Changing Patterns of Lakes on The Southern Tibetan Plateau Based on Multi ‐ Source Satellite Data

: More than 1100 lakes covering an area greater than 4500 km 2 are located on the Tibetan Plateau, and these lakes are important regulators of several large and famous rivers in Asia. The determination of hydrological changes that have occurred in these lakes can reflect climate change and supply scientific data to plateau environmental research. Data from high frequency (moderate ‐ resolution imaging spectro ‐ radiometer) MODIS images, altimetry, and the Hydroweb database collected during 2000–2015 were integrated in this study to delineate the detailed hydrological changes of 15 lakes in three basins—Inner Basin, Indus Basin, and Brahmaputra Basin—on the southern Tibetan Plateau. Seven of the 10 lakes in the Inner Basin presented increasing trends with various intensities, and the increasing rates in the area of four lakes (Nam Co, Selin Co, Zhari ‐ namco, and Ngangze) were 1.62, 28.81, 2.27, and 3.70 km 2 /yr, respectively. The yearly increases in volume of the four lakes were 3.6, 9.44, 6, and 2.36 km 3 , respectively. A water balance equation was established for the four lakes based on lake volume changes to illustrate the contributions of precipitation, ground runoff, evaporation, and other factors. The results revealed that surface runoff was the major contributor to expansion, and lake surface evaporation was almost 2.76–3.86 times that of lake surface precipitation. The two lakes in Indus Basin, Rakshastal and Mapam Yumco, presented a slight retreat. As a representative of Brahmaputra Basin, Yamzho Yumco underwent a retreat of –3.49 km 2 /yr in area, –0.39 m/yr in level, and –0.19 km 3 /yr in volume. Decreasing precipitation, increasing evaporation, and the operation of a hydrological project were the main causes of its constant retreat.


Introduction
Hydrologic changes in lakes reflect the comprehensive influences of climate, land-surface processes, and human activities on the water cycle and ecosystem. As the world's highest and largest plateau, the Tibetan Plateau (TP) is located at 26°00′ N-39°47′ N, 73°19′ E-104°47′ E in Central and East Asia. The area of this plateau is approximately 2.5 million km 2 , with an average elevation of higher than 4500 m. The TP is known as the "Water tower of Asia", and it is the source of several large rivers, such as the Yellow River, Yangtze River, Ganges River, India River, Brahmaputra River, and so on, providing domestic water supply for hundreds of millions of populations. On the TP, the number of lakes greater than 1 km 2 in size is nearly 1100, and the total area of these lakes is nearly 45,000 km 2 . Due to the remoteness and harsh environment of the TP, the intensity of agricultural and industrial activities is low. Therefore, the hydrologic changes of lakes on the plateau can be treated as a sensitive indictor of climate change. Many researchers have studied the changes in lakes on the plateau. The research has reported that lakes greater than 10 km 2 in size on the plateau showed an increasing trend during the last decade, and the total area increased from 37,111 to 40,015 km 2 , at a rate of 7.82% [1]. The water level of the two largest lakes in the central Tibet, Nam Co and Selin Co, has risen by 4.37 m and 2.03 m, respectively, in the last decade [2]. The lake volume in the Inner Basin on the TP increased at a rate of 7.72 Gt/yr during 2003-2009, mainly due to precipitation contribution and glacier mass loss [3]. These research results were mostly based on remote sensing techniques as the gauged observations on the plateau were very limited. There are only several hydrological stations in the large Tibetan region, and these stations present a decentralized and punctate distribution. Satellite data play an important role in delineating the hydrologic changes of these plateau lakes with high efficiency and at a low cost [4,5].
Altimeter data have been widely used to continuously monitor water level changes [6,7]. Starting from the 1990s, there are 30 years of altimeter data covering the globe with a frequency of 10-30 days. This includes Topex/Poseidon (T/P), Jason, ENVIromental SATellite (Envisat), Ice, Cloud and Land Elevation Satellite (ICESat), and Geosat Follow on (GFO) data. T/P data collected during 1992-2002 were applied to the six largest lakes in China [8]. Zheng et al. [9] used T/P and Envisat data during 1992-2010 to monitor water level changes of Hulun Lake in northeast China and found that the lake presented a decreasing trend. T/P data collected during 1992-1999 were applied on rivers with a width of more than 1 km in the Amazon watershed [7]. ICESat data collected during 2003-2009 were employed on 56 lakes in China, showing different change patterns of lake surface level [10]. At present, there are four kinds of water level databases for large rivers, lakes, and reservoirs that are derived from altimeter data: the Database for Hydrological Time Series of Inland Waters (DAHITI) [11], the Global Reservoir and Lake Monitor (GRLM) [12], the River Lake Hydrology product (RLH) [13], and the Hydroweb [14].
Remotely sensed images are able to capture lake area fluctuations occurring over short periods or decades. Multi-source remote sensing images were employed to delineate the monthly spatial distribution of global land surface water bodies in 1993-2004 [15,16]. Feng et al. [17] used moderateresolution imaging spectro-radiometer (MODIS) images to monitor dynamic changes in Poyang Lake in 2000-2010 and found that the area in the wet season was four times of that in the dry season. Sun et al. [1] used MODIS images to study the inundation changes of more than 600 large lakes in China in 2000-2010. Some researches have combined altimeter data and satellite images to study volume changes of inland water bodies. Water mass changes of the Negro River basin were revealed by synthetic aperture radar (SAR), T/P and in situ water level observations [18]. ICESat data and Landsat images were used in conjunction to construct area-level curves for 30 lakes on the TP to study their volume changes and the result showed that there was an increase in water volume of 92.43 km 3 for the 30 lakes from the 1970s to 2011 [19]. Cai et al. [20] constructed the area-volume models for 128 lakes and 108 reservoirs in Yangtze River watershed according to gauged measurements and MODIS images, and found that 53.91% of lakes were shrinking, while the reservoirs were expanding. Sun et al. [21] used water level data derived from altimeter data to construct the bottom topography of the Aral Sea and obtained its water volume changes. Medina et al. [22] applied gauged water level measurements, and Envisat and Advanced SAR images to estimate storage changes in Lake Izabal. In addition, Gravity Recovery and Climate Experiment (GRACE) data were applied to the temporal-spatial terrestrial water mass changes [23].
At present, detailed changes in the level, area and storage of lakes on the southern TP have not been clear. An illustration of seasonal and inter-annual hydrological changes of the main lakes on the southern TP would be beneficial to improve the understanding of water cycles and the climate changes occurring on the TP. As altimeter data are in the form of footprints and are obtained by recording values along sparse ground tracks, the derivation of level fluctuations for lakes in the large southern TP areas requires several kinds of altimeter data to be combined. In addition, high frequency remotely sensed images are capable of capturing the dynamics of lakes in a short period of time. This paper used the aforementioned altimeter data and MODIS images to illustrate the dynamics of 15 lakes greater than 100 km 2 in size on the southern TP, and the contributions of climate factors were discussed for the lakes with obvious changes. The results may serve as important information for land-surface process models and may be used in plateau researches.

