Article 1

Prediction of Potential Areas of Geothermal Disaster along the 2

Yunnan–Tibet Railway Project based on Niche Models 3

Zhe Chen 1,3,6, Ruichun Chang 1,6,7,* , Huadong Guo 2,4,6,8, Xiangjun Pei 3,5,6, Wenbo Zhao 4,8, Zhengbo Yu 1,6,7 and Lu 4
Zou 1,6,7 5

1 College of Mathematics and Physics, Chengdu University of Technology, Chengdu 610059, China; 6
chenz115@stu.cdut.edu.cn (Z.C.); 2021021076@stu.cdut.edu.cn (Z.Y.); lucille@stu.cdut.edu.cn (L.Z.) 7
2 International Research Centre of Big Data for Sustainable Development Goals (CBAS), Beijing 100094, 8
China; hdguo@radi.ac.cn (H.G.) 9
3 State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of 10
Technology, Chengdu 610059, China; dclixa@chd.edu.cn (X.P.) 11
4 Key Laboratory of Digital Earth Science, Aerospace Information Research Institute, Chinese Academy of 12
Sciences, Beijing 100094, China; zhaowb@aircas.ac.cn (W.Z.) 13
5 College of Ecological Environment, Chengdu University of Technology, Chengdu 610059, China; 14
6 Digital Hu Line Research Institute, Chengdu University of Technology, Chengdu 610059, China; 15
7 Geomathematics Key Laboratory of Sichuan Province, Chengdu University of Technology, Chengdu 16
610059, China; 17
8 University of Chinese Academy of Sciences, Beijing 100049, China 18
* Correspondence: changruichun08@cdut.edu.cn 19
Citation: Lastname, F.; Lastname,
F.; Lastname, F. Title. Remote Sens.
Abstract: With the continuous advancement of China's railway engineering construction to the 20
2021, 13, x.
Qinghai-Tibet Plateau, the number of deep-buried long tunnels increases daily. Tunnel geothermal 21
https://doi.org/10.3390/xxxxx
disaster has become a common geological disaster in underground engineering. Predicting the po- 22

Academic Editors: Nicola Casagli tential areas of geothermal disaster along the Yunnan-Tibet railway project is conducive to the plan- 23
ning and construction of railway projects and the realization of the industry, innovation and infra- 24
Received: date structure goal (SDG 9) of the United Nations Sustainable Development Goals (SDGs). Taking the 25
Accepted: date Yunnan-Tibet railway project as the study area, based on Landsat-8 image and other spatial data, 26
Published: date focusing on the causes and distribution law of geothermal disaster, the collinearity diagnosis of 27
environmental variables is carried out. 12 environmental variables such as land surface temperature 28
Publisher’s Note: MDPI stays neu- are selected to predict the potential area of geothermal disaster through four niche models (MaxEnt, 29
tral with regard to jurisdictional
Bioclim, Domain and GARP), and the prediction results are divided into four levels. The research 30
claims in published maps and insti-
shows that the prediction results of the four models have different characteristics. Among them, the 31
tutional affiliations.
Area under receiver operating characteristic curve (AUC) and kappa values of the MaxEnt model 32
are the highest, which are 0.84 and 0.63 respectively. The prediction accuracy is the highest, and the 33
algorithm results are more suitable for the prediction of geothermal disasters. The prediction results 34
Copyright: © 2021 by the authors. show that the geothermal disaster in Markam-Deqen, Zuogong-Zayu and Baxoi-Zayu along the 35
Submitted for possible open access Yunnan-Tibet railway project is highly potential. Through the analysis of jackknife, it is concluded 36
publication under the terms and that the land surface temperature, active faults, water system distribution and Moho depth are the 37
conditions of the Creative Commons key environmental variables to predict the potential area of geothermal disaster. The research re- 38
Attribution (CC BY) license sults can provide a reference for the design and construction of the Yunnan-Tibet railway project 39
(http://creativecommons.org/li- and sustainable development. 40
censes/by/4.0/).

Keywords: Yunnan-Tibet railway; geothermal disaster; sustainable development; Landsat-8; 41

niche model 42
43

1. Introduction 44
China's railway construction continues to advance to the Qinghai-Tibet Plateau, 45
which is of great significance to promoting economic development, national unity and 46

Remote Sens. 2021, 13, x. https://doi.org/10.3390/xxxxx www.mdpi.com/journal/remotesensing

Remote Sens. 2021, 13, x FOR PEER REVIEW 2 of 24

strengthening national defense construction along the line [1]. Tunnel geothermal disaster 47
is a common geological disaster in deep and long tunnels. It will not only lead to the de- 48
terioration of the construction environment, threaten the health and safety of constructors 49
and the efficiency and progress of tunnel engineering, but also be detrimental to the safety 50
and durability of lining structure and affect the stability and safety after railway operation 51
[2]. Yunnan-Tibet railway leads northward from Shangri-La station, crosses Jinsha River, 52
Baima Snow Mountain to Deqen County, then crosses Meili Snow Mountain and Yuqu in 53
the northwest, crosses Nujiang River in the south of Gura Township, then crosses boshula 54
ridge in the northwest to Rawu East, passes along the north side of Rawu lake, and then 55
flows downstream along Parlung Zangbo, passes Yupu, Songzong and Bome county to 56
connect with the Sichuan-Tibet railway under construction, with a total length of about 57
490km [1]. The line crosses the famous "Three Parallel Rivers" area of the Hengduan 58
Mountains. The terrain and geological conditions of the whole line are extremely complex, 59
the ecological environment is fragile, and the project is huge and arduous. It can be called 60
"the most difficult and dangerous railway project in the world" (Figure 1). Restricted by 61
the topographic elevation difference and geomorphic pattern, the Yunnan-Tibet railway 62
mainly passes through the mountain canyon area in the form of a tunnel. On the one hand, 63
the construction of a deep-buried long tunnel will greatly shorten the traffic distance and 64
obtain a large drop in a short distance, but it will inevitably bring a series of geological 65
disasters, especially the problem of geothermal disaster [3]. Achieving the United Nations 66
Sustainable Development Goals (SDGs) is one of the goals shared by governments and 67
humanity. The Industry, Innovation and Infrastructure Goals (SDG 9) emphasize building 68
disaster-resistant infrastructure; The sustainable cities and communities goal (SDG 11) 69
calls for building inclusive, safe, disaster-resilient and sustainable cities and human set- 70
tlements. Therefore, infrastructure construction such as railways plays a vital role in 71
achieving the goals of the sustainable development agenda [4,5]. 72

Location Topography

Figure 1. Location and topography along the Yunnan-Tibet railway. 73

At present, there are many studies on the geothermal disaster of plateau linear engi- 74
neering such as the Yunnan-Tibet railway. For example, Yan Jian studied the characteris- 75
tics of high ground temperature in the Sangzhuling tunnel of the Sichuan-Tibet Railway 76
in 2019, and analyzed the characteristics of high ground temperature and its impact on 77
tunnel engineering [6]; Luo Feng analyzed and predicted the ground temperature charac- 78
teristics of an underground project in the north wall of Lhari fault on the Qinghai-Tibet 79
Plateau in 2021, and summarized the relationship between tunnel ground temperature 80
characteristics and Lhari fault [7]. In addition, the research on the influencing factors of 81
geothermal disaster shows that the characteristics of ground temperature and ground 82
temperature gradient are closely related to the thermal process in the earth, tectonism, 83
Remote Sens. 2021, 13, x FOR PEER REVIEW 3 of 24

climate change and formation lithology [8–11]. However, most of the above studies focus 84
on the influencing factors of ground temperature distribution and the impact of the geo- 85
thermal disaster on the project, while there are few studies on the prediction of potential 86
areas of geothermal disaster. Relying on the GIS platform and spatial information tech- 87
nology, prediction models such as evidence weight method, deterministic coefficient 88
method, maximum entropy and random forest have been widely used in many fields such 89
as metallogenic prediction, landslide disaster prediction, species suitable area prediction 90
and geothermal resources prediction and evaluation [12–15]. For example, Zhang Shuai 91
Used a maximum entropy and random forest model to predict the potential of gold de- 92
posits in the West Qinling region of China in 2018, and defined the prospecting target area 93
[16]; In 2021, Zhao Wenbo used index coverage, entropy weight and evidence weight in- 94
formation model to predict the potential geothermal anomaly areas in eastern Tibet and 95
Western Sichuan, and delineated the geothermal anomaly areas along the Sichuan-Tibet 96
railway. [17]. The niche model was initially applied to species prediction and analysis. In 97
recent years, it has been widely used in predicting the distribution range and changes of 98
disasters such as landslides, debris flow and floods [16]. Using the known distribution 99
data of disasters and disaster-causing environmental variables, the model is constructed 100
according to certain algorithm operations, and the operation results are projected into dif- 101
ferent times and spaces to predict the potential prone areas of disasters [18]. When the 102
niche model is applied to the prediction of potential areas of geothermal disaster, the 103
known geothermal spots in the study area are equivalent to the known distribution data 104
of disasters, and the corresponding environmental variables are the structural prediction 105
variables. At present, there are many commonly used niche models. Each model can in- 106
dependently predict potential prone areas, but each model has a certain preference 107
(Araujo and New, 2007). Using the idea of an ensemble prediction system and integrating 108
the prediction results of various models, the false negative or false positive impact caused 109
by empirically selecting one model can be reduced as much as possible. At the same time, 110
the defects of one model may be made up by another model to improve the scientificity 111
of prediction [20,21]. 112
Based on niche models with four different algorithms (MaxEnt, Bioclim, Domain and 113
GARP), taking 15 disaster causing factors as environmental variables and combined with 114
the ArcGIS platform, this study predicts the potential areas of geothermal disaster along 115
the proposed Yunnan-Tibet railway project, compares the prediction results of different 116
models, analyzes the relationship between geothermal disaster and environmental varia- 117
bles, and reveals the dominant factors affecting the potential distribution of geothermal 118
disaster. It can provide scientific basis and technical support for the engineering planning, 119
route optimization and sustainable development of the Yunnan-Tibet railway, and can 120
provide a reference for the survey, design and engineering construction of complex and 121
dangerous mountainous railways such as Sichuan-Tibet, China-Nepal and Xinjiang-Tibet. 122

