Long-Term Changes of Open-Surface Water Bodies in the Yangtze River Basin Based on the Google Earth Engine Cloud Platform

The spatiotemporal changes of open-surface water bodies in the Yangtze River Basin (YRB) have profound influences on sustainable economic development, and are also closely relevant to water scarcity in China. However, long-term changes of open-surface water bodies in the YRB have remained poorly characterized. Taking advantage of the Google Earth Engine (GEE) cloud platform, this study processed 75,593 scenes of Landsat images to investigate the long-term changes of open-surface water bodies in the YRB from 1984 to 2018. In this study, we adopted the percentile-based image composite method to collect training samples and proposed a multiple index water detection rule (MIWDR) to quickly extract the open-surface water bodies. The results indicated that (1) the MIWDR is suitable for the long-term and large-scale Landsat water bodies mapping, especially in the urban regions. (2) The areas of permanent water bodies and seasonal water bodies were 29,076.70 km2 and 21,526.24 km2, accounting for 57.46% and 42.54% of the total open-surface water bodies in the YRB, respectively. (3) The permanent water bodies in the YRB increased along with the decreases in the seasonal water bodies from 1984 to 2018. In general, the total open-surface surface water bodies in the YRB experienced an increasing trend, with an obvious spatial heterogeneity. (4) The changes of open-surface water bodies were associated with the climate changes and intense human activities in the YRB, however, the influences varied among different regions and need to be further investigated in the future.


Introduction
Water resources and services play a critical role in economic growth and environmental sustainability. As the essential components of water resources, open-surface water bodies, including lakes, reservoirs, rivers, streams and ponds, provide a series of ecosystem services, such as water supply and regulation, climate regulation, and food production [1,2]. However, from 1984 to 2015, permanent water bodies decreased by almost 90,000 km 2 due to climate changes and human activities [3]. China has the largest population in the world and has experienced fast development over the past years. Whereas, the uneven spatial distribution of water resources has been an important constraining factor which has made the water scarcity more severe in China, especially in the northern part of Thus, the objectives of this study are (1) to propose a new framework to quickly collect samples and detect open-surface water bodies in the GEE platform and (2) to analyze the long-term changes of the open-surface water bodies in the YRB from 1984 to 2018 based on the GEE platform and all available Landsat images. Furthermore, the study can reinforce the understanding of long-term changes of open-surface water bodies in the YRB and facilitate the development of water resource management in the YRB.

Study Area
The YRB is located in southern China (90-121°E and 24-36°N) ( Figure 1). It covers about 1.8 × 10 6 km 2 , representing one-fifth of China's land area. The YRB is characterized by complex geographical environments, with plateaus and mountains in the western region, basins, and hills in the middle part, and plains in the eastern area. In this study, the YRB was divided into four sub-basins: the source of the YRB (SYRB), the upper reaches of the YRB (UYRB), the middle reaches of the YRB (MYRB), and the lower reaches of the YRB (LYRB). The SYRB is located in the Tibetan Plateau, and the climate in SYRB is characterized by the low annual precipitation and cold average temperature. The URYB mainly consists of western mountains and eastern Sichuan Basin. The precipitation in the UYRB varies from 500 mm to the 1000 mm. The precipitation in MRYB and SYRB varies from 1000 mm to 1900 mm ( Figure S1). Apart from the western YRB in the Tibetan Plateau, the climate in the eastern UYRB, MYRB, and LYRB is mostly controlled by the Indian summer monsoon and the East Asian summer monsoon, which lead to the annual cycle of dry and wet seasons [42]. Overall, precipitations from May to October were estimated to be about 70%-90% of the annual total precipitation in the YRB [43].

