly temperature and precipitation dataset for China from 1901 to 2017. Earth System Sci. Data 2019, 11, 1931–1946. [Google Scholar] [CrossRef]
  43. Poggio, L.; de Sousa, L.M.; Batjes, N.H.; Heuvelink, G.B.M.; Kempen, B.; Ribeiro, E.; Rossiter, D. SoilGrids 2.0: Producing soil information for the globe with quantified spatial uncertainty. Soil 2021, 7, 217–240. [Google Scholar] [CrossRef]
  44. Pourghasemi, H.R.; Sadhasivam, N.; Kariminejad, N.; Collins, A.L. Gully erosion spatial modelling: Role of machine learning algorithms in selection of the best controlling factors and modelling process. Geosci. Front. 2020, 11, 2207–2219. [Google Scholar] [CrossRef]
  45. Qin, Y.; Zhang, P.; Liu, W.; Guo, Z.; Xue, S. The application of elevation corrected MERRA2 reanalysis ground surface temperature in a permafrost model on the Qinghai-Tibet Plateau. Cold Reg. Sci. Technol. 2020, 175, 103067. [Google Scholar] [CrossRef]
  46. Qingbai, W.; Yongzhi, L. Ground temperature monitoring and its recent change in Qinghai–Tibet Plateau. Cold Reg. Sci. Technol. 2004, 38, 85–92. [Google Scholar] [CrossRef]
  47. Qiu, J. China: The third pole. Nature 2008, 454, 393–396. [Google Scholar] [CrossRef]
  48. Ran, Y.; Li, X.; Cheng, G.; Che, J.; Aalto, J.; Karjalainen, O.; Hjort, J.; Luoto, M.; Jin, H.; Obu, J.; et al. New high-resolution estimates of the permafrost thermal state and hydrothermal conditions over the Northern Hemisphere. Earth Syst. Sci. Data 2022, 14, 865–884. [Google Scholar] [CrossRef]
  49. Ribeiro, M.T.; Singh, S.; Guestrin, C. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. arXiv 2016, arXiv:1602.04938. [Google Scholar] [CrossRef]
  50. Schuur, E.a.G.; McGuire, A.D.; Schädel, C.; Grosse, G.; Harden, J.W.; Hayes, D.J.; Hugelius, G.; Koven, C.D.; Kuhry, P.; Lawrence, D.M.; et al. Climate change and the permafrost carbon feedback. Nature 2015, 520, 171–179. [Google Scholar] [CrossRef]
  51. Sørensen, R.; Zinko, U.; Seibert, J. On the calculation of the topographic wetness index: Evaluation of different methods based on field observations. Hydrol. Earth Syst. Sci. 2006, 10, 101–112. [Google Scholar] [CrossRef]
  52. Thi Ngo, P.T.; Panahi, M.; Khosravi, K.; Ghorbanzadeh, O.; Kariminejad, N.; Cerda, A.; Lee, S. Evaluation of deep learning algorithms for national scale landslide susceptibility mapping of Iran. Geosci. Front. 2021, 12, 505–519. [Google Scholar] [CrossRef]
  53. Wang, K.; Tian, J.; Zheng, C.; Yang, H.; Ren, J.; Liu, Y.; Han, Q.; Zhang, Y. Interpretable prediction of 3-year all-cause mortality in patients with heart failure caused by coronary heart disease based on machine learning, S.H.A.P. Comput. Biol. Med. 2021, 137, 104813. [Google Scholar] [CrossRef]
  54. Wang, P.; de Jager, J.; Nauta, A.; van Huissteden, J.; Trofim, M.C.; Limpens, J. Exploring near-surface ground ice distribution in patterned-ground tundra: Correlations with topography, soil and vegetation. Plant Soil. 2019, 444, 251–265. [Google Scholar] [CrossRef]
  55. Wang, R.; Guo, L.; Yang, Y.; Zheng, H.; Jia, H.; Diao, B.; Li, H.; Liu, J. Thermokarst lake susceptibility assessment using machine learning models in permafrost landscapes of the Arctic. Sci. Total Environ. 2023, 900, 165709. [Google Scholar] [CrossRef]