2. Materials and Methods 123

2.1. Study Area 124
The area along the Yunnan-Tibet railway is located between 95 ° 15'-100 ° 15'E and 125
27 ° 25'-30 ° 05'N in Southwest China. It mainly includes the southeast of Tibet Tibetan 126
Autonomous Region and the north of Yunnan Province, with a total area of about 127
143009km2 (Figure 1). It is dominated by a plateau humid monsoon climate, with an an- 128
nual average temperature of 6.3 ~ 12.5 ℃, a large daily temperature difference and a min- 129
imum temperature of - 27.4 ℃; the whole year is dry and wet, and the rainfall is mainly 130
concentrated from May to September, with an average annual rainfall of 449.19 ~ 726.4 131
mm. The line passes through four geomorphic units: Zhongdian fault depression basin 132
area, Hengduan Mountain high mountain canyon area, Southeast Tibet high mountain 133
deep canyon area and Parlung Zangbo River Valley area. The trend of the mountain range 134
Remote Sens. 2021, 13, x FOR PEER REVIEW 4 of 24

is consistent with that of the regional structural belt. The landform is alternating with gul- 135
lies and mountains, and the terrain is generally high in the northwest and low in the south- 136
east. The elevation of the river valley is 2700 ~ 3300m, the elevation of the mountains is 137
3000 ~ 6700m, and the relative height difference is 1000 ~ 3000m. The canyon area is mostly 138
in a "V" shape, with a cutting depth of 1000 ~ 2000m and a bank slope gradient of 30 ° ~ 70 139
° on both sides. Meili Snow Mountain passed by the line is the extreme peak of the whole 140
line, with an altitude of about 6740m [1]. The basic topographic conditions are shown in 141
Figure 1. 142
The eastern part of the study area is the Deqen Weixi stratigraphic division of the 143
Qiangtang Qamdo stratigraphic area, and the central and western part is the Lhorong 144
stratigraphic division and Lhasa Bome stratigraphic division of Gangdise Nianqing 145
Tanggula stratigraphic area. The strata are exposed from Quaternary to Precambrian, and 146
the distribution of strata is mainly controlled by structure. The Quaternary strata are 147
mainly sandy soil, gravel soil and cohesive soil; the underlying sedimentary rocks, meta- 148
morphic rocks and intrusive rocks are distributed alternately, and the lithology is mixed 149
and changeable [22]. Affected by the strong collision and compression of plates and the 150
uplift of Qinghai-Tibet Plateau, the fault zones in the region are extremely developed, 151
including the Jinsha River fault zone, Lancang River fault zone, Bangong Lake Nujiang 152
suture zone, Jiali fault zone and Yarlung Zangbo River suture zone. From west to east, the 153
Yunnan-Tibet railway passes through 6 faults including the Deqin Zhongdian fault, 154
Jinshajiang fault, Lancang River fault, Basu fault, Nujiang fault and Lhari fault respec- 155
tively. It is characterized by large scale, fault bandwidth, strong tectonic environment, 156
complex structure and strong activities since the Holocene. It is the seismogenic structure 157
of medium-strong earthquakes and mega earthquakes in the region, with the risk of 158
strong earthquakes, and the engineering geological conditions are extremely complex [23] 159
(Figure 2). The engineering problems caused by active faults mainly include catastrophic 160
structural creep and dislocation, induced earthquakes, resulting in secondary disasters 161
such as slope instability and destructive debris flow, and the harm of high-temperature 162
hot springs distributed along the fault zone. 163

Figure 2. Overview of the study area. 164

Remote Sens. 2021, 13, x FOR PEER REVIEW 5 of 24

The line runs in the Yunnan-Tibet geothermal zone in the great bend geothermal ac- 165
tivity area of Yarlung Zangbo and belongs to the Himalayan geothermal zone with the 166
strongest geothermal activity in mainland China. Geothermal is mainly distributed in 167
beads along active fault zones or exposed in fault basins and fault valleys. The geothermal 168
outcrops along the line are mainly temperature springs and hot springs with low salinity. 169
The maximum water temperature exceeds 80 ℃, the water temperature is relatively sta- 170
ble, and is less affected by the climate. It is an extremely high-temperature area [24] (Fig- 171
ure 2). The line crosses many seismic zones such as Xianshuihe East Yunnan seismic zone, 172
Southwest Yunnan seismic zone, central Tibet seismic zone and Himalayan seismic zone. 173
It is the largest seismic area in China and belongs to an area with extremely active seismic 174
activity and frequent high-intensity earthquakes [25] (Figure 2). 175
The distribution of the river water system along the line is controlled by the structure, 176
and the flow direction of the water system is consistent with the structural trend (Figure 177
2). The line crosses the Jinsha River, Lancang River, Nujiang River, Parlung Zangbo River 178
and other rivers from east to west. Its common feature is that it has abundant water and 179
is affected by the seasons. The runoff in summer is the largest, the river slope is large, and 180
the water flow is turbulent. It is mainly recharged by precipitation, groundwater and ice 181
meltwater [1]. 182
The shallow geothermal energy resources in the area are abundant. Its generation is 183
closely related to the strong solar radiation, complex geological and geomorphic condi- 184
tions and extreme natural climate in the area. At the same time, it is also restricted by 185
geological conditions such as stratigraphic structure, structure, groundwater distribution 186
and rock thermophysical properties [26–28]. Under certain circumstances, with the influ- 187
ence of deep heat sources and fractures, the places with abundant shallow ground tem- 188
perature energy can also be transformed into the geothermal disasters of the tunnel [11]. 189

2.2. Materials 190

2.2.1. Geothermal Sample Points 191
The Yunnan-Tibet tropical zone is located at the collision junction of the Indian Ocean 192
and the Eurasian plate. It has the characteristics of long extension distance, grand scale 193
and the strongest geothermal activity. Its hot springs account for more than half of the 194
total hot springs in China. The distribution of abnormal areas of geothermal activity along 195
the line is related to the activity of active faults, and the water temperature is mainly a hot 196
spring. The hot springs exposed on the surface can be regarded as the ground characteri- 197
zation of geothermal disasters. Within an area of 10km2, they may belong to the same 198
geothermal system, and high-temperature hot springs are likely to be concentrated. There- 199
fore, the hot spring points exposed on the surface in the study area are taken as the known 200
geothermal sample points of the prediction model[12]. The data of geothermal sample 201
points comes from the 1:4 million distribution map of geothermal resources in China[29]. 202
After statistics, 108 representative geothermal sample points are finally selected (Figure 203
2). Among them, 75% of the sample points are randomly selected as the training set, and 204
the remaining 25% of the sample points and the background points 10 times the total sam- 205
ple points are randomly selected as the test set. To compare the differences in the pre- 206
dicted results of the four models, 10 groups of training data sets and corresponding test 207
data sets are randomly generated. The training set is used for model prediction and the 208
test set is used for model verification. 209

2.2.2. Selection and Processing of Environment Variables 210

Combined with the geological structure, geophysics, natural climate, landform and 211
hydrogeological conditions of the study area[12], 15 factors are selected as environmental 212
variables: land surface temperature, buffer distance to fault, fault density, combined en- 213
tropy of geological formation, earthquake peak acceleration, epicentral nucleus density, 214
aeromagnetic anomaly, Bouguer gravity anomaly, Moho depth, terrestrial heat flow, near- 215
Remote Sens. 2021, 13, x FOR PEER REVIEW 6 of 24

surface temperature, snow depth, the degree of permafrost, amount of precipitation and 216
buffer distance to river (Table 1). 217

Table 1. Environment variables. 218

Code Environment variable Unit

Bio1 Land surface temperature ℃
Bio2 Buffer distance to fault km
Bio3 Fault density km/km²
Bio4 Combined entropy of geological formation —
Bio5 Earthquake peak acceleration g
Bio6 Epicentral nucleus density —
Bio7 Aeromagnetic anomaly nT
Bio8 Bouguer gravity anomaly mgal
Bio9 Moho depth km
Bio10 Terrestrial heat flow mW/m²
Bio11 Near-surface temperature ℃
Bio12 Snow depth m
Bio13 The degree of permafrost —
Bio14 Amount of precipitation mm
Bio15 Buffer distance to river km