Data
This study used all available Landsat 5, 7, and 8 surface reflectance images of the entire YRB in the GEE platform from 1984 to 2018. The Landsat 5 and 7 surface reflectance datasets were generated from the Landsat ecosystem disturbance adaptive processing system (LEDAPS) algorithm, and the Landsat 8 surface reflectance products were generated from the Landsat surface reflectance code (LaSRC) algorithm [44]. Six spectral bands, including blue band, green band, red band, near-infrared band, short-wave infrared band 1, and short-wave infrared band 2, had a spatial resolution of 30 m and were

Data
This study used all available Landsat 5, 7, and 8 surface reflectance images of the entire YRB in the GEE platform from 1984 to 2018. The Landsat 5 and 7 surface reflectance datasets were generated from the Landsat ecosystem disturbance adaptive processing system (LEDAPS) algorithm, and the Landsat 8 surface reflectance products were generated from the Landsat surface reflectance code (LaSRC) algorithm [44]. Six spectral bands, including blue band, green band, red band, near-infrared band, short-wave infrared band 1, and short-wave infrared band 2, had a spatial resolution of 30 m and were used to extract water bodies. The total number of these images was 75,593, including 35,642 Moreover, the Sentinel-2 images with a spatial resolution of 10 meters were used to evaluate the accuracy of the extracted water bodies in the YRB. PERSIANN-CDR daily precipitation products from 1984 to 2018 with a spatial resolution of 0.25° were used to analyze the temporal trend of precipitation in the YRB [45]. The JRC Global Surface Water Occurrence Layer was also used as ancillary data to filter the misclassifications, such as hill shadow [3]. These data can be accessed through the GEE's public data catalog. In addition, FROM-GLC data [46] and SRTM DEM data were collected to display the land cover and elevation in the YRB, respectively.

Sample Collection Based on Percentile Composite Images
A total of 14 out of 122 Landsat tiles that overlap with the YRB were first selected. The selected tiles were relatively evenly located in the YRB, with different land cover types ( Figure 3). Moreover, these tiles cover the typical open-surface water bodies, such as Dongting Lake, urban lakes, plateau Moreover, the Sentinel-2 images with a spatial resolution of 10 m were used to evaluate the accuracy of the extracted water bodies in the YRB. PERSIANN-CDR daily precipitation products from 1984 to 2018 with a spatial resolution of 0.25 • were used to analyze the temporal trend of precipitation in the YRB [45]. The JRC Global Surface Water Occurrence Layer was also used as ancillary data to filter the misclassifications, such as hill shadow [3]. These data can be accessed through the GEE's public data catalog. In addition, FROM-GLC data [46] and SRTM DEM data were collected to display the land cover and elevation in the YRB, respectively.