  56. Wang, R.; Guo, L.; Yang, Y.; Zheng, H.; Liu, L.; Jia, H.; Diao, B.; Liu, J. Thermokarst Lake Susceptibility Assessment Induced by Permafrost Degradation in the Qinghai–Tibet Plateau Using Machine Learning Methods. Remote Sens. 2023, 15, 3331. [Google Scholar] [CrossRef]
  57. Wei, Z.; Du, Z.; Wang, L.; Lin, J.; Feng, Y.; Xu, Q.; Xiao, C. Sentinel-Based Inventory of Thermokarst Lakes and Ponds Across Permafrost Landscapes on the Qinghai-Tibet Plateau. Earth Space Sci. 2021, 8, e2021EA001950. [Google Scholar] [CrossRef]
  58. Wen, Z.; Niu, F.; Yu, Q.; Wang, D.; Feng, W.; Zheng, J. The role of rainfall in the thermal-moisture dynamics of the active layer at Beiluhe of Qinghai-Tibetan plateau. Environ. Earth Sci. 2014, 71, 1195–1204. [Google Scholar] [CrossRef]
  59. Wu, T.; Wang, D.; Mu, C.; Zhang, W.; Zhu, X.; Zhao, L.; Li, R.; Hu, G.; Zou, D.; Chen, J.; et al. Storage, patterns, and environmental controls of soil organic carbon stocks in the permafrost regions of the Northern Hemisphere. Sci. Total Environ. 2022, 828, 154464. [Google Scholar] [CrossRef]
  60. Yamazaki, D.; Ikeshima, D.; Tawatari, R.; Yamaguchi, T.; O’Loughlin, F.; Neal, J.C.; Sampson, C.C.; Kanae, S.; Bates, P.D. A high-accuracy map of global terrain elevations. Geophys. Res. Lett. 2017, 44, 5844–5853. [Google Scholar] [CrossRef]
  61. Yang, M.; Wang, X.; Pang, G.; Wan, G.; Liu, Z. The Tibetan Plateau cryosphere: Observations and model simulations for current status and recent changes. Earth-Sci. Rev. 2019, 190, 353–369. [Google Scholar] [CrossRef]
  62. Yi, S.; Arain, M.A.; Woo, M.-K. Modifications of a land surface scheme for improved simulation of ground freeze-thaw in northern environments. Geophys. Res. Lett. 2006, 33. [Google Scholar] [CrossRef]
  63. Yin, G.; Luo, J.; Niu, F.; Lin, Z.; Liu, M. Machine learning-based thermokarst landslide susceptibility modeling across the permafrost region on the Qinghai-Tibet Plateau. Landslides 2021, 18, 2639–2649. [Google Scholar] [CrossRef]
  64. Yin, G.; Luo, J.; Niu, F.; Zhou, F.; Meng, X.; Lin, Z.; Liu, M. Spatial Analyses and Susceptibility Modeling of Thermokarst Lakes in Permafrost Landscapes along the Qinghai–Tibet Engineering Corridor. Remote Sens. 2021, 13, 1974. [Google Scholar] [CrossRef]
  65. You, Q.; Cai, Z.; Pepin, N.; Chen, D.; Ahrens, B.; Jiang, Z.; Wu, F.; Kang, S.; Zhang, R.; Wu, T.; et al. Warming amplification over the Arctic Pole and Third Pole: Trends, mechanisms and consequences. Earth-Sci. Rev. 2021, 217, 103625. [Google Scholar] [CrossRef]
  66. Zhang, G.; Chen, W.; Xie, H. Tibetan Plateau’s Lake Level and Volume Changes From NASA’s ICESat/ICESat-2 and Landsat Missions. Geophys. Res. Lett. 2019, 46, 13107–13118. [Google Scholar] [CrossRef]
  67. Zhang, G.; Yao, T.; Xie, H.; Yang, K.; Zhu, L.; Shum, C.K.; Bolch, T.; Yi, S.; Allen, S.; Jiang, L.; et al. Response of Tibetan Plateau lakes to climate change: Trends, patterns, and mechanisms. Earth-Sci. Rev. 2020, 208, 103269. [Google Scholar] [CrossRef]