The land surface temperature (LST) can provide high-quality and high-efficiency 219
heat information of the land acquisition surface, and provide a basis for the prediction of 220
geothermal disasters. The terrain of the study area is complex and covers a large area. One 221
of the most effective methods is to obtain the surface temperature through thermal infra- 222
red remote sensing inversion [12]. The Landsat-8 satellite of the United States has a ther- 223
mal infrared sensor with 100 meter spatial resolution, which can extract the land surface 224
temperature anomaly in the study area with high quality and efficiency. Thermal infrared 225
data are from the U.S. Geological Survey (https://earthexplorer.usgs.gov/)[12]. Finally, the 226
single window algorithm is used to retrieve the surface temperature, and the formula is: 227

Ts = {a(1-Ci-Di) + [b(1-Ci-Di)+Ci+Di]Tb-DiTa}/Ci, (1)

Ci = εi τi , (2)

Di = (1-τi )[1+τi (1-εi )], (3)

where a and b are the linear regression coefficients, which are related to the temperature 228
range of the study area; C and D are intermediate variables; Ta is the average value at 229
atmospheric temperature (k); Tb is the brightness temperature (k) obtained by the sensor; 230
Ts is LST; τ is the atmospheric transmittance; and ε is the surface emissivity[12]. 231
In order to reduce the influence of complex terrain on LST inversion in the study area, 232
terrain correction needs to be added. The empirical statistical method is selected in this 233
study, and the formula is: 234

cos(i) = cos(z)cos(S)+sin(z)sin(S)cos(Φx -Φn ), (4)

LT = mcos(i) + b, (5)

LH = LT-[mcos(i)+ b]+ LT, (6)

Remote Sens. 2021, 13, x FOR PEER REVIEW 7 of 24

Where S is the tilt angle of the pixel; i is the effective incidence angle of the sun; Φn is the 235
tilt angle of the pixel; Φx is the azimuth of the sun; z is the zenith angle of the sun; LT is 236
the radiation value of the ground object before correction; LH is the corrected radiation 237
value of the ground feature; LT is the theoretical radiation value of ground objects in a 238
flat area without topographic relief; b and m are the parameters obtained by regression 239
analysis[12]. 240
In this study, a total of 90 long-time series Landsat-8 data images from 2013 to 2021 241
were downloaded, and the multi-year average winter LST was calculated. 242
In terms of geological structure, geothermal disasters are prone to occur in areas with 243
complex fault structures and high degree of rock fragmentation [12,29]. The fault zone 244
forms a channel to the underground heat source, so that the groundwater can be heated 245
and gushed back to the surface, causing geothermal disasters [12,30]. Through the pro- 246
cessing of fault data on ArcGIS platform, the fault buffer distance and fault line density 247
can be obtained. The fault data are from the 1:5 million structural map of China and its 248
adjacent areas. The formation combination entropy can be calculated from the formation 249
lithology data. It is a basic form of geological anomaly [17], which represents the entropy 250
anomaly of different properties of the same geological body or the combination of differ- 251
ent geological bodies in a volume or unit area [12,31]. The formation lithology data comes 252
from the 1:4 million geological map of China. 253
In terms of Geophysics, epicenter density and ground motion peak acceleration can 254
measure the level of seismic activity and hydrothermal activity in the study area to a cer- 255
tain extent [12,32]. The seismic peak acceleration and epicenter data are from the 1:5 mil- 256
lion seismic peak acceleration zoning map of China and the seismic epicenter distribution 257
map of China, respectively. Aeromagnetic anomaly distribution is often used to describe 258
areas with large changes in tectonic load in groundwater thermal activity areas [12,33]. 259
Aeromagnetic anomaly data are from the 1:6 million aeromagnetic anomaly map of China 260
and adjacent sea areas. Gravity anomaly is caused by the uneven distribution of under- 261
ground rock mass and mineral density, or the density difference between geological body 262
and surrounding rock. Bouguer gravity anomaly map can be used to understand regional 263
structures, delineate large fault structures and local anomalies that may be related to ge- 264
othermal system [12,34]. Bouguer gravity anomaly data comes from 1:4 million Bouguer 265
gravity anomaly map of China. Moho depth distribution characteristics are of great sig- 266
nificance for lithospheric structure, crust mantle tectonic evolution and geothermal distri- 267
bution[35]. Moho depth data are from Moho depth map of China Sea and land. Terrestrial 268
heat flow refers to the heat transmitted from the earth's interior to the earth's surface in 269
the form of heat conduction per unit time and unit area, and then emitted into space. Ac- 270
cording to the definition, the earth heat flow itself contains the thermal information of the 271
earth from deep to shallow[35]. The terrestrial heat flow data is from the compilation of 272
terrestrial heat flow data in Chinese Mainland (Fourth Edition). The above geological and 273
geophysical data are from the geological cloud of China Geological Survey (https://geo- 274
cloud.cgs.gov.cn/). 275
In terms of natural climate and hydrology, the factors of near-ground temperature, 276
snow depth and frozen soil distribution are the material basis for the formation of shallow 277
ground temperature fields, which together constitute thermal energy storage sites and 278
transportation channels [26]. Among them, the average annual near-surface temperature 279
data comes from the China Ecological Science Data Center（http://www.nesdc.org.cn/） 280
Data set of near-surface temperature in China from 1979 to 2018 [36]. The data on perma- 281
frost and average annual snow depth are from the China Qinghai-Tibet Plateau Scientific 282
Data Center（http://data.tpdc.ac.cn/）Newly drawn frozen soil distribution map of Qing- 283
hai-Tibet Plateau [37] and data set of 0.05 °daily snow depth of Qinghai-Tibet Plateau from 284
2000 to 2018 [38]. The study area is rich in precipitation and numerous rivers, which can 285
provide a water source for the formation of geothermal anomalies. The average annual 286
precipitation data are from China Ecological Science Data Center （ 287
http://www.nesdc.org.cn/）Spatial interpolation data set of fine grid meteorological data 288
Remote Sens. 2021, 13, x FOR PEER REVIEW 8 of 24

every 8 days for 1km in China from 2000 to 2018 [39]. The river buffer distance is obtained 289
by processing the hydrological data. The hydrology comes from the geological cloud of 290
the China Geological Survey (https://geocloud.cgs.gov.cn/) 1:3 million National hydroge- 291
ological Atlantes. 292

2.2.3. Collinearity Diagnosis of Environmental Variables 293

To avoid the overfitting of the model prediction results caused by the culinary be- 294
tween environmental factors [40], this study will first input all variables into the MaxEnt 295
model to obtain the preliminary model simulation results and the contribution of each 296
factor, and then conduct Spearman rank correlation analysis on environmental factors 297
[41], screen out the environmental variables with low correlation (Spearman coefficient < 298
0.75), combined with the contribution of each factor in the simulation results of all factors 299
And the environmental conditions of geothermal samples used in this paper, excluding 300
the factors with high correlation, low contribution and low impact [42] (Figure 3), and 301
finally 12 variables are obtained (Figure 4). 302

Figure 3. Heat map of correlation coefficient of the environment variable.

Remote Sens. 2021, 13, x FOR PEER REVIEW 9 of 24

Bio1 Bio2

Bio4 Bio4

Bio5 Bio6
Remote Sens. 2021, 13, x FOR PEER REVIEW 10 of 24

Bio9 Bio10

Bio11 Bio12

Bio13 Bio15

Figure 4. Environment variables.

2.3. Methods 303

2.3.1. Prediction Models 304
In recent years, the maximum entropy (MaxEnt) model has attracted extensive atten- 305
tion in the field of machine learning [16]. It simulates the geographical distribution of spe- 306
cies based on the correlation between environmental variables with spatial location and 307
the distribution of target species [43]. The maximum entropy theory holds that on the 308
premise of meeting the existing conditions, the real state of the research object should be 309
the state when the system entropy is maximum, that is, when predicting the unknown 310
Remote Sens. 2021, 13, x FOR PEER REVIEW 11 of 24

distribution according to the known species samples and meeting the constraints of the 311
known species samples, the real distribution of the target species in the study area is when 312
the unknown distribution entropy is maximum, to obtain the ecological environment dis- 313
tribution or climate suitability of the target species [44]. When the maximum entropy 314
model is applied to the prediction of geothermal disaster, the known geothermal hot 315
spring points in the study area are equivalent to the records of species, and the corre- 316
sponding influencing factors are the environmental variables of the model. The calcula- 317
tion formula is: 318