Sample Collection Based on Percentile Composite Images
A total of 14 out of 122 Landsat tiles that overlap with the YRB were first selected. The selected tiles were relatively evenly located in the YRB, with different land cover types ( Figure 3). Moreover, these tiles cover the typical open-surface water bodies, such as Dongting Lake, urban lakes, plateau lakes, and rivers. For each tile, all the images from 1984 to 2018 can be regarded as an image collection. Firstly, for each image in the image collection, the cloud, cloud shadow, and snow/ice were masked by Landsat quality band generated by the Fmask algorithm [47]. Each image has six Remote Sens. 2019, 11, 2213 5 of 22 spectral bands, including blue band, green band, red band, near-infrared band, short-wave infrared band 1, and short-wave infrared band 2. Then, the study calculated the percentile value of surface reflectance for each pixel based on the single spectral bands in the image collection. Finally, six single spectral bands were composited into a single image. The composited image was called the cloud-free percentile composite image. The percentile-based image composite method was easily conducted in the GEE platform through the built-in function "ee.Reducer.percentile". The surface reflectance exhibits intra-annual changes due to phenology and sun angle differences [48]. Therefore, the percentile composite method can be used to generate season-variation images. For example, Donchyts et al. [49] revealed that the percentile range of 15%-55% of all Landsat 8 top-of-atmosphere reflectance products from 2013-2015 was suitable for permanent water detection over the Murray-Darling Basin. In this study, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90% percentile composite images were calculated for each tile. As shown in Figure 3, the percentile images show the seasonal variations, and the use of lower percentiles has a higher chance to represent a larger amount of surface water bodies.
through the built-in function "ee.Reducer.percentile". The surface reflectance exhibits intra-annual changes due to phenology and sun angle differences [48]. Therefore, the percentile composite method can be used to generate season-variation images. For example, Donchyts et al. [49] revealed that the percentile range of 15%-55% of all Landsat 8 top-of-atmosphere reflectance products from 2013-2015 was suitable for permanent water detection over the Murray-Darling Basin. In this study, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, and 90% percentile composite images were calculated for each tile. As shown in Figure 3, the percentile images show the seasonal variations, and the use of lower percentiles has a higher chance to represent a larger amount of surface water bodies.
The 90% percentile composite image collection (14 Landsat tiles are regarded as an image collection) represented the minimum extent of surface water bodies and were used to randomly generate water samples. The water samples were expected to be identified as water bodies in the other eight percentile composite image collections. Moreover, the 10% percentile composite image collection indicated the minimum extent of non-water bodies, and was used to randomly generate non-water samples. The non-water samples were expected to be identified as non-water bodies in the other eight percentile composite image collections. Then, by further visual interpretation, the samples which had a stable type but a changing spectral reflectance were retained. Finally, for each percentile composite image collection, the study collected 10,665 samples, including 1427 water samples and 9238 non-water samples. Therefore, the 9 percentile image collections result in 95,985 samples, including 12,843 water samples and 83,142 non-water samples. The number of samples was enough to establish a water detection rule. . The spatial distribution of sampling plots and 10%, 50%, and 90% percentile composite images for three sampling plots. The number in sampling plots indicates the Landsat path (the first three digits) and row (The last three digits).
The 90% percentile composite image collection (14 Landsat tiles are regarded as an image collection) represented the minimum extent of surface water bodies and were used to randomly generate water samples. The water samples were expected to be identified as water bodies in the other eight percentile composite image collections. Moreover, the 10% percentile composite image collection indicated the minimum extent of non-water bodies, and was used to randomly generate non-water samples. The non-water samples were expected to be identified as non-water bodies in the other eight percentile composite image collections. Then, by further visual interpretation, the samples which had a stable type but a changing spectral reflectance were retained. Finally, for each percentile composite image collection, the study collected 10,665 samples, including 1427 water samples and 9238 non-water samples. Therefore, the 9 percentile image collections result in 95,985 samples, including 12,843 water samples and 83,142 non-water samples. The number of samples was enough to establish a water detection rule.

Water Extraction Based on Water Detection Rule
Based on Zou et al. [28], the study used NDWI, AWEI, NDVI, and EVI to develop a multiple index water detection rule (MIWDR). In particular, AWEI has two versions, including AWEI nsh and AWEI sh [32]. The AWEI nsh is suited for these areas where shadows are not a major problem. AWEI sh can effectively eliminate shadow pixels but misclassify highly reflective surfaces as water [32]. If a pixel meets the following criteria: (AWEI nsh -AWEI sh > −0.1) and (MNDWI > NDVI or MNDWI > EVI), then it was classified as water body. Figure 4 shows the scatter density plots of all samples. For water samples, only 1.21% were misclassified into non-water bodies. As for non-water samples, only 1.47% were misclassified into water bodies.
Remote Sens. 2019, 11, x FOR PEER REVIEW 6 of 22 Figure 3. The spatial distribution of sampling plots and 10%, 50%, and 90% percentile composite images for three sampling plots. The number in sampling plots indicates the Landsat path (the first three digits) and row (The last three digits).