  68. Zhao, L.; Wu, Q.; Marchenko, S.S.; Sharkhuu, N. Thermal state of permafrost and active layer in Central Asia during the international polar year. Permafr. Periglac. Process. 2010, 21, 198–207. [Google Scholar] [CrossRef]
  69. Zhou, G.; Ren, H.; Liu, T.; Zhou, L.; Ji, Y.; Song, X.; Lv, X. A new regional vegetation mapping method based on terrain-climate-remote sensing and its application on the Qinghai-Xizang Plateau. Sci. China Earth Sci. 2023, 66, 237–246. [Google Scholar] [CrossRef]
  70. Zou, D.; Zhao, L.; Sheng, Y.; Chen, J.; Hu, G.; Wu, T.; Wu, J.; Xie, C.; Wu, X.; Pang, Q.; et al. A new map of permafrost distribution on the Tibetan Plateau. Cryosphere 2017, 11, 2527–2542. [Google Scholar] [CrossRef]
Atmosphere 15 00788 g001
Figure 1. Spatial distribution of the study area. (a) Location of the QTP in the world; (b) location of sample points and permafrost distribution across the QTP [39].
Figure 1. Spatial distribution of the study area. (a) Location of the QTP in the world; (b) location of sample points and permafrost distribution across the QTP [39].
Atmosphere 15 00788 g001
Atmosphere 15 00788 g002
Figure 2. Conditioning factors for thermokarst hazard susceptibility modeling: (a) elevation, (b) aspect, (c) slope, (d) TWI, (e) MAGT, (f) ALT, (g) solar radiation, (h) NDVI, (i) FSC, and (j) precipitation.
Figure 2. Conditioning factors for thermokarst hazard susceptibility modeling: (a) elevation, (b) aspect, (c) slope, (d) TWI, (e) MAGT, (f) ALT, (g) solar radiation, (h) NDVI, (i) FSC, and (j) precipitation.
Atmosphere 15 00788 g002
Atmosphere 15 00788 g003
Figure 3. Correlation heat maps of conditioning factors.
Figure 3. Correlation heat maps of conditioning factors.
Atmosphere 15 00788 g003
Atmosphere 15 00788 g004
Figure 4. Flowchart of the stacking model.
Figure 4. Flowchart of the stacking model.
Atmosphere 15 00788 g004
Atmosphere 15 00788 g005
Figure 5. Flowchart of the proposed methodology for thermokarst hazard susceptibility mapping.
Figure 5. Flowchart of the proposed methodology for thermokarst hazard susceptibility mapping.
Atmosphere 15 00788 g005
Atmosphere 15 00788 g006
Figure 6. Relationship between thermokarst hazard and conditioning factor: (a) elevation, (b) aspect, (c) slope, (d) TWI, (e) MAGT, (f) ALT, (g) solar radiation, (h) NDVI, (i) FSC, and (j) precipitation.
Figure 6. Relationship between thermokarst hazard and conditioning factor: (a) elevation, (b) aspect, (c) slope, (d) TWI, (e) MAGT, (f) ALT, (g) solar radiation, (h) NDVI, (i) FSC, and (j) precipitation.
Atmosphere 15 00788 g006
Atmosphere 15 00788 g007
Figure 7. Classification map of the thermokarst hazard susceptibility: (a) stacking, (b) CatBoost, (c) EXT, (d) RF, (e) XGBoost and (f) GBDT.
Figure 7. Classification map of the thermokarst hazard susceptibility: (a) stacking, (b) CatBoost, (c) EXT, (d) RF, (e) XGBoost and (f) GBDT.
Atmosphere 15 00788 g007
Atmosphere 15 00788 g008
Figure 8. Global interpretability of the stacking model. (a) Bar plot; (b) beeswarm plot.
Figure 8. Global interpretability of the stacking model. (a) Bar plot; (b) beeswarm plot.