̂) = − ∑x∈X   π
H(π ̂(x) ⋅ ln π
̂(x), (7)
The unknown probability distribution is defined as π. For the finite element (X) in the 319
study area, each element point (x) is assigned a non-negative probability value, and the 320
sum of the probability values is 1; The estimated probability distribution is 𝜋̂, Then H (𝜋̂) 321
Is entropy. 322
In this study, the MaxEnt model is used to simulate the potential distribution of geo- 323
thermal disasters. Import 10 groups of training data sets and environmental variable data 324
sets into the model respectively, set the maximum number of iterations to 500 times and 325
the maximum number of background points to 10000, and test the weight by the jackknife 326
method. 327
The Bioclim model is a framework model that extracts the limited range of environ- 328
mental factors from known species distribution areas, and then summarizes the environ- 329
mental needs of species into the environmental envelope [45]. By studying the climate 330
parameters of the known distribution areas of species, the Bioclim model summarizes var- 331
ious ecological characteristics of species into the rectangular environmental envelope. Fi- 332
nally, many single envelopes will form a group of the environmental envelope system, 333
and then project this envelope system to the prediction target area, and compare each 334
climate variable in the target area with the environmental envelope system. If the position 335
of a point in the envelope system space is within the environmental envelope, the model 336
determines that the point is a potential distribution point [46]. The more the number of 337
species distribution points, the higher the accuracy of the model in predicting the suitable 338
habitat of species. However, due to the characteristics of the prediction process of the Bi- 339
oclim model, when the species distribution presents a discrete trend, the prediction results 340
will have errors, which will often enlarge the suitable habitat of species. 341
The Domain model uses the environmental similarity matrix between the points for 342
simulation and prediction [47], and its core idea is the Gower distance from a point to 343
point, which is the distance between two points in Euclidean geometric space. Then it was 344
introduced into applied ecology to judge the similarity between the target area and the 345
clear distribution area of the species. The Gower distance is treated by variance standard- 346
ization or range standardization to ensure that the contribution of this value in each di- 347
mension is the same. In the case of different sampling, variance standardization can better 348
avoid the error caused by it [48]. It only uses the existing data of species distribution, and 349
has a good effect when there are few prediction variables. 350
Based on the DIVA-GIS platform, this study simulates the potential distribution of 351
geothermal disaster, imports 10 groups of training data sets and environmental variable 352
data sets into Bioclim and Domain models respectively in Modeling, and then forecasts 353
them in turn [48]. 354
With the development of machine learning, the genetic algorithm for the rule set pro- 355
duction (GARP) model has been formally proposed. The operation principle of GARP is 356
based on a regular genetic algorithm. During the operation of the model, the environmen- 357
tal factors affecting species distribution are automatically selected, and all the selected 358
factors are mapped to a certain area. Through superposition analysis, the nonrandom re- 359
lationship between the known distribution of target species and the target research area 360
is explored, the basic niche of the research target is determined, and finally the potential 361
geographical distribution area of the research target is simulated and predicted [49,50]. In 362
Remote Sens. 2021, 13, x FOR PEER REVIEW 12 of 24

general, it is based on the rule combination of genetic algorithm to model the local envi- 363
ronment space and predict the species distribution [51], and can well predict the suitable 364
habitat of discrete species, which has advantages in this regard. 365
This study simulates the potential distribution of geothermal disasters based on the 366
Desktop GARP platform. Import 10 groups of training data sets and environmental vari- 367
able data sets into the model respectively, then randomly create 10 repetitions, set the 368
convergence limit to 0.01 and the maximum number of iterations to 1000, and finally add 369
and stack them in ArcGIS to obtain the predicted distribution map of geothermal disaster. 370

2.3.2. Test of Model Prediction Results 371

Using 10 groups of test sets obtained in the random segmentation step of the predic- 372
tion link, the receiver operating characteristic (ROC) and kappa coefficient are calculated 373
to verify the prediction accuracy of the model. 374
The area under the receiving operator curve (AUC) is the area covered by the ROC 375
curve. Based on the AUC value, it is not affected by the diagnostic threshold and is not 376
sensitive to the incidence. At present, it is recognized as the best evaluation index and has 377
been widely used in the accuracy test and evaluation of niche models. The numerical 378
range of the AUC value distribution is 0 ~ 1. The closer it is to 1, the greater the correlation 379
between environmental variables and the distribution of simulation objects, that is, the 380
better the simulation results. AUC value is considered to be "more accurate" between 0.7 381
~ 0.8, "very accurate" between 0.8 ~ 0.9, and "very accurate" beyond 0.9 [52]. 382
Kappa coefficient is usually used to measure the consistency between the simulation 383
results and the real situation [53,54]. The distribution interval the of kappa coefficient is 384
usually 0 ~ 1. The closer it is to 1, the more consistent the simulation results are with the 385
real situation, that is, the more realistic the simulation results are. The Kappa coefficient 386
is considered "moderate" at 0.4 ~ 0.6, and "significant" when the kappa value is greater 387
than 0.6. The larger the value, the higher the prediction accuracy [55–57]. " 388

3. Result 389
3.1. Predicted Results of Models 390
From the 10 groups of prediction maps of each model, the map with the largest AUC 391
is selected as the base map, and the potential areas of geothermal disaster are classified by 392
using the natural discontinuity classification method (Jenks), to obtain the potential area 393
distribution map of each model, as shown in Figure 5. The high potential area of geother- 394
mal disaster in the figure indicates that there is a high possibility of geothermal disaster 395
in the area; a Low potential area means that there is a low possibility of geothermal disas- 396
ter in the area; a Non-potential area means that there is no geothermal disaster in the area. 397
From the classification chart of prediction potential and statistical results, it can be seen 398
that the distribution trend of the overall prediction results of the four models is relatively 399
close, and the area size and specific details are different. The predicted distribution map 400
of the MaxEnt model (Figure 5) shows that the area with a high potential for geothermal 401
disaster is the smallest, mainly distributed in eastern Tibet and Western Sichuan at 29° - 402
31°N, mainly concentrated in the north of Medog County, Zayu County, the east of Bome 403
County, Nyingchi City, Tibet Autonomous Region, the South and east of Markam County, 404
Qamdo City, the South and west of Zuogong county and the east of Baxoi County; North- 405
west of Deqen County, Diqing Tibetan Autonomous Prefecture, Yunnan Province; West 406
of Batang County and southwest of Litang County, Garze Tibetan Autonomous Prefec- 407
ture, Sichuan Province. The predicted distribution map of the Bioclim model (Figure 5) 408
shows that the high potential areas of geothermal disaster are mainly distributed in east- 409
ern Tibet, Western Sichuan and Northwest Yunnan at 28° - 31°N and 97° - 100°E, mainly 410
in the north of Medog County, Nyingchi City, the northeast of Zayu County, the south- 411
west of Bome County, Markam County, Zuogong County, the East and west of Qamdo 412
Remote Sens. 2021, 13, x FOR PEER REVIEW 13 of 24

city and the east of Baxoi County; Northwest of Deqen County, Diqing Tibetan Autono- 413
mous Prefecture, Yunnan Province; The West and east of Batang County, Garze Tibetan 414
Autonomous Prefecture, Sichuan Province, the south of Litang County, Xiangcheng 415
County and Derong county. The predicted distribution map of the Domain model (Figure 416
5) shows that the high potential areas of geothermal disaster are mainly distributed in 417
eastern Tibet, Western Sichuan and Northwest Yunnan at 28° - 31°N and 97° - 100°E, 418
mainly in the north of Medog County, Nyingchi City, the east of Zayu County, Markam 419
County, Zuogong County, the East and west of Qamdo city and the east of Baxoi County, 420
Tibet Autonomous Region; Deqen County, Diqing Tibetan Autonomous Prefecture, Yun- 421
nan Province; The West and east of Batang County, Garze Tibetan Autonomous Prefec- 422
ture, Sichuan Province, the south of Litang County, the southwest of Xiangcheng County 423
and Derong county. The predicted distribution map of the GARP model (Figure 5) shows 424
that the geothermal disaster and high potential area is the largest, which is mainly distrib- 425
uted in eastern Tibet, Western Sichuan and Northwest Yunnan at 28° - 31°N and 97° - 426
100°E, mainly in the northeast of Zayu County, Nyingchi City, Tibet Autonomous Region, 427
Markam County, Zuogong County, Eastern and Western Qamdo City, and Eastern Baxoi 428
County; Deqen County and Shangri La County, Diqing Tibetan Autonomous Prefecture, 429
Yunnan Province; The West and east of Batang County, Garze Tibetan Autonomous Pre- 430
fecture, Sichuan Province, the south of Litang County, Xiangcheng County and Derong 431
county. 432