Water Extraction Based on Water Detection Rule
Based on Zou et al. [28], the study used NDWI, AWEI, NDVI, and EVI to develop a multiple index water detection rule (MIWDR). In particular, AWEI has two versions, including AWEInsh and AWEIsh [32]. The AWEInsh is suited for these areas where shadows are not a major problem. AWEIsh can effectively eliminate shadow pixels but misclassify highly reflective surfaces as water [32]. If a pixel meets the following criteria: (AWEInsh-AWEIsh > -0.1) and (MNDWI > NDVI or MNDWI > EVI), then it was classified as water body. Figure 4 shows the scatter density plots of all samples. For water samples, only 1.21% were misclassified into non-water bodies. As for non-water samples, only 1.47% were misclassified into water bodies. In order to obtain better water bodies mapping in the YRB, we compared the MIWDR with the method in Zou et al. [28]. In general, both of two rules showed good performance on the open-surface water bodies, such as large lakes and rivers ( Figure S2). The MIWDR performed better than Zou et al. [28] in urban environments ( Figure S3). However, the rule of Zou et al. [28] performed better than the MIWDR for these water bodies in the plateau regions ( Figure S4). Therefore, we used the Landsat path as an index to determine the suitable method to extract open-surface water bodies. If the Landsat path of an image was less than 131, then the MIWDR was used. If the Landsat path of an image was greater than or equal to 131, then the water detection rule proposed by Zou et al. [28] was used. The Landsat path of 131 was chosen by the terrain in the YRB (Figures 1 and 2). Moreover, according to the scatter density plots of water samples ( Figure S5), the study also observed that the MIWDR In order to obtain better water bodies mapping in the YRB, we compared the MIWDR with the method in Zou et al. [28]. In general, both of two rules showed good performance on the open-surface water bodies, such as large lakes and rivers ( Figure S2). The MIWDR performed better than Zou et al. [28] in urban environments ( Figure S3). However, the rule of Zou et al. [28] performed better than the MIWDR for these water bodies in the plateau regions ( Figure S4). Therefore, we used the Landsat path as an index to determine the suitable method to extract open-surface water bodies. If the Landsat path of an image was less than 131, then the MIWDR was used. If the Landsat path of an image was greater than or equal to 131, then the water detection rule proposed by Zou et al. [28] was used. The Landsat path of 131 was chosen by the terrain in the YRB (Figures 1 and 2). Moreover, according to the scatter density plots of water samples ( Figure S5), the study also observed that the MIWDR performed better for these water samples in the low-altitude YRB (Landsat path < 131) than for these water samples in the high-altitude YRB (Landsat path ≥ 131).
The extraction of open-surface water bodies was supported by the GEE cloud platform. Firstly, according to the quality band of Landsat surface reflectance products generated from the Fmask algorithm, cloud, cloud shadow, and snow were masked. It is noted that the Landsat 7 images after 2003 had the scan line errors. Therefore, these pixels with scan line errors also were masked. The unmasked pixels were regarded as good observations. Then according to the Landsat path, the suitable water detection rule was applied into each image to obtain a preliminary mapping of open-surface water bodies. Lastly, the study used the JRC global surface water occurrence layer to filter hill shadow and other misclassifications. All in all, the workflow of this study is shown in Figure 5. performed better for these water samples in the low-altitude YRB (Landsat path < 131) than for these water samples in the high-altitude YRB (Landsat path ≥ 131).
The extraction of open-surface water bodies was supported by the GEE cloud platform. Firstly, according to the quality band of Landsat surface reflectance products generated from the Fmask algorithm, cloud, cloud shadow, and snow were masked. It is noted that the Landsat 7 images after 2003 had the scan line errors. Therefore, these pixels with scan line errors also were masked. The unmasked pixels were regarded as good observations. Then according to the Landsat path, the suitable water detection rule was applied into each image to obtain a preliminary mapping of opensurface water bodies. Lastly, the study used the JRC global surface water occurrence layer to filter hill shadow and other misclassifications. All in all, the workflow of this study is shown in Figure 5.

Accuracy Assessment Based on Sentinel-2 Images
This study used the Sentinel-2 images to evaluate the accuracy of the extracted open-surface water bodies in the YRB. Considering the strong dynamical variations in the extent of open-surface water bodies, the Landsat images and Sentinel images should have the same date, with a tolerance of no more than 2 days. Moreover, the cloud coverage of these images should be less than 15%. Finally, a total of 24 Sentinel images were used (Table S1). In the 24 images, 9000 test samples, including 5424 water samples and 3576 non-water samples, were randomly generated within the extent of the extracted Landsat open-surface water bodies and another 60 meters buffer which was created based on the boundary of the extracted Landsat open-surface water bodies. Finally, the study adopted a visual inspection to check the test samples.

Accuracy Assessment Based on Sentinel-2 Images
This study used the Sentinel-2 images to evaluate the accuracy of the extracted open-surface water bodies in the YRB. Considering the strong dynamical variations in the extent of open-surface water bodies, the Landsat images and Sentinel images should have the same date, with a tolerance of no more than 2 days. Moreover, the cloud coverage of these images should be less than 15%. Finally, a total of 24 Sentinel images were used (Table S1). In the 24 images, 9000 test samples, including 5424 water samples and 3576 non-water samples, were randomly generated within the extent of the extracted Landsat open-surface water bodies and another 60 m buffer which was created based on the boundary

Change Analysis of an Open-Surface Water Body
According to the long-term open-surface water bodies in the YRB, the water inundation frequency (WIF) of each pixel can be calculated from the following: where N indicates the number of all the good observations in a specific period, and w is a binary variable (w = 1 indicates water body and w = 0 indicates non-water body). WIF ranges from 0% to 100%. According to WIF, open-surface water bodies in the YRB can be divided into two types: permanent water bodies (WIF > 75%) and seasonal water bodies (25% < WIF ≤ 75%) [18,28]. Moreover, the maximum water bodies in one year were defined as the collection of the pixels with WIF > 0. Here WIF was calculated for each year. Table 1 shows the accuracy of extracted open-surface water bodies in the YRB. The overall accuracy and kappa coefficient was 95.03% and 0.895, respectively (Table 1), which indicated that the detected water products in the YRB had higher accuracy and then can be used for further analysis.

Spatial Distribution of the Open-Surface Water Bodies in the YRB
The open-surface water bodies in the YRB showed a clearly uneven spatial distribution ( Figure 6 and Table 2). About 49.99% of water bodies in the YRB were located in the MYRB, followed by 28.88% in the LYRB. A total of 13.25% and 7.87% of water bodies in the YRB were distributed in the UYRB and SYRB, respectively. In general, the areas of permanent water bodies and seasonal water bodies were 29,076.70 km 2 and 21,526.24 km 2 , accounting for 57.46% and 42.54% of the total surface water bodies in YRB, respectively. The results indicated that the YRB was dominated by permanent water bodies. However, it can be observed that the LYRB had far more permanent water bodies than seasonal water bodies. The ratio of permanent water bodies to seasonal water bodies in the LYRB was 2.16, and the corresponding ratio for the SYRB, UYRB, and MYRB were 1.38, 1.21, and 1.06, respectively.

Monthly Variations of the Surface Water Bodies in the YRB
For each climatological month, the study also investigated the temporal dynamics of permanent and seasonal water bodies ( Figure 7). The open-surface water bodies in the YRB showed a clear seasonality. The areas in wet months were obviously greater than those in dry months. Especially, the minimum extent of open-surface water bodies occurred in January, with a total area of 41,901.03 km 2 . The maximum extent of open-surface water bodies occurred in July, with a total area of 50,927.08 km 2 . If from the perspective of different four sub-basin, we can find similar characteristics of monthly variation ( Figure S6). (a-f) indicates the source regions of the YRB, Wuhan city, Taihu Lake, the Three Gorges Dam (TGD), Dongting Lake, and Poyang Lake, respectively.