Atmosphere 15 00788 g008
Atmosphere 15 00788 g009
Figure 9. SHAP dependence plots for (a) slope and TWI; (b) elevation and slope; (c) TWI and slope; (d) precipitation and TWI.
Figure 9. SHAP dependence plots for (a) slope and TWI; (b) elevation and slope; (c) TWI and slope; (d) precipitation and TWI.
Atmosphere 15 00788 g009
Atmosphere 15 00788 g010
Figure 10. Interpretation of single samples based on SHAP: (a) thermokarst hazard and (b) non-thermokarst hazard.
Figure 10. Interpretation of single samples based on SHAP: (a) thermokarst hazard and (b) non-thermokarst hazard.
Atmosphere 15 00788 g010
Atmosphere 15 00788 g011
Figure 11. Accumulated local effects for the four dominant factors: (a) slope, (b) elevation, (c) TWI, and (d) precipitation.
Figure 11. Accumulated local effects for the four dominant factors: (a) slope, (b) elevation, (c) TWI, and (d) precipitation.
Atmosphere 15 00788 g011
Atmosphere 15 00788 g012
Figure 12. Interpretation of single samples based on LIME: (a) thermokarst hazard and (b) non-thermokarst hazard.
Figure 12. Interpretation of single samples based on LIME: (a) thermokarst hazard and (b) non-thermokarst hazard.
Atmosphere 15 00788 g012
Atmosphere 15 00788 g013
Figure 13. Potential risk evaluation based on stacking model: (a) main infrastructures; (b) the SOC density above 3 m; and (c) vegetation types.
Figure 13. Potential risk evaluation based on stacking model: (a) main infrastructures; (b) the SOC density above 3 m; and (c) vegetation types.
Atmosphere 15 00788 g013
Atmosphere 15 00788 g014
Figure 14. The potential risk evaluation based on the stacking model: (a) thermokarst hazard susceptibility evaluation in the QTP; (b) thermokarst hazard susceptibility evaluation in the QTEC; (c) distribution of susceptibility classes along the QTH in the QTEC; (d) distribution of susceptibility classes along the QTR in the QTEC (VL: very low, L: low, M: moderate, H: high, VH: very high, NP: non-permafrost).
Figure 14. The potential risk evaluation based on the stacking model: (a) thermokarst hazard susceptibility evaluation in the QTP; (b) thermokarst hazard susceptibility evaluation in the QTEC; (c) distribution of susceptibility classes along the QTH in the QTEC; (d) distribution of susceptibility classes along the QTR in the QTEC (VL: very low, L: low, M: moderate, H: high, VH: very high, NP: non-permafrost).
Atmosphere 15 00788 g014
Table 1. Conditioning factors and the results of multicollinearity analysis.
Table 1. Conditioning factors and the results of multicollinearity analysis.
Conditioning FactorsCollinearity Statistics
ToleranceVIF
ALT0.2813.553
Aspect0.9941.006
Elevation0.6141.630
NDVI0.3362.975
Precipitation0.3093.237
Slope0.4982.008
FSC0.4472.240
Solar radiation0.4652.150
TWI0.5041.986
MAGT0.3872.582
Table 2. Model evaluation indexes.
Table 2. Model evaluation indexes.
ModelsAUCAccuracyRecallPrecisionF1-Score
Stacking0.93220.86270.90400.83340.8673
CatBoost0.93160.86050.89320.83670.8641
RF0.92860.85680.87810.84060.8589
XGBoost0.92760.85630.88450.83570.8594
EXT0.92750.86000.88720.83980.8628
GBDT0.91950.84480.87510.82340.8484
Table 3. Characteristics of different susceptibility classes based on the stacking model.
Table 3. Characteristics of different susceptibility classes based on the stacking model.
ClassesArea Covered (%)Thermokarst Hazard Covered (%)
Very Low51.250.96
Low20.843.66
Moderate7.824.18
High10.4914.36
Very High9.5976.83
Table 4. Characteristics of different susceptibility classes based on the CatBoost model.
Table 4. Characteristics of different susceptibility classes based on the CatBoost model.