Maxent model Bioclim model

Domain model GARP model

Figure 5. Classification maps of prediction. 433

Although the geothermal sample data and environmental variables selected by each 434
model are the same, there are great differences in the prediction results of different niche 435
models on the distribution of potential areas of geothermal disaster [58,59]. The prediction 436
Remote Sens. 2021, 13, x FOR PEER REVIEW 14 of 24

differences are mainly reflected in the junction areas of Southeast Tibet Autonomous Re- 437
gion, Northwest Yunnan Province and Southwest Sichuan Province, which is mainly 438
caused by the different algorithm mechanisms of each model. The prediction area of the 439
MaxEnt model is the smallest, but it is more clear at the local level of detail because the 440
MaxEnt model algorithm infers the environmental demand of geothermal disaster and 441
simulates the distribution of geothermal disaster based on the principle of maximum en- 442
tropy. It focuses on eliminating the Commission in the simulation [60]. During the opera- 443
tion, the entropy increases with the input of environmental variables associated with each 444
geothermal sample data and the increase of iteration times, the research results tend to 445
the real niche [61], that is, under the real conditions, the distribution range of geothermal 446
disaster, and the prediction results are more delicate. The prediction area of GARP model 447
is large and the overall performance is good because use GARP model algorithm is a ge- 448
netic algorithm. It searches the environmental variables related to the research samples 449
and selects the optimal rule set to predict the distribution area of geothermal disaster. It 450
focuses on eliminating the omission error in the simulation, that is, the selection of the 451
optimal model is based on the minimum omission rate, and the research results are biato- 452
wardards the basic niche [62], that is, in the ideal state without interference, The maximum 453
distribution range that may be occupied by geothermal disaster [63], which is predicted 454
to expand the range of geothermal disaster area, and MaxEnt model algorithm can make 455
a good trade-off between omission rate and recording error rate. The prediction range of 456
Bioclim and Domain models is between MaxEnt and GARP because the Bioclim model is 457
based on the principle of the environmental envelope and the Domain model uses the 458
Gower algorithm. They are greatly affected by geothermal sample points. Almost all 459
places with sample distribution points can have a certain area of potential areas. Geother- 460
mal sample distribution data mainly comes from previous field survey mapping, and the 461
sample data has a certain preference, scientific researchers generally take samples accord- 462
ing to their own research needs or simply collect data for different work areas. The sample 463
information is scattered lacks lack systematicness and representativeness; Secondly, less 464
geothermal sample data may lead to the reduction of niche space, and due to the complex 465
geological and climatic conditions in some areas, the niche of geothermal hot spots in dif- 466
ferent regions may also drift [64]. Therefore, different models have their advantages and 467
disadvantages. They can refer to each other, comprehensively compare and select an ap- 468
propriate prediction result. 469
The four models used in this study have their advantages. MaxEnt prediction results 470
are selected as the final results, but the fitting degree of geothermal disaster is high. When 471
replaced with other geological disasters, the results may be different. Therefore, when 472
predicting the potential area of a geological disaster, multiple models should be used for 473
prediction, and then the best model should be selected. At present, most studies choose 474
potential area analysis without model screening. Due to the uncertainty of model predic- 475
tion, the results are unstable, so the combination of multiple models will improve the re- 476
liability and stability of the results. 477

3.2. Evaluation of Prediction Accuracy of Different Models 478

In this paper, 10 groups of training data and test data are used to analyze the ROC 479
curve and kappa consistency test of the four models (Figure 6). The average AUC of 480
MaxEnt model is 0.842, indicating that there is a great correlation between 12 environ- 481
mental variables and the distribution of geothermal sample points, and the prediction re- 482
sults are very accurate; The average AUC of Bioclim model is 0.693, indicating that the 483
correlation between environmental variables and the distribution of simulated objects in 484
general, and the prediction results are general; The average AUC of Domain model is 485
0.822, which shows that there is a great correlation between 12 environmental variables 486
and the distribution of geothermal sample points, and the prediction result is very accu- 487
rate; The average AUC of GARP model is 0.783 (Table 2), indicating that there is a large 488
Remote Sens. 2021, 13, x FOR PEER REVIEW 15 of 24

correlation between 12 environmental variables and the distribution of geothermal sam- 489
ple points, and the predicted results are more accurate; The average AUC of the four mod- 490
els is higher than that of the random model (AUC = 0.5), indicating that the four models 491
have good prediction effects on geothermal disaster, and the average AUC of MaxEnt 492
model is the largest. 493
By analyzing the kappa values of the four models, it can be seen that the average 494
kappa value of the MaxEnt model is 0.620, and the average kappa values of Bioclim, Do- 495
main and GARP models are 0.403, 0.552 and 0.524 respectively, indicating that the con- 496
sistency of MaxEnt model is significant, and the other three models are moderate, which 497
can be used to predict the potential area of geothermal disaster. It can be seen from Figure 498
6 that the AUC average and kappa average of the MaxEnt and Domain models are large 499
and have good consistency, while the AUC average and kappa average of the MaxEnt 500
model are slightly higher than that of the Domain model. Therefore, for geothermal dis- 501
asters, the MaxEnt model can be used as the best model to predict its potential distribu- 502
tion. 503

Figure 6. AUC and Kappa values .

Table 2. The AUC and Kappa values of the four models. 504

The area under receiver operating

Consistency test statistics (Kappa)
The area characteristic curve（AUC）
Maxent Bioclim Domain GARP Maxent Bioclim Domain GARP
1 0.894 0.730 0.872 0.814 0.707 0.452 0.628 0.605
2 0.839 0.699 0.787 0.766 0.621 0.371 0.441 0.485
3 0.848 0.754 0.811 0.768 0.586 0.436 0.560 0.490
4 0.821 0.586 0.787 0.780 0.534 0.271 0.546 0.449
5 0.868 0.763 0.835 0.820 0.698 0.525 0.592 0.570
6 0.825 0.718 0.807 0.782 0.576 0.417 0.519 0.544
7 0.795 0.654 0.786 0.750 0.579 0.366 0.460 0.490
8 0.860 0.696 0.868 0.811 0.706 0.408 0.628 0.645
Remote Sens. 2021, 13, x FOR PEER REVIEW 16 of 24

9 0.810 0.597 0.823 0.754 0.567 0.316 0.531 0.453

10 0.861 0.733 0.845 0.787 0.629 0.466 0.615 0.512
Average 0.842 0.693 0.822 0.783 0.620 0.403 0.552 0.524

3.3. Analysis of Environment Variables Affecting the Potential Area of Geothermal Disaster 505

The percent contribution of each environmental variable to the prediction of the dis- 506
tribution of potential areas of geothermal disaster is shown in Figure 7. Among the 12 507
environmental variables in the word cloud diagram (Figure 7), the fault buffer distance 508
(bio 2) has the greatest contribution to the distribution of geothermal disaster, and the LST 509
(bio 1), water buffer distance (bio 9) and Moho depth (bio 15) also have a great impact on 510
the distribution of geothermal disaster. 511

Percent contribution of environmental variables Word cloud of environmental variables

Figure 7. Environmental variables contributions. 512

From the test results of the jackknife (Figure 8), it can be seen that when only a single 513
environmental factor is used, the four factors that have the greatest impact on the normal- 514
ization training gain are LST, fault buffer distance, river buffer distance and Moho depth. 515
From the combination of the two, it can be concluded that the LST, fault buffer distance, 516
river buffer distance and Moho depth are the most important environmental variables 517
that lead to the distribution of geothermal disasters along with the Yunnan-Tibet railway 518
project. The higher the LST, the closer it is to the active fault and river, and the deeper the 519
Moho depth, the more prone it is to geothermal disaster. The line passage in this area 520
should be fully considered in the planning and design of the railway project. 521
Remote Sens. 2021, 13, x FOR PEER REVIEW 17 of 24

Figure 8. Environmental variables jackknife test.

3.4. Analysis of Key Areas of Geothermal Disaster along Yunnan-Tibet Railway 522
By overlaying the prediction results of the Yunnan-Tibet railway line and MaxEnt 523
model on the ArcGIS platform (Figure 9), the corresponding key areas of geothermal dis- 524
aster are delineated in the prediction map, and through three-dimensional modeling and 525
overlaying the local magnification of Landsat-8 remote sensing image (Figure 10), it is 526
obtained that the railway line passes through three key areas of geothermal disaster: (1) 527
Markam-Deqen key areas of geothermal disaster: distributed at the junction of eastern 528
Tibet and northwestern Yunnan at 28 ° 53 ′ N and 98 ° 37 ′ E, Specifically, it is located in 529
the canyon area where the Lancang River flows at the junction of the south of Markam 530
County, Qamdo City, Tibet Autonomous Region and the north of Deqen County, Diqing 531
Tibetan Autonomous Prefecture, Yunnan Province. It is about 1500-4000m above sea level, 532
north-south, nearly vertically distributed with the Lancang River Fault Zone, and located 533
in the docking zone between the Gondwana plate and the South China plate. There are 534
many high-temperature hot springs around, which is the second-largest key area along 535
the Yunnan-Tibet railway. (2) Zuogong-Zayu geothermal disaster key area: it is distrib- 536
uted in eastern Tibet at 29 ° 18 ′ N and 98 ° 10 ′ e. it is specifically located in the canyon 537
zone at the junction of the south of Zuogong County, Qamdo City, Tibet Autonomous 538
Region and the north of Zayu County, Nyingchi city. It is about 2000-5000m above sea 539
level and runs northwest-southeast. Nujiang River flows through it and is distributed 540
nearly parallel to the Nujiang fault zone. It is located in the block docking zone between 541
the Tengchong block and Baoshan block of the Gondwana plate, there are many high- 542
temperature hot springs around, which is the largest key area along the Yunnan-Tibet 543
railway. (3) Baxoi-Zayu geothermal disaster key area: it is distributed in eastern Tibet at 544
29 ° 20 ′ N and 96 ° 55 ′ e. it is specifically located at the junction of the southeast of Baxoi 545
County, Qamdo City, Tibet Autonomous Region and the north of Zayu County, Nyingchi 546
City, with an altitude of about 4500-5000m. It is northwest-southeast trending and distrib- 547
uted nearly parallel to the Lhari-Zayu fault zone, which is controlled by the east direction 548
of the Qinghai-Tibet Plateau and the clockwise rotation around the East Himalayan tec- 549
tonic knot, Since the Holocene, it has strong activity, mainly dextral strike-slip [58], and 550
Remote Sens. 2021, 13, x FOR PEER REVIEW 18 of 24

this area is the smallest key area along the Yunnan-Tibet railway. For the key areas of 551
geothermal disaster the railway route selection should follow the disaster reduction prin- 552
ciple of avoiding around first, and then passing through the low potential area or non- 553
potential area quickly in a short distance based on finding out the distribution character- 554
istics of geothermal anomalies [24]. 555

Figure 9. Superposition analysis of Yunnan-Tibet railway and the MaxEnt model predic-
tion results.

Map of key areas of geothermal disaster

Remote Sens. 2021, 13, x FOR PEER REVIEW 19 of 24

3D map of key areas of geothermal disaster

Figure 9. key areas of geothermal disaster along Yunnan-Tibet railway.

4. Discussion 556
Based on Landsat-8 image and niche model, this paper forecasts the potential areas 557
of geothermal disaster along the Yunnan-Tibet railway project, and draws the following 558
conclusions: 559
(1) The terrain along the Yunnan-Tibet railway project is difficult and dangerous, and 560
the area covered by ice and snow is large, so it is very difficult to carry out the conven- 561
tional large-area ground geothermal survey. Through Landsat-8 satellite image, using the 562
advantages of a large amount of information carried by thermal infrared remote sensing, 563
high monitoring accuracy and small restriction by ground conditions, we can effectively 564
perceive the ground temperature anomaly in the study area. Based on the mechanism of 565
spatial information acquisition and analysis, as an important part of earth big data, earth 566
observation data and other means can quickly, accurately and macroscopically reflect the 567
key information such as the spatial location of high-temperature heat disaster causing fac- 568
tors, and play a basic role in the construction of plateau railway engineering. 569
(2) Environmental variables are closely related to the spatial distribution of geother- 570
mal sample points. Combined with the contribution rate of environmental variables and 571
the test results of the knife cutting method, the LST, fault buffer distance, river buffer dis- 572
tance and Moho depth are the most important environmental variables that lead to the 573
distribution of geothermal disaster along the Yunnan-Tibet railway project. The higher 574
the LST is, the closer it is to the active fault and river, and the deeper the Moho depth is, 575
The more prone to geothermal disaster, the railway engineering should fully consider the 576
line passage in these areas in the planning and design. 577
(3) Four niche models (MaxEnt, Bioclim,Domain and GARP) are introduced. The 578
known distribution data of geothermal samples and environmental variables caused by 579
geothermal disasters are used to construct the model according to different algorithm op- 580
erations. Combined with the idea of an ensemble prediction system, the prediction results 581
of various models are comprehensively compared, which can reduce the false negative or 582
false positive impact caused by empirically selecting a model as much as possible, to im- 583
prove the scientificity of prediction, Each model is screened for 10 times to find the best 584
prediction results. The analysis of the ROC curve and kappa coefficient shows that the 585
four models have good prediction effects on the geothermal disaster. Among them, the 586
average value of the AUC and kappa coefficient of the MaxEnt model is the largest and 587
the accuracy is the highest. Finally, combined with the preferences of different model al- 588
gorithms, ROC curve and kappa coefficient, it is concluded that the MaxEnt model can be 589
used as the best model to predict the potential distribution of geothermal disasters. 590
(4) According to the results of the MaxEnt model, the geothermal disaster and high 591
potential areas in the study area are mainly distributed in eastern Tibet and Western Si- 592
chuan at 29° - 31°N, mainly in the north of Medog County, Zayu County, the east of Bome 593
County, Nyingchi City, Tibet Autonomous Region, the South and east of Markam County, 594
Remote Sens. 2021, 13, x FOR PEER REVIEW 20 of 24

Qamdo City, the South and west of Zuogong county and the east of Baxoi county; North- 595
west of Deqen County, Diqing Tibetan Autonomous Prefecture, Yunnan Province; West 596
of Batang County and southwest of Litang County, Garze Tibetan Autonomous Prefec- 597
ture, Sichuan Province. Among them, there are three key areas of geothermal disaster 598
along the Yunnan-Tibet railway project, namely, Markam-Deqen key area of geothermal 599
disaster, Zuogong-Zayu key area of geothermal disaster and Baxoi-Zayu key area of geo- 600
thermal disaster. The LST of these three key areas is high, and there are active fault zones 601
and a large number of geothermal hot springs in the area. Among them, Lancang River 602
and Nujiang river flow through Markam-Deqen and Zuogong-Zayu key areas of geother- 603
mal disaster respectively, there are many disasters causing factors. For the key areas of 604
geothermal disaster, the railway route selection should follow the disaster reduction prin- 605
ciple of avoiding first, and then passing through the low potential area or non-potential 606
area quickly in a short distance based on finding out the distribution characteristics of 607
geothermal anomalies. 608

5. Conclusions 609
Achieving the United Nations Sustainable Development Goals (SDGs) is one of the 610
common goals of governments and all mankind. The industry, innovation and infrastruc- 611
ture goal (SDG 9) emphasize the construction of disaster-resilient infrastructure; the sus- 612
tainable cities and communities goal (SDG 11) calls for building inclusive, safe, disaster- 613
resilient and sustainable cities and human settlements. Therefore, infrastructure construc- 614
tion such as railways plays a vital role in achieving the goals of the sustainable develop- 615
ment agenda. 616
Yunnan-Tibet railway is located in the parallel flow area of Hengduanshan and three 617
rivers in China. The internal and external dynamic geological processes are significant, 618
and the unfavorable geology such as high tectonic stress, a strong earthquake, active fault, 619
gravity unfavorable geology and thermal altered rock mass are developed, which brings 620
great challenges to the railway engineering route selection. Therefore, the route selection 621
of Railway Engineering in the survey section should not only follow the principles of 622
"topographic route selection" and "geological route selection", but also reduce the impact 623
of geological disasters on Railway Engineering from the source, and implement the con- 624
cept of "disaster reduction and route selection" of natural disasters on the risk regulation 625
of railway life cycle. In this paper, four niche models are introduced into the engineering 626
disaster prediction to predict the potential areas of geothermal disaster along the Yunnan- 627
Tibet railway project, and good results are obtained, which provides the scientific basis 628
and technical support for the engineering planning, route optimization and sustainable 629
development of the Yunnan-Tibet railway project. 630
Future research will use SDGSAT-1 data to inverse the LST of the Qinghai-Tibet Plat- 631
eau, identify regional ground temperature anomalies with high quality and efficiency, 632
and then quantitatively predict and evaluate the potential areas of geothermal disaster of 633
the Qinghai-Tibet Plateau in combination with the latest in-depth learning method, which 634
can provide reference for the survey, design and engineering construction of Railways in 635
complex and dangerous mountainous areas such as Sichuan Tibet, China Nepal and Xin- 636
jiang Tibet. 637

Author Contributions: Z.C.: methodology, writing-original draft preparation and editing, and soft- 638
ware. H.G.: writing-review and editing. X.P.: validation and formal analysis. W.Z.: investigation 639
and data curation. Z.Y.: resources and visualization. L.Z.: supervision and project administration. 640
R.C.: funding acquisition. All authors have read and agreed to the published version of the manu- 641
script. 642

Funding: Please add: This research was funded by the Strategic Priority Research Program of the 643
Chinese Academy of Sciences (grant numbers XDA19030000, XDA19090000 and XDA19090121), the 644
Remote Sens. 2021, 13, x FOR PEER REVIEW 21 of 24

Key Research and Development Program of the Sichuan Provincial Science and Technology Depart- 645
ment (grant number 2022YFS0486) and the National Natural Science Foundation of China (grant 646
number 42002298). 647

Institutional Review Board Statement: Not applicable. 648

Informed Consent Statement: Not applicable. 649

Data Availability Statement: Not applicable. 650

Acknowledgments: The authors are grateful for helpful comments from many researchers and col- 651
leagues. 652

Conflicts of Interest: The authors declare that they have no known competing financial interests or 653
personal relationships that could have appeared to influence the work reported in this paper. 654

References 655

1. Wang, J. Research on Main Engineering Geological Problems and Principles of Geological Route Selection for 656

Yunnan-Tibet Railway. Railw. Stand. Des. 2021, 65, 7–13, doi:10.13238/j.issn.1004-2954.202007220004. 657

2. Liu, Z. Analysis and Prediction of Geothermal Characteristics of Sanbaishan Tunnel. J. Railw. Sci. Eng. 2022, 1– 658

10, doi:10.19713/j.cnki.43-1423/u.T20211502. 659

3. Wang, X.; Huang, R.; Huang, G. Analysis of Effects of Groundwater Geotemperature Anomaly in Deeply 660

Lying Longon and Big Tunnel. J. Geol. Hazards Environ. Preserv. 1996, 7, 24–28, doi:CNKI:SUN:DZHB.0.1996-04- 661

004. 662

4. Guo, H.; Chen, F.; Sun, Z.; Liu, J.; Liang, D. Big Earth Data: A Practice of Sustainability Science to Achieve the 663

Sustainable Development Goals. Sci. Bull. 2021, 66, 1050–1053, doi:10.1016/j.scib.2021.01.012. 664

5. Guo, H.; Liang, D.; Chen, F.; Sun, Z.; Liu, J. Big Earth Data Facilitates Sustainable Development Goals. Bull. 665

Chinese Acad. Sci. 2021, 36, 874–884, doi:10.16418/j.issn.1000-3045.20210707006. 666

6. Yan, J.; He, C.; Zeng, Y.; Wang, B.; Zhang, J. Cooling Technology and Effect Analysis for High Geothermal 667

Tunnel on Sichuan-Tibet Railway. China Railw. Sci. 2019, 40, 53–62, doi:10.3969/j.issn.1001-4632.2019.05.08. 668

7. Luo, F.; Du, S.; Meng, X.; Zhang, Z.; Dong, Q.; Hao, L.; Zhang, W.; Huang, Y.; Zhang, X.; Wang, Y. Research on 669

the “Space-Aero-Ground” Remote Sensing Geological Survey Technology of Sichuan-Tibet Railway. J. Railw. 670

Eng. Soc. 2021, 38, 1–7. 671

8. Gruber, S.; Haeberli, W. Permafrost in Steep Bedrock Slopes and Its Temperature-Related Destabilization 672

Following Climate Change. J. Geophys. Res. 2007, 112, F02S18, doi:10.1029/2006JF000547. 673

9. Isaksen, K.; Sollid, J.L.; Holmlund, P.; Harris, C. Recent Warming of Mountain Permafrost in Svalbard and 674

Scandinavia. J. Geophys. Res. 2007, 112, F02S04, doi:10.1029/2006JF000522. 675

10. Liu, G.; Zhao, L.; Xie, C.; Pang, Q.; Du, E.; Qiao, Y. Variation Characteristics and Impact Factors of the Depth of 676

Zero Annual Amplitude of Ground Temperature in Permafrost Regions on the Tibetan Plateau. J. Glaciol. 677

Geocryol. 2016, 38, 1189–1200, doi:10.7522/j.issn.1000-0240.2016.0139. 678

11. Zhang, N.; Zhang, W.; Li, L.; Lan, H. Geothermal Characteristics of Qinghai-Tibet Plateau and Risk Analysis of 679

Deep Underground Geothermal Hazards for Major Linear Sichuan-Tibet Railway Project. J. Eng. Geol. 2021, 1– 680

17, doi:10.13544/j.cnki.jeg.2021-0118. 681

12. Chen, Z.; Chang, R.; Zhao, W.; Li, S.; Guo, H.; Xiao, K.; Wu, L.; Hou, D.; Zou, L. Quantitative Prediction and 682

Evaluation of Geothermal Resource Areas in the Southwest Section of the Mid-Spine Belt of Beautiful China. 683

Int. J. Digit. Earth 2022, 15, 748–769, doi:10.1080/17538947.2022.2061055. 684

13. Xiao, K.; Xiang, J.; Fan, M.; Xu, Y. 3D Mineral Prospectivity Mapping Based on Deep Metallogenic Prediction 685

Theory: A Case Study of the Lala Copper Mine, Sichuan, China. J. Earth Sci. 2021, 32, 348–357, 686

doi:10.1007/s12583-021-1437-8. 687
Remote Sens. 2021, 13, x FOR PEER REVIEW 22 of 24

14. Chen, J.; Xiang, J.; Hu, Q.; Yang, W.; Lai, Z.; Hu, B.; Wei, W. Quantitative Geoscience and Geological Big Data 688

Development: A Review. Acta Geol. Sin. - English Ed. 2016, 90, 1490–1515, doi:10.1111/1755-6724.12782. 689

15. Zhang, S.; Carranza, E.J.M.; Xiao, K.; Wei, H.; Yang, F.; Chen, Z.; Li, N.; Xiang, J. Mineral Prospectivity 690

Mapping Based on Isolation Forest and Random Forest: Implication for the Existence of Spatial Signature of 691

Mineralization in Outliers. Nat. Resour. Res. 2021, doi:10.1007/s11053-021-09872-y. 692

16. Zhang, S.; Xiao, K.; Carranza, E.J.M.; Yang, F. Maximum Entropy and Random Forest Modeling of Mineral 693

Potential: Analysis of Gold Prospectivity in the Hezuo–Meiwu District, West Qinling Orogen, China. Nat. 694

Resour. Res. 2019, 28, 645–664, doi:10.1007/s11053-018-9425-0. 695

17. Zhao, W.; Dong, Q.; Chen, Z.; Feng, T.; Wang, D.; Jiang, L.; Du, S.; Zhang, X.; Meng, D.; Bian, M.; et al. 696

Weighted Information Models for the Quantitative Prediction and Evaluation of the Geothermal Anomaly 697

Area in the Plateau: A Case Study of the Sichuan–Tibet Railway. Remote Sens. 2021, 13, 1606, 698

doi:10.3390/rs13091606. 699

18. Gengping, Z.; Guoqing, L.; Wenjun, B.; Yubao, G. Ecological Niche Modeling and Its Applications in 700

Biodiversity Conservation. Biodivers. Sci. 2013, 21, 90–98, doi:10.3724/SP.J.1003.2013.09106. 701

19. Araujo, M.; New, M. Ensemble Forecasting of Species Distributions. Trends Ecol. Evol. 2007, 22, 42–47, 702

doi:10.1016/j.tree.2006.09.010. 703

20. Wang, G.; Gen, Q.; Xiao, M.; Zhang, M.; Zhang, Y.; Wang, Z. Predicting Pseudolarix Amabilis Potential Habitat 704

Based on Four Niche Models. Acta Ecol. Sin. 2020, 40, 6096–6104, doi:10.5846/stxb201907021390. 705

21. Zhang, Q.; Zhang, D.F.; Wu, M.L.; Guo, J.; Sun, C.Z.; Xie, C.X. Predicting the Global Areas for Potential 706

Distribution of Gastrodia Elata Based on Ecological Niche Models. Chinese J. Plant Ecol. 2017, 41, 770–778, 707

doi:10.17521/cjpe.2016.0380. 708

22. Ming, Q. A Study on The Neotectonic Division＆ Environment Evolution of Qing-Zang Plateau＆ Three 709

Parallel Riversarea. 2007, 4, 387–396, doi:10.3969/j.issn.1004-1885.2007.04.002. 710

23. Zheng, Y.; Yang, Y.; Jiang, L. The Particularity of Engineering Geology on Yunnan-Tibet Route in Glacier 711

Snowfield Area. J. Railw. Eng. Soc. 2005, A1, 216–221, doi:10.3969/j.issn.1006-2106.2005.z1.028. 712

24. Song, Z.; Zhu, Y.; Wei, Y.; Jiang, L.; Zhang, G. Research on the Engineering Geological Hazards-Reduction 713

Route Selection in Three Parallel Rivers Regions for the Yunnan-Tibet Railway. J. Railw. Eng. Soc. 2019, 36, 1–6. 714

25. Zhang, J.; Zhang, Y.; Qu, Y.; Cao, H. In Depth Discussions on The Engineering Geological Problems Along 715

Sanjiang Sector of Yunan-Tibet Railway Line. J. Eng. Geol. 2010, 18, 781–789, doi:10.3969/j.issn.1004- 716

9665.2010.05.028. 717

26. Liu, A.; Zheng, J.; Li, J.; Du, J.; Li, F. Comparison Study on Resource Evaluation Methods of Shallow 718

Geothermal Energy and Traditional Geothermal Energy. Urban Geol. 2018, 13, 37–41, doi:10.3969/j.issn.1007- 719

1903.2018.02.006. 720

27. Zhang, L.; Li, G. Shallow Geothermal Resources. West. Resour. 2019, 1, 178–180, doi:10.16631/j.cnki.cn15- 721

1331/p.2019.01.072. 722

28. Wang, X.; Huang, R. Evaluation and Prediction of Temperature Field in Deep Buried Long Tunnel. Hydrogeol. 723

Eng. Geol. 1996, 23, 6–10, doi:CNKI:SUN:SWDG.0.1996-06-001. 724

29. Chen, Z.; Qing, D.; Chen, J.; Zhao, W.; Jiang, L.; Zhang, G.; Feng, T.; Wang, D.; Bi, X.; Bian, M.; et al. Research 725

on Quantitative Prediction and Evaluation of Geo- Thermal Anomaly Area in Qamdo-Nyingchi Section of 726

Sichuan-Tibet Railway. Remote Sens. Technol. Appl. 2021, 36, 1368–1378, doi:10.11873/j.issn.1004- 727

0323.2021.6.1368. 728
Remote Sens. 2021, 13, x FOR PEER REVIEW 23 of 24

30. Revil, A.; Pezard, P.A. Streaming Electrical Potential Anomaly along Faults in Geothermal Areas. Geophys. Res. 729

Lett. 1998, 25, 3197–3200, doi:10.1029/98GL02384. 730

31. Xiao, K.; Pirajno, F.; Xu, Y.; Zhang, S.; Sun, L.; Cong, Y.; Fan, J.; Yin, J.; Zhu, Y.; Wei, H. Study on the Major 731

Minerals Potential in China. Ore Geol. Rev. 2020, 127, 103816, doi:10.1016/j.oregeorev.2020.103816. 732

32. Bayrak, Y.; Bayrak, E. Regional Variations and Correlations of Gutenberge-Richter Parameters and Fractal 733

Dimension for the Different Seismogenic Zones in Western Anatolia. J. Asian Earth Sci. 2012, 58, 98–107, 734

doi:10.1016/j.jseaes.2012.06.018. 735

33. Aydogan, D. Extraction of Lineaments from Gravity Anomaly Maps Using the Gradient Calculation: 736

Application to Central Anatolia. Earth, Planets Sp. 2011, 63, 903–913, doi:10.5047/eps.2011.04.003. 737