Study Area
The environment of the TP is very harsh due to the high altitude. It has dry and thin air, strong solar radiation, and low temperatures. The climate of the TP varies greatly from region to region because of the complex terrain. The annual average temperature is 20 ℃ in the warm and humid southeast, while it drops to below -6 ℃ in the cold and dry northwest. The annual precipitation ranges from 100 to 1300 mm, decreasing from southeast to northwest. The studied area is bordered to the east by the Tanggula Mountains and to the south by the Himalayan range, and the Brahmaputra River is also in this region. The elevation is higher than 4000 m and several large glaciers are distributed in the surrounding high mountains.
The researched 15 lakes are located in three basins on the southern TP. Ten lakes including Selin Co and Nam Co are located in the Inner Basin, which occupies 70% of the area of the TP. The Inner Basin is located in the Qiangtang steppe and has a semi-arid climate. The average temperature of the Inner Basin is about 0-2℃, and the annual rainfall is 150-200 mm. Two lakes, Mapam Yumco and Rakshastal, are in the Hindu Basin and the other three lakes, including Yamzho Yumco, are in the Brahmaputra Basin. Nam Co, Selin Co, and Yamzho Yumco are known as the "Three holy lakes" on the TP. All of the 15 lakes are greater than 100 km 2 in size. Selin Co and Nam Co are the two largest lakes; they are greater than 1800 km 2 in size. Selin Co is a Tectonic lake, with an open lakeside and dense water plants, and it is the main region for livestock. Selin Co has shown continuous expansion and has been the largest lake on the TP since 2003 [1]. Nam Co is the highest lake in the world with an elevation of around 4725 m. It is close to Tanggula Mountain and its southern shoreline is parallel to the Tanggula ridge. The area of Zhari-namco is nearly 1000 km 2 in size. Yamzho Yumco lies 10 km south of the Brahmaputra River. Yamzho Yumco and several small lakes around it are considered to be the largest group of lakes on the southern TP. The west two adjacent lakes, Mapam Yumco and Rakshastal, are between two high mountains-the Gangdise Mountains (5500 m) and Namunani peak (7728 m)-where mountain glaciers are widely distributed. Therefore, the water supply from melted glaciers or snow is very rich for these two lakes. The spatial distribution of the 15 lakes is shown in Figure 1, and characteristics of the lakes are listed in Table 1.

MODIS Images
Long-term and high frequency MODIS images have been widely used to monitor dynamic changes in land cover [1]. The level 3, 8-day composited product MOD09A, with 500 m resolution, was downloaded from the Earth Observing System (https://reverb.echo.nasa.gov/reverb/). The product was atmospheric corrected and consisted of 7 bands (0.648-2.130 μm). As the fifth band (centered at 1.24 μm) contained stripe noises, it was excluded from the data processing. The data used in this research started on the 49th day of the year 2000 (2000-049) and ended on the 361st day of the year 2015 (2015-361). It was worth mentioning that, in some images, lakes were covered by thick clouds or snow, or were frozen, especially in winter seasons. In these cases, lake boundaries were not easy to distinguish. Therefore, these kinds of images were discarded from the following process. The largest available number of images was 573, for Yamzho Yumco, while the smallest was 322, for Nam Co. The detailed information about the start/end dates and the numbers of available series for each lake is listed in Table 2.

Altimeter Data
The applications of T/P and Jason series were very extensive due to their short revisiting periods (10 days). However, the interval between the adjacent ground tracks of these altimeter data is very large, around 1 km. Thus, less data were available for the 15 studied lakes and only Envisat and ICESat were used in this research. Envisat carried a radar altimeter. The radar altimeter sent microwave pulses and was able to carry on monitoring under various weather conditions. As a successor of the European Remote Sensing satellites (ERS), Envisat was launched in March 2002 by the European Space Agency (ESA). Envisat revisited the same cycle over 35 days, covering a latitude of ±81.5 degrees, with a footprint of around 2-10 km in diameter. Envisat RA-2 Geographical Data Records (GDR) were used in this research. ICESat sent laser pulses and was easily interfered by clouds. ICESat was launched in January 2003 and stopped working in 2009. ICESat covered a latitude of ±86 degrees, and the diameter of its ground footprint was only 70 m. The frequency of ICESat data was about 3-4 months. The product Global Land Surface Altimetry Data (GLA14) was downloaded from the National Snow and Ice Data Center (NSIDC). Fourteen lakes were covered by Envisat data and 12 lakes were covered by ICESat data, as Table 2 shows.

Hydroweb
Hydroweb is a database created by LEGOS/GOHS (Laboratoire d'Études en Géophysique et Océanographie Spatiale/Equipe Geodesie, Oceanograhie et Hydrologie Spatiale), which combines T/P, Jason-1, Jason-2, Envisat, and GFO altimetry data to present the water level of 150 large global lakes/reservoirs [14]. The available period in the database is from the 1990s until the present. Compared with gauged data, the derived accuracy of the water surface for lakes/reservoirs is 5-24 cm. Seven of the studied lakes were included in this database, as shown in Table 2.

Meteorological Data
The daily gauged precipitation, evaporation, and temperature data of the southern TP during the studied period were obtained from the China Meteorological Data Sharing Service System (http://cdc.cma.gov.cn/). There are 19 metrological stations in the study area, and the spatial distributions of them are shown in Figure 1. To match the hydrological results of this study, daily metrological data were accumulated into a monthly scale format. The monthly precipitation for the whole study area was estimated by kriging interpolation of the measured data. In addition, Global Land Data Assimilation System (GLDAS) data (https://ldas.gsfc.nasa.gov/gldas) collected monthly in 2000-2015 were downloaded to estimate the land evaporation in sub-basins.

Inundation Extraction
Water indexes such as the normalized differenced water index (NDWI), modified NDWI (MNDWI), and automated water extraction index (AWEI) were not suitable for use in this research, as their key thresholds were easily influenced by imaging environments, such as aerosol interference and viewing geometry [17]. The use of various suitable thresholds was not practical. Therefore, an accurate water-land discriminating method was applied in this research. In the procedure, accurate and automatic extraction of training data was firstly carried out for each image, and then the images were classified by the support vector machine (SVM) classifier [24]. Finally, surface inundations of the 15 lakes were extracted according to their spatial locations [21]. With respect to the training data selection, six rules were constructed to select initial training data for six classes: water bodies, bare soil (including urban area), vegetation, ice, snow, and clouds. Then, the initial training samples were refined based on an iterative procedure integrating the k-means algorithm and the AWEI, whose threshold is more stable and fluctuates in a smaller range than those of NDWI and MNDWI. As the threshold of AWEI was suggested to be around -0.15 to 0.045 [25], the threshold of water clusters in the proposed method was determined to be -0.1. Taking the initial water training samples as an example, in each iteration process, the samples were divided into 10 clusters, and the AWEI of each cluster was computed. If the AWEI of a cluster was greater than -0.1, then this cluster was considered to be a typical water training set. If not, this cluster was continuously divided into 10 clusters by kmeans, and the corresponding AWEIs of newly created sub-clusters were calculated for the following assessment. Ten iterations of this process were performed. When the iterations ended, if the AWEI of some sub-cluster was still less than -0.1, it was removed from the water training sets. For other nonwater training data, the procedure was similar, except that the AWEI had to be less than -0.1 and the outliers were removed from non-water training data at the end of the iterations. This automatic water extraction method was used in high-frequency water surface delineations of the Aral Sea and Poyang Lake in 2000-2015. The omission errors were 0.9%-1.5% and commission errors were 2.94%-4.23% [21].

Water Level
An altimeter assembled on a satellite sends a series of pulses and receives reflected signals. Therefore, the distance between the satellite and water surface can be derived by counting the time taken for the pulses to return. The altitude represents the orbit height of the satellite and the range indicates the distance between altimeters and the water surface. The height of the water surface relative to some referenced ellipsoid is equal to the altitude minus the range [26]. The influences of the solid earth tide and pole tide need to be removed, as shown in Formula (1). In addition, the range needs to include various environmental and geophysical corrections, as listed in Formula (2).
Sea Surface Height = Altitude -Corrected Range -Solid earth tide -Pole tide (1) Corrected Range = Range + Wet troposphere correction + Dry troposphere correction + Ionosphere correction (2) Envisat and ICESat data were processed based on the above formulas. It is worth mentioning that the Ice-1 retracking algorithm was chosen to process the Envisat RA-2 product as it had higher accuracy when deriving the range [9]. Dry and wet troposphere corrections were derived based on the European Centre for Medium-Range Weather Forecasts (ECMWF) model, and the ionosphere correction was derived from the electron content from the Global Ionosphere Maps (GIM). The processing of the ICESat GLA14 product was similar to that applied by Zhang et al. [2].
Altimeter data were in the form of footprints and the measured values were recorded along the ground tracks of the altimeter. As lakes may undergo dynamics, footprints may fall on the lakeside when a lake had a small extent as Figure 2 shows. To remove the contamination of land signals, only footprints falling on the lake surface were filtered by the coastline on the same date. Therefore, the high frequency lake boundaries created in Section 3.2.1 were applied as filters. Then, the 3-sigma rule was applied to remove outliers from the filtered results, and the mean value of the remaining results was calculated to represent the lake surface elevation on a particular day. Finally, the daily results were cumulated to obtain the time series water level of each lake. The reference ellipsoid used for the ICESat data was Topex/Poseidon, while Envisat used the World Geodetic System 84 (WGS84) as a reference. Thus, ICESat data were converted to the WGS84 reference to maintain consistency. System differences existed between different altimeters in the orbit height, orbit inclination, and revisit period. There were also some differences between Hydroweb outcomes and the two derived altimetry results. Therefore, for lakes covered by two or three kinds of water level source, it was necessary to transform the water level results into one base value to create a long-term series. Correlations between different kinds of results were constructed to accomplish this goal.
A correlation was first constructed based on results obtained during overlapping periods. The correlation of two kinds of water results was considered to be credible only when more than 10 pairs of data were present during overlapping periods and the R 2 value of the correlation was greater than 0.88. In this case, the altimeter results were converted to the base value of the results with the longest time range to create a more intensive level series. If the R 2 value was less than 0.88, the level results with the longer time span were chosen for the next step of the analysis. The available altimeter and Hydroweb data for the 15 lakes are presented in Table 2. Seven lakes with ID numbers 13, 14, 15, 2, 3, 4, and 5 were all covered by all the three sources. Lakes with ID numbers 1, 6, 7, and 9 were covered by both ICESat and Envisat data. As the number of overlapping records between Envisat and ICESat data was less than 10 for these four lakes, only Envisat results were used. Lakes Rakshastal and Shuru Co were only covered by Envisat series, and the levels of Daggayai Co and Pegu Cao were only covered by ICESat data. Figure 3 displays the three sources of lake level results for Selin Co, Nam Co and Tangra-yumco. The ICESat results were usually 2 m higher than those from the other two datasets. Selin Co had more ICESat-and Envisat-derived results, and the three sources of lake level data had consistent fluctuations. The correlation between ICESat results and Hydroweb had an R 2 value of 0.96. The R 2 value for the comparison between the Hydroweb and Envisat results was 0.92. For Selin Co, as the correlations between altimeter data and Hydroweb were strong, altimeter results were converted to the Hydroweb base to make the data dense. There were few available altimeter results for Nam Co and Tang-ro Yum Co, and the R 2 values between the altimeter results and Hydroweb were around 0.80-0.96, as shown in Figure 3. As the threshold of R 2 was defined as 0.88 in this research, for Nam Co, only the ICEsat lake level data were converted to the Hydroweb base due to the high R 2 value of 0.96. For Tangra-yumco, only the Envisat results were converted to the Hydroweb base due to the R 2 value of 0.89. The water level results of the other lakes were also determined based on correlations among the three datasets.

Lake Storage Changes
The variation in lake storage can be estimated according to the pairs of lake level and area data when the lake is assumed to be a conical frustum [9,27,28]. In this research, the changes in lake storage were derived from the changes in level and area, as shown in Formula (3).
where ∆ means the changed lake storage from one state with level and area to another state with level and area . It is worth mentioning that lakebeds on the southern TP have been less influenced by human actions, such as dredging activities. Therefore, changes in the lake bottom topography can be ignored. In the study of lake volume changes, the bottom topography under the smallest extent in 2000-2015 was assumed to be unchanged. In addition, for each lake, the area and level time series data were of different lengths. The amount of water level data was less than that of inundation results, and hence the length of storage time series depended on matched results between the water level and area series.

Fluctuation of the Water Surface
Water surface fluctuations of the studied 15 lakes showed different patterns of change: expansion, retreat, and no obvious change. Usually, lakes in the same basin presented similar trends as presented in Figure 1. Table 3 presents the information about inundation changes for the 15 lakes. Figure 4 shows the dynamic changes of the five lakes with obvious changes. The first left column is the overlapped results of the available series of lake surfaces, showing the inundated frequencies with the range from 1 to 580. The light yellow shows rarely inundated regions and dark blue indicates frequently inundated regions. The second column presents changes in the bank during the research period. The third and fourth columns display large views of some special changed regions of the lakes, and the curve plots present the area changes in each lake that occurred in 2000-2015.
In the Inner Basin, seven of the ten lakes underwent obvious expansions with various intensities. Two lakes Ngangla-ringco and Daggayai Co showed no obvious changes, and Gyaring Co showed a slight retreat with a rate of -0.16 km 2 /yr. Ngangze Co, Zhari-namco, Nam Co, and Selin Co had increasing rates of greater than 1.60 km 2 /yr. Nam Co had maximum and minimum areas of 2042.48 km 2 and 1889.21 km 2 , which occurred in October 2008 and July 2000, respectively. The lake gradually increased at a rate of 11.26 km 2 /yr from 2000 to 2005 and presented no apparent variability in the following years (Figure 4a). The rate in 2000-2015 was 1.62 km 2 /yr. The lake is shaped from northeast to southwest, and its southeast bank is restricted by ranges. Therefore, it enlarged towards the west and east (Figure 4a), especially in the year 2005. December until the next July was the dry season, and the inundated area was usually the smallest in July. Then, the area began to increase from August and reached its maximum in November.
Selin Zhari-namco had a clear increasing trend of 2.27 km 2 /yr and its changing pattern showed stage differences. Its largest area was 1019.84 km 2 in February 2010 and its smallest area was 940.85 km 2 in April 2000. Zhari-namco presented an increasing pattern in 2000-2010 and the average annual area increased from 955.83 km 2 in 2000 to 997.07 km 2 in 2010, with a rate of 4.01 km 2 /yr. Then, the lake began retreating, with the mean area being around 990 km 2 and the corresponding decreasing trend being -1.77 km 2 /y in 2011-2015. The north and south regions of the lake are restricted by rift zones, leading to lake banks being parallel to several large ranges. Therefore, the expansion mainly appeared in the west portion (Figure 4c). Zhari-namco showed apparent inter-month variability. The little expansion usually occurred in April-July and then the lake began expanding, reaching its largest size in October. From November to April, the inundated area gradually decreased.
Ngangze Co had an obvious expanding trend of 3.70 km 2 /yr. The maximum area was 470.75 km 2 in March 2015, the minimum area was 385.05 km 2 in April 2000, and the average area in the studied period was 427.55 km 2 . Ngangze Co is located in an alluvial fan with alkaline land. Several large feeding rivers flow into the western and southern parts of the lake. The lake showed great enlargements toward the west, south and southeast (Figure 4d). Rapid expansion occurred in 2008-2011, causing the area to enlarge to 18.467 km 2 . The intra-seasonal variability indicated that the smallest inundated area occurred in April and then the area extent gradually increased and reached the largest size in October. After that, the lake area remained stable and then began shrinking until the next February.
The other three lakes in the Inner Basin showed similar seasonal patterns during the studied period. April-July was the dry season when the area tended to be small, while the lakes were in better situations from September to the next March. Tangra-yumco showed an increasing pattern with fluctuations, and had a slight changing trend of 0.61 km 2 /yr. The maximum area was 848.33 km 2 in March 2011, and the minimum area was 792.52 km 2 in August 2001. Taro Co and Shuru Co had small increasing trends, as shown in Table 1.
Three lakes in the Brahmaputra Basin showed similar seasonal characteristics, while they presented big differences in their changing trends. Yanzho Yumco presented fluctuations and had an overall contraction. In 2000-2005, Yamzho Yumco presented a fluctuating expansion, and the area increased from 498.80 km 2 to 539.81 km 2 , at a rate of 7.61 km 2 /yr. Then, its area decreased from 513.08 to 485.99 km 2 , at a rate of -3.42 km 2 /yr in 2006-2015. The total rate of the decreasing trend was -3.27 km 2 /yr. The largest area was 587.09 km 2 in October 2004 and the smallest was 402.49 km 2 in June 2015.
As Yamzho Yumco is surrounded by mountains, its inundation extent was restricted by terrain, resulting in complex shapes with several tributaries (Figure 4e). Yumzho Yumco endured a retreat in all directions in 2015, and some of the tributaries dried up. In terms of seasonality, Yamzho Yumco remained stable in January-April, when it tended to be in frozen state. It began retreating from May and reached its minimum size in July. Then, it increased from August to October and began decreasing again in November. In general, Pegu Co remained stable with a slight decreasing trend. The maximum area of Pegu Co was 274.55 km 2 in November 2005, and the minimum area was 254.8 km 2 in April 2009. The inundation extent after 2006 was less than that of the early years. Smaller areas were usually present in July, and larger surface areas appeared in October. The small lake Puma Yumco showed no apparent change and was almost stable with an area of around 280 km 2 during the 16 years.
Two lakes in Indus Basin, Rakshastal and Puma Yumco endured slight retreats. Rakshastal had an annual decreasing rate of -0.32 km 2 /a. The minimum area was 233.  Figure 5 presents the lake level and volume changes of 12 lakes with a series of data collected during the studied period. Three lakes-Rakshastal, Puma Yumco, and Daggayai Co-did not have any matched records to create volume change curves. Seven lakes (lake IDs 13,14,15,2,3,4,and 5) were covered by three sources of water level data. Therefore, they had more level and volume records.

Accuracy Assessment
To check the accuracy of the lake surface results, six series of 30 m interpreted results of the studied lakes based on Landsat images from the years 2000, 2005, 2010, 2013, 2014, and 2015 were collected. The accuracy of the interpreted lake maps was 96% [29]. To ensure images were consistent with acquisition dates, MODIS results from the nearest dates to the 30 m Landsat results were chosen. Theoretically, for 15 lakes, 90 results could be selected in the six years for comparison. As there was some mismatch between the acquisition dates of the Landsat and MODIS images, 75 pairs of data were finally determined to assess the accuracy. Figure 6 shows that the correlation between MODIS results and interpretation results was high with an R 2 value of 0.99. The 30 m results were higher, as Landsat images can clearly delineate the trivial transition of coastlines. Boundary differences between the two sets of results for some respective lakes are shown in Figure 6. The area differences were between 2.83% and 3.56%, which meant the research results were convincing and can be used to study lake inundation changes. Water levels of more than 50 major lakes on the TP from 2000 to 2017 were presented in [30], and nine studied lakes of this research were included. Long time series of the water levels of the nine common lakes were respectively compared and the correlations were around 0.91-0.99. Usually, larger lakes had higher correlation coefficients. Moreover, daily gauged water level observations of Zhari-namco in 2010-2015 were obtained. The in situ data did not represent the real water level and only captured level fluctuations relative to the mean value of each year, around 0.1-0.8 m as shown in Figure 7. For Zhari-namco, the derived results and matched the in situ lake level fluctuations well. The lake level results in this paper had a high degree of accuracy and were able to reflect lake dynamics. Figure 7. Comparisons between studied results and gauged records of Zhari-namco.

Driving Forces
Based on climate data and collected materials, driving forces were analyzed on four expanded and one retreating lakes as their obvious changes.
The four expanded lakes-Nam Co, Selin Co, Zhari-namco, and Ngangze-were in small enclosed sub-basins, as shown in Figure 1. As the interference from human actions can be ignored, the water balance equation considering precipitation, ground runoff, evaporation, and other factors such as glacier and snow meltwater was established for them. The equation is as follows: where A is the area of the studied lake at time t, P is the corresponding precipitation, and thus A P is the equivalent volume of rainfall on the lake surface. E is the evaporation on the lake surface. R is the accumulated ground runoff that is determined as precipitation minus evaporation over land in the sub-basin. Regarding ground runoff, as there were several other lakes in each sub-basin, the ground runoff allocated to a studied lake was determined according to the area ratios among lakes in one sub-basin. The area ratios of the four expanded lakes-Nam Co, Selin Co, Ngangze, and Zharinamco-were 99.57%, 85.00%, 99.43%, and 85.70%, respectively. W is an overall supplement that is made up of other components: glacier meltwater, snow water, permafrost, and some groundwater outflow. When W is negative, it indicates that W is the outflow. V and V represent the lake storage at two moments. The net precipitation was defined as the precipitation minus evaporation over the lake surface, and then with the allocated ground runoff in the sub-basin (A P E R). In most years, the net precipitation was usually less than 0 due to the strong land evaporation on the southern TP. In addition, the infiltration of each lake was ignored.
Daily precipitation and evaporation measurements were transformed to the monthly frequency to match the water level and volume changes data. Then, the monthly precipitation in the whole study area was determined according to kriging interpolation of gauged observations on 19 nearby stations. The land evaporation of each sub-basin in 2000-2015 was extracted from GLDAS, which was on the monthly scale and had a resolution of 1°. The evaporation of each lake was estimated from the nearest in situ station according to the Penman-Monteith equation, which was illustrated in detail in [31]. Based on several variables and lake volume changes, the W of each lake was derived from the water balance equation.
For the four sub-basins, the precipitation showed no obvious changing patterns and the land evaporation increased with a slight trend during 2000-2015. The storages of the four lakes increased, while the net precipitation decreased over the studied period in Figure 8. This implied that besides precipitation supplements, the contribution of W increased. As Figure 8 shows, the annual average lake volume of Nam Co experienced an expansion with a rate of 0.29 km 3 /yr. The volume rose rapidly, increasing by 5.11 km 3 in 2000-2005 and then was maintained at around 5.71 km 3 after 2005. The land evaporation in its basin increased at a rate of 0.04 km 3 /yr and the highest was 3.60 km 3 in 2011. The contribution of the yearly net precipitation was less than 0. It fluctuated and decreased from -0.31 km 3 in 2000 to -3.27 km 3 in 2015 at a rate of -0.14 km 3 /yr. From the view of water balance, the yearly supplement of W increased from 1.00 to 3.00 km 3 during the study period, with a mean value of 1.68 km 3 in 2000-2015. The increased intensity of Zhari-namco was less than that of Selin Co. Zhari-namco had a slight change of 0.14 km 3 /yr, and the volume rose by 2.19 km 3 over the 16 year period. The net precipitation fluctuated with average and variation values of 1.66 km 3 and 0.72 km 3 , respectively. The highest net precipitation was 0.97 km 3 in 2002 and the lowest was −1.56 km 3 in 2015. Negative net precipitation indicated that evaporation induced outflow was more than precipitation induced inflow. As the lake storage fluctuated, the supplement of W was around −0.50 to 1.68 km 3 , with a mean value of 0.42 km 3 .
Ngangze also expanded gradually and the volume increased by 1.89 km 3 at a rate of 0.12 km 3 /yr. The net precipitation presented a decreasing tendency with a slight fluctuation in the years 2007, 2008, and 2014. The highest net precipitation was 0.56 km 3 in 2008, and the average annual value was -0.16 km 3 . As the average annual variation in the lake volume was 1.15 km 3 , the average yearly contribution of W was 0.32 km 3 during the study period.
Yearly quantitative supplements of several components for the four lakes are shown in Table 4. The evaporated lake water was usually 2.5-3.5 times that of lake surface rainfall, and precipitation induced runoff was the main cause for lake expansion. The low W of Selin Co implied that precipitation supplied most in its expansion and some sources of outflow, such as groundwater flow existed to counterbalance the feed. Yamzho Yumco is an inland lake, and its supplements are snow meltwater, spring water, and precipitation induced surface runoff. Lake surface evaporation is usually greater than precipitation. As there is a hydrological project in the basin, the water balance equation is not suitable for Yamzho Yumco. The analysis of the effects of climate variables on lake retreat, fluctuations of yearly precipitation, lake surface evaporation, and temperature, in 2000-2015, derived from five nearby metrological stations, are presented in Figure 9. Yamzho Yumco had a decrease in annual average volume from 1.79 km 3 in 2002 to 0.28 km 3 in 2010. The value of evaporation minus precipitation on the lake surface increased from 0.16 km 3 in 2000 to 0.48 km 3 in 2015 at a rate of 0.01 km 3 /yr. The annual inflow to Yamzho Yumco was 0.95 km 3 , which was nearly equal to the evaporation over the lake surface, and this was capable of allowing the lake to maintain a stable state. However, a hydropower station was established in the year 1998 with yearly lake water consumption of 0.18 km 3 . The project broke the lake balance and became a major contributor to lake retreat. Some research has pointed out that the water level increase before 2005 was mainly caused by precipitation and surface runoff, though human activities had a negative effect [32]. After 2005, fluctuations in precipitation, evaporation, and temperature were not in agreement with lake dynamic changes, and climate factors were not able to fully explain the cause of lake contraction. With increasing evaporation and aggravating influences of the hydropower project, YamzhoYumco experienced a drastic retreat. In addition to metrological factors, glacier/snow cover also contributed to lake volume variation. With an increasing temperature and the retreat of glaciers, potential supplements of meltwater will decrease, and then the lake will be at a risk of constant recession.

Conclusions
Several findings were obtained from the analysis of hydrological changes of 15 lakes on the southern TP. Lakes in the Inner basin usually underwent expansion with rates from 0.17 to 29.87 km 2 /yr between 2000 and 2015, especially Selin Co, Nam Co, Ngangze, and Zhari-namco. For the four lakes, with the aid of volume changes, contributions of the driving factors were derived based on the water balance equation. The precipitations in sub-basins presented no clear changes, while the land evaporations increased slightly. The negative effect of lake surface evaporation was almost 2.76-3.88 times that of lake surface precipitation. Surface runoff was the main cause of lake expansion, while it decreased over the studied period, and this implied that the contributions from other factors, such as snow/glacier meltwater, permafrost increased. Two lakes in the Indus Basin experienced retreats. Three lakes in the Brahmaputra Basin showed different changing trends. Yamzho Yumco endured a considerable retreat, and climate factors were not enough to explain this. Actually, the hydrological project has been a major cause of lake level decline since 2005. With increasing evaporation, decreasing glacier meltwater, and the constant effect of the hydropower station, the situation of Yamzho Yumco will get worse.
In this research, MODIS images were selected to document inundation changes as this instrument scanned the earth two times a day. Some other high-resolution measurements were discarded as they had low frequent observations, such as the widely applied Landsat images (30 m). Discarding the scenes contaminated by clouds or snow, the available Landsat images were few, lowing the density of lake storage data. The usable MODIS images can guarantee at least one measurement available in a month. As our aim was to study lake hydrological changes on annual and monthly scales, such frequency of observations was adequate. In addition, high frequency of lake area data derived from MODIS images can match more lake level data.
Regarding driving forces, some researches have tried to analyze the impact of climate and land cover change on lake fluctuations on the TP. However, only the correlation analysis was carried out between independent variables and hydrologic traits. Hydrologic models such as Soil and Water Assessment Tool (SWAT) and Variable Infiltration Capacity (VIC) [33,34] can simulate hydrological processes of lakes or basins and illustrate the driving forces of lake dynamics, while sufficient measured variables is a key prerequisite. Therefore, hydrological models are not practical on the TP due to the insufficient gauged measurements. In contrast, the water balance equation applied in this research can quantitatively derive contributions of precipitation, lake surface evaporation, land evaporation, surface runoff, and other factors. The results have implications to assess the effect of climate change and glacier retreat and it is possible to perform the same task in other enclosed basins.
In addition, three hydrological traits were presented with high levels of accuracy. Therefore, the procedure employed in this paper can be expanded to other lakes without in situ observations. The high frequency and long series of lake coastlines can be treated as contours of lake bathymetry. The lake's bottom topography can be derived based on these contours, especially for the lakes endured large dynamics, such as Selin Co. On the whole, this research obtained changing details and quantified potential contributors to lakes on the southern TP. The results can help us to better understand the regional water cycle and climate change on the TP that have occurred in this century and also can be a foundation for hydrologic modelling of the TP.

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