Monthly Variations of the Surface Water Bodies in the YRB
For each climatological month, the study also investigated the temporal dynamics of permanent and seasonal water bodies (Figure 7). The open-surface water bodies in the YRB showed a clear seasonality. The areas in wet months were obviously greater than those in dry months. Especially, the minimum extent of open-surface water bodies occurred in January, with a total area of 41,901.03 km 2 . The maximum extent of open-surface water bodies occurred in July, with a total area of 50,927.08 km 2 . If from the perspective of different four sub-basin, we can find similar characteristics of monthly variation ( Figure S6). Remote Sens. 2019, 11, x FOR PEER REVIEW 10 of 22  Moreover, we also investigated the temporal trends of open-surface water bodies in the four sub-basins ( Figure 8). From 1984 to 2018, the seasonal water bodies in the SYRB had an increasing trend, while they showed a decreasing trend in the UYRB, MYRB, and LYRB. Especially, the MRYB experienced the most losses (1192.42 km 2 ), followed by the UYRB (279.83 km 2 ) and the LYRB (119.13 km 2 ). When it comes to permanent water bodies, all the sub-basins revealed an increasing trend. The increased areas for the SYRB, UYRB, MYRB, and LRB were 594.15 km 2 , 1144.25 km 2 , 522.15 km 2, and 471.79 km 2 , respectively. In general, the total open-surface water bodies in the SYRB, UYRB, and LYRB experienced an increasing trend, with an increasing percentage of 19.38%, 14.78%, and 2.53%. At the same time, the total open-surface water bodies in the MYRB suffered from a decreasing trend, with a decreasing percentage of 2.71%.   Moreover, we also investigated the temporal trends of open-surface water bodies in the four sub-basins ( Figure 8). From 1984 to 2018, the seasonal water bodies in the SYRB had an increasing trend, while they showed a decreasing trend in the UYRB, MYRB, and LYRB. Especially, the MRYB experienced the most losses (1192.42 km 2 ), followed by the UYRB (279.83 km 2 ) and the LYRB (119.13 km 2 ). When it comes to permanent water bodies, all the sub-basins revealed an increasing trend. The increased areas for the SYRB, UYRB, MYRB, and LRB were 594.15 km 2 , 1144.25 km 2 , 522.15 km 2, and 471.79 km 2 , respectively. In general, the total open-surface water bodies in the SYRB, UYRB, and LYRB experienced an increasing trend, with an increasing percentage of 19.38%, 14.78%, and 2.53%. At the same time, the total open-surface water bodies in the MYRB suffered from a decreasing trend, with a decreasing percentage of 2.71%.  Figure 9 further shows the conversion among non-water bodies, seasonal water bodies and permanent water bodies in the YRB. During the first two stages (1984-1999 and 2000-2009), the major increases in seasonal water bodies resulted from the conversion from non-water bodies with the conversion area of 4502.54 km 2 . Moreover, 2848.72 km 2 permanent water bodies were converted into seasonal water bodies. At the same time, the major increases in permanent water bodies were caused by the conversion from seasonal water bodies with the conversion area of 2848.73 km 2 . Only 924.89 km 2 non-water bodies were converted into permanent water bodies. During the last two phases (2000-2009 and 2010-2018), the most increases in permanent water bodies were still caused by the conversion from seasonal water bodies, with a conversion area of 4105.51 km 2 . However, the more seasonal water bodies (4303.78 km 2 ) were converted into non-water bodies. A large amount of conversion in seasonal water bodies led to the decreases in the seasonal water bodies, though there were 6260.93 km 2 new seasonal water bodies. Overall, from 1984 to 2018, the areas of new permanent water bodies were 7090.15 km 2 , and 59.87% of the increases in new permanent water bodies were caused by the conversion from seasonal water bodies. The areas of disappeared permanent water bodies were 4359.02 km 2 , and 77.01% of the decreases in permanent water bodies were converted into seasonal water bodies. Moreover, the areas of new seasonal water bodies were 8015.61 km 2 , and 58.09% of the increases in new seasonal water bodies resulted from the conversion from non-water bodies. The areas of disappeared seasonal water bodies were 9577.67 km 2 , and 55.68% of the decreases in seasonal water bodies were converted into non-water bodies.  Figure 9 further shows the conversion among non-water bodies, seasonal water bodies and permanent water bodies in the YRB. During the first two stages (1984-1999 and 2000-2009), the major increases in seasonal water bodies resulted from the conversion from non-water bodies with the conversion area of 4502.54 km 2 . Moreover, 2848.72 km 2 permanent water bodies were converted into seasonal water bodies. At the same time, the major increases in permanent water bodies were caused by the conversion from seasonal water bodies with the conversion area of 2848.73 km 2 . Only 924.89 km 2 non-water bodies were converted into permanent water bodies. During the last two phases (2000-2009 and 2010-2018), the most increases in permanent water bodies were still caused by the conversion from seasonal water bodies, with a conversion area of 4105.51 km 2 . However, the more seasonal water bodies (4303.78 km 2 ) were converted into non-water bodies. A large amount of conversion in seasonal water bodies led to the decreases in the seasonal water bodies, though there were 6260.93 km 2 new seasonal water bodies. Overall, from 1984 to 2018, the areas of new permanent water bodies were 7090.15 km 2 , and 59.87% of the increases in new permanent water bodies were caused by the conversion from seasonal water bodies. The areas of disappeared permanent water bodies were 4359.02 km 2 , and 77.01% of the decreases in permanent water bodies were converted into seasonal water bodies. Moreover, the areas of new seasonal water bodies were 8015.61 km 2 , and 58.09% of the increases in new seasonal water bodies resulted from the conversion from non-water bodies. The areas of disappeared seasonal water bodies were 9577.67 km 2 , and 55.68% of the decreases in seasonal water bodies were converted into non-water bodies.  (Figures 10 and 11). In the SYRB, the conversion areas from non-water bodies to seasonal water bodies were 568.19 km 2 , and accounted for the largest proportion, while the conversion areas from permanent water bodies to seasonal water bodies were 122.28 km 2 , and took up the smallest proportion. In the UYRB, the conversion from non-water bodies to permanent water bodies was the most frequent, with the conversion areas of 867.21 km 2 . However, the conversion from permanent water bodies to non-water bodies was the least frequent, with the conversion areas of 88.33 km 2 . As for the MYRB, the conversion from permanent water bodies to nonwater bodies still was the least frequent, with the conversion areas of 413.58 km 2 . However, the most frequent conversion was the conversion from seasonal water bodies to non-water bodies, with the conversion area of 3079.40 km 2 . When it comes to the LYRB, the conversion areas from non-water bodies to seasonal water bodies were 1461.24 km 2 , and accounted for the largest proportion, while the conversion areas from permanent water bodies to non-water bodies were 324.22 km 2 , and took up the smallest proportion. Overall, the conversions of open-surface water bodies in the YRB revealed the obvious spatial heterogeneity.   (Figures 10 and 11). In the SYRB, the conversion areas from non-water bodies to seasonal water bodies were 568.19 km 2 , and accounted for the largest proportion, while the conversion areas from permanent water bodies to seasonal water bodies were 122.28 km 2 , and took up the smallest proportion. In the UYRB, the conversion from non-water bodies to permanent water bodies was the most frequent, with the conversion areas of 867.21 km 2 . However, the conversion from permanent water bodies to non-water bodies was the least frequent, with the conversion areas of 88.33 km 2 . As for the MYRB, the conversion from permanent water bodies to non-water bodies still was the least frequent, with the conversion areas of 413.58 km 2 . However, the most frequent conversion was the conversion from seasonal water bodies to non-water bodies, with the conversion area of 3079.40 km 2 . When it comes to the LYRB, the conversion areas from non-water bodies to seasonal water bodies were 1461.24 km 2 , and accounted for the largest proportion, while the conversion areas from permanent water bodies to non-water bodies were 324.22 km 2 , and took up the smallest proportion. Overall, the conversions of open-surface water bodies in the YRB revealed the obvious spatial heterogeneity.  (Figures 10 and 11). In the SYRB, the conversion areas from non-water bodies to seasonal water bodies were 568.19 km 2 , and accounted for the largest proportion, while the conversion areas from permanent water bodies to seasonal water bodies were 122.28 km 2 , and took up the smallest proportion. In the UYRB, the conversion from non-water bodies to permanent water bodies was the most frequent, with the conversion areas of 867.21 km 2 . However, the conversion from permanent water bodies to non-water bodies was the least frequent, with the conversion areas of 88.33 km 2 . As for the MYRB, the conversion from permanent water bodies to nonwater bodies still was the least frequent, with the conversion areas of 413.58 km 2 . However, the most frequent conversion was the conversion from seasonal water bodies to non-water bodies, with the conversion area of 3079.40 km 2 . When it comes to the LYRB, the conversion areas from non-water bodies to seasonal water bodies were 1461.24 km 2 , and accounted for the largest proportion, while the conversion areas from permanent water bodies to non-water bodies were 324.22 km 2 , and took up the smallest proportion. Overall, the conversions of open-surface water bodies in the YRB revealed the obvious spatial heterogeneity.   Figure 12 shows the time series of yearly maximum surface water bodies and annual precipitation in the YRB from 1984 to 2018. Moreover, the study estimated the temporal trends of maximum water bodies and annual precipitation, and calculated the correlation coefficient between them (Table 4). During the period of 1984-1999, the maximum water bodies showed the increasing trend, though the trend was not significant (P-value = 0.126). At the same time, the precipitation also showed a significantly increasing trend at the significance level of 0.1 (P-value = 0.052). During this period, the correlation coefficient between the area and precipitation was 0.635 (P-value = 0.019). Moreover, the study observed the largest area of maximum water bodies occurred in 1998, when the annual precipitation reached the peak in this period. During the period of 2000-2009, the maximum water bodies and annual precipitation exhibited the decreasing trend. Especially, the trend for the former was significant at the significance level of 0.1 (P-value = 0.092). In this period, the correlation coefficient decreased to 0.569 (P-value =0.086). In this period, the peak of maximum water bodies and precipitation appeared in 2002. During the period of 2010-2018, the maximum water bodies and annual precipitation kept a slightly increasing trend. In this period, the correlation coefficient further decreased to 0.221 (P-value = 0.568). In this period, the peak of maximum water bodies and precipitation occurred simultaneously in 2010 and 2016. (a-f) indicates the source regions of the YRB, Wuhan city, Taihu Lake, the Three Gorges Dam (TGD), Dongting Lake and Poyang Lake, respectively. Figure 12 shows the time series of yearly maximum surface water bodies and annual precipitation in the YRB from 1984 to 2018. Moreover, the study estimated the temporal trends of maximum water bodies and annual precipitation, and calculated the correlation coefficient between them (Table 4). During the period of 1984-1999, the maximum water bodies showed the increasing trend, though the trend was not significant (P-value = 0.126). At the same time, the precipitation also showed a significantly increasing trend at the significance level of 0.1 (P-value = 0.052). During this period, the correlation coefficient between the area and precipitation was 0.635 (P-value = 0.019). Moreover, the study observed the largest area of maximum water bodies occurred in 1998, when the annual precipitation reached the peak in this period. During the period of 2000-2009, the maximum water bodies and annual precipitation exhibited the decreasing trend. Especially, the trend for the former was significant at the significance level of 0.1 (P-value = 0.092). In this period, the correlation coefficient decreased to 0.569 (P-value =0.086). In this period, the peak of maximum water bodies and precipitation appeared in 2002. During the period of 2010-2018, the maximum water bodies and annual precipitation kept a slightly increasing trend. In this period, the correlation coefficient further decreased to 0.221 (P-value = 0.568). In this period, the peak of maximum water bodies and precipitation occurred simultaneously in 2010 and 2016. Remote Sens. 2019, 11, x FOR PEER REVIEW 14 of 22  Table 4 The trend and correlation in maximum water bodies and precipitation in the YRB.

Changes in Open-Surface Water Bodies due to Climate and Human Factors
The maximum-water bodies had a good correlation with annual precipitation, especially in the period of 1984-1999 and 2000-2010. The results indicated the precipitation was the important factor that controls the open-surface water bodies in the YRB. Moreover, for the whole YRB, the precipitation showed a slightly increasing trend from 1984 to 2018 (Table 4). More precipitation suggested that there were more chances for the conversions form non-water bodies/seasonal water bodies to permanent water bodies. Therefore, the permanent water bodies in the YRB increased along with the decreases in the seasonal water bodies from 1984 to 2018. However, the precipitation trend varied much in the YRB (Figure 13). The SYRB and LYRB experienced the increasing precipitation, which suggested the water bodies increased. However, the MYRB suffered from the decreasing

Changes in Open-Surface Water Bodies due to Climate and Human Factors
The maximum-water bodies had a good correlation with annual precipitation, especially in the period of 1984-1999 and 2000-2010. The results indicated the precipitation was the important factor that controls the open-surface water bodies in the YRB. Moreover, for the whole YRB, the precipitation showed a slightly increasing trend from 1984 to 2018 (Table 4). More precipitation suggested that there were more chances for the conversions form non-water bodies/seasonal water bodies to permanent water bodies. Therefore, the permanent water bodies in the YRB increased along with the decreases in the seasonal water bodies