34. Wang, D.F. Aeromagnetic Series Maps and Specifications of Qinghai–Tibet Plateau and Adjacent Areas. 738

Beijing, China Geol. Publ. House 2013. 739

35. Qiu, N.; Tang, B.; Zhu. Chuanqing Deep Thermal Background of Hot Spring Distribution in the Chinese 740

Continent. Acta Geol. Sin. 2022, 96, 195–207, doi:10.19762/j.cnki.dizhixuebao.2022022. 741

36. Fang, S.; Mao, K.; Xia, X.; Wang, P.; Shi, J.; Bateni, S.M.; Xu, T.; Cao, M.; Heggy, E.; Qin, Z. Dataset of Daily 742

Near-Surface Air Temperature in China from 1979 to 2018. Earth Syst. Sci. Data 2022, 14, 1413–1432, 743

doi:10.5194/essd-14-1413-2022, 2022. 744

37. 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 745

Permafrost Distribution on the Tibetan Plateau. Cryosph. 2017, 11, 2527–2542, doi:10.5194/tc-11-2527-2017. 746

38. Yan, D.; Ma, N.; Zhang, Y. Development of a Fine-Resolution Snow Depth Product Based on the Snow Cover 747

Probability for the Tibetan Plateau: Validation and Spatial–temporal Analyses. J. Hydrol. 2022, 604, 127027, 748

doi:10.1016/j.jhydrol.2021.127027. 749

39. Wang, J.; Wang, J.; Ye, H. An Interpolated Temperature and Precipitation Dataset at 1-Km Grid Resolution in 750

China (2000-2012). China Sci. Data 2017, 2, doi:10.11922/csdata.170.2016.0112. 751

40. Graham, M.H. Confronting Multicollinearity in Ecological Multiple Regression. Ecology 2003, 84, 809–2815, 752

doi:10.1890/02-3114. 753

41. Zhang, M.G.; Slik, J.W.F.; Ma, K.P. Using Species Distribution Modeling to Delineate the Botanical Richness 754

Patterns and Phytogeographical Regions of China. Sci. Rep. 2016, 6, 22400, doi:10.1038/srep22400. 755

42. Mai, J.; Xian, Y.; Liu, G. Predicting Potential Rainfall-Triggered Landslides Sites in Guangdong Province 756

(China) Using MaxEnt Model under Climate Changes Scenarios. J. Geo-Information Sci. 2021, 23, 2042–2054, 757

doi:10.12082/dqxxkx.2021.210182. 758

43. Franklin, J.; Davis, F.W.; Ikegami, M.; Syphard, A.D.; Flint, L.E.; Flint, A.L.; Hannah, L. Modeling Plant Species 759

Distributions under Future Climates: How Fine Scale Do Climate Projections Need to Be? Glob. Chang. Biol. 760

2013, 19, 473–483, doi:10.1111/gcb.12051. 761

44. Phillips, S.J.; Anderson, R.P.; Schapire, R.E. Maximum Entropy Modeling of Species Geographic Distributions. 762

Ecol. Modell. 2006, 190, 231–259, doi:10.1016/j.ecolmodel.2005.03.026. 763

45. Booth, T.H.; Nix, H.A.; Busby, J.R.; Hutchinson, M.F. Bioclim : The First Species Distribution Modelling 764

Package, Its Early Applications and Relevance to Most Current MaxEnt Studies. Divers. Distrib. 2014, 20, 1–9, 765

doi:10.1111/ddi.12144. 766

46. Busby, J.R. BIOCLIM - a Bioclimate Analysis and Prediction System. Nat. Conserv. cost Eff. Biol. Surv. data Anal. 767

1991, 64–68. 768

47. Guisan, A.; Zimmermann, N.E. Predictive Habitat Distribution Models in Ecology. Ecol. Modell. 2000, 135, 147– 769

186, doi:10.1016/S0304-3800(00)00354-9. 770

Remote Sens. 2021, 13, x FOR PEER REVIEW 24 of 24

48. Yunsheng, W. Application of ROC Curve Analysis in Evaluating the Performance of Alien Species’ Potential 771

Distribution Models. Biodivers. Sci. 2007, 15, 365, doi:10.1360/biodiv.060280. 772

49. Higgins, S.I.; Richardson, D.M.; Cowling, R.M. Modeling Invasive Plant Spread: The Role of Plant- 773

Environment Interactions and Model Structure. Ecology 1996, 77, 2043–2054, doi:10.2307/2265699. 774

50. Higgins, S.I.; Richardson, D.M.; Cowling, R.M.; Trinder-Smith, T.H. Predicting the Landscape-Scale 775

Distribution of Alien Plants and Their Threat to Plant Diversity. Conserv. Biol. 1999, 13, 303–313, 776

doi:10.1046/j.1523-1739.1999.013002303.x. 777

51. Stockwell, D.; Peters, D. The GARP Modelling System: Problems and Solutions to Automated Spatial 778

Prediction. Int. J. Geogr. Inf. Sci. 1999, 13, 143–158, doi:10.1080/136588199241391. 779

52. Swets, J.A. Measuring the Accuracy of Diagnostic Systems. Science (80-. ). 1988, 240, 1285–1293, 780

doi:10.1126/science.3287615. 781

53. Chen, F.; Du, Y.; Niu, S.; Zhao, J. Modeling Forest Lightning Fire Occurrence in the Daxinganling Mountains of 782

Northeastern China with MAXENT. Forests 2015, 6, 1422–1438, doi:10.3390/f6051422. 783

54. West, A.M.; Kumar, S.; Brown, C.S.; Stohlgren, T.J.; Bromberg, J. Field Validation of an Invasive Species 784

Maxent Model. Ecol. Inform. 2016, 36, 126–134, doi:10.1016/j.ecoinf.2016.11.001. 785

55. Xu, W.N.; Wang, P.X.; Han, P.; Yan, T.L.; Zhang, S.Y. Application of Kappa Coefficient to Accuracy 786

Assessments of Drought Forecasting Model: A Case Study of Guanzhong Plain. J. Nat. Disasters 2011, 20, 81–86, 787

doi:10.1631/jzus.B1000265. 788

56. Manel, S.; Williams, H.C.; Ormerod, S.J. Evaluating Presence-Absence Models in Ecology: The Need to 789

Account for Prevalence. J. Appl. Ecol. 2001, 38, 921–931, doi:10.1046/j.1365-2664.2001.00647.x. 790

57. Segurado, P.; Araújo, M.B. An Evaluation of Methods for Modelling Species Distributions. J. Biogeogr. 2004, 31, 791

1555–1568, doi:10.1111/j.1365-2699.2004.01076.x. 792

58. Qin, Z.; Zhang, J.; DiTommaso, A.; Wang, R.; Wu, R. Predicting Invasions of Wedelia Trilobata (L.) Hitchc. 793

with Maxent and GARP Models. J. Plant Res. 2015, 128, 763–775, doi:10.1007/s10265-015-0738-3. 794

59. Hernandez, P.A.; Graham, C.H.; Master, L.L.; Albert, D.L. The Effect of Sample Size and Species Characteristics 795

on Performance of Different Species Distribution Modeling Methods. Ecography (Cop.). 2006, 29, 773–785, 796

doi:10.1111/j.0906-7590.2006.04700.x. 797

60. Townsend, P.A.; Papeş, M.; Eaton, M. Transferability and Model Evaluation in Ecological Niche Modeling: A 798

Comparison of GARP and Maxent. Ecography (Cop.). 2007, 30, 550–560, doi:10.1111/j.2007.0906-7590.05102.x. 799

61. Kong, W.Y.; Li, X.H.; Zou, H.F. Optimizing MaxEnt Model in the Prediction of Species Distribution. Ying yong 800

sheng tai xue bao = J. Appl. Ecol. 2019, 30, 2116–2128, doi:10.13287/j.1001-9332.201906.029. 801

62. Zhang, C.; Chen, L.; Tian, C.M.; Li, T.; Wang, R.; Yang, Q.Q. Predicting the Distribution of Dwarf Mistletoe 802

(Arceuthobium Sichuanense) with GARP and MaxEnt Models. Beijing Linye Daxue Xuebao/Journal Beijing For. 803

Univ. 2016, 13, 23–32, doi:10.13332/j.1000-1522.20150516. 804

63. Elith, J.; H. Graham, C.; P. Anderson, R.; Dudík, M.; Ferrier, S.; Guisan, A.; J. Hijmans, R.; Huettmann, F.; R. 805

Leathwick, J.; Lehmann, A.; et al. Novel Methods Improve Prediction of Species’ Distributions from 806

Occurrence Data. Ecography (Cop.). 2006, 29, 129–151, doi:10.1111/j.2006.0906-7590.04596.x. 807

64. Phillips, S.J.; Dudík, M.; Schapire, R.E. A Maximum Entropy Approach to Species Distribution Modeling. In 808

Proceedings of the Twenty-first international conference on Machine learning - ICML ’04; ACM Press: New 809

York, New York, USA, 2004; p. 83. 810

811