ClassesArea Covered (%)Thermokarst Hazard Covered (%)
Very Low59.732.10
Low13.793.74
Moderate8.786.49
High8.1314.53
Very High9.5773.14
Table 5. Characteristics of different susceptibility classes based on the RF model.
Table 5. Characteristics of different susceptibility classes based on the RF model.
ClassesArea Covered (%)Thermokarst Hazard Covered (%)
Very Low48.621.08
Low19.943.24
Moderate12.956.71
High10.2916.62
Very High8.1972.35
Table 6. Distribution characteristics of main infrastructures in stacking model map.
Table 6. Distribution characteristics of main infrastructures in stacking model map.
ClassesBuildingRailway StationRailway (km)Highway (km)
Very Low46165.84782.30
Low19379.62487.48
Moderate7260.77206.04
High75141.11399.65
Very High88247.01624.57
Table 7. The distribution characteristics of the SOC storage in the stacking model map.
Table 7. The distribution characteristics of the SOC storage in the stacking model map.
ClassesArea Covered (%)Average SOC Density (kg/m2)SOC Storage (Pg C)
Very Low51.2513.446.85
Low20.8413.802.86
Moderate7.8213.241.03
High10.4912.691.32
Very High9.5912.641.21
Table 8. The distribution characteristics of vegetation in high and very high susceptibility areas.
Table 8. The distribution characteristics of vegetation in high and very high susceptibility areas.
Vegetation TypesAlpine GrasslandAlpine MeadowAlpine VegetationAlpine Desert
Area (km2)115,45765,3978857772
Table 9. The distribution characteristics of the QTH and QTR in the QTEC in the stacking model map.
Table 9. The distribution characteristics of the QTH and QTR in the QTEC in the stacking model map.
Length (km)Susceptible Class
Very LowLowModerateHighVery High
QTH70.7159.9539.25113.70222.45
QTR40.9367.0552.15122.83222.16
Table 10. Susceptibility map accuracy assessment based on training data and additional data.
Table 10. Susceptibility map accuracy assessment based on training data and additional data.
ClassesNumber of
Training Data Points		Training Data Covered (%)Number of Additional Data PointsAdditional Data Covered (%)
Very Low002740.96
Low0010923.81
Moderate90.099673.38
High6176.17349612.12
Very High937493.7422,82379.73
Table 11. The results of the 10-fold cross-validation model.
Table 11. The results of the 10-fold cross-validation model.
ModelsAverage Accuracy (%)Maximum Accuracy (%)
CatBoost86.0987.10
EXT85.8087.45
RF85.6187.70
XGBoost85.3386.85
GBDT84.2085.30
AdaBoost83.1584.65
LR80.9582.50
Bayes79.5880.95
CART79.4480.90
KNN74.0975.20
SVM68.1570.25
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Yang, Y.; Wang, J.; Mao, X.; Lu, W.; Wang, R.; Zheng, H. Susceptibility Modeling and Potential Risk Analysis of Thermokarst Hazard in Qinghai–Tibet Plateau Permafrost Landscapes Using a New Interpretable Ensemble Learning Method. Atmosphere 2024, 15, 788. https://doi.org/10.3390/atmos15070788

AMA Style

Yang Y, Wang J, Mao X, Lu W, Wang R, Zheng H. Susceptibility Modeling and Potential Risk Analysis of Thermokarst Hazard in Qinghai–Tibet Plateau Permafrost Landscapes Using a New Interpretable Ensemble Learning Method. Atmosphere. 2024; 15(7):788. https://doi.org/10.3390/atmos15070788

Chicago/Turabian Style

Yang, Yuting, Jizhou Wang, Xi Mao, Wenjuan Lu, Rui Wang, and Hao Zheng. 2024. "Susceptibility Modeling and Potential Risk Analysis of Thermokarst Hazard in Qinghai–Tibet Plateau Permafrost Landscapes Using a New Interpretable Ensemble Learning Method" Atmosphere 15, no. 7: 788. https://doi.org/10.3390/atmos15070788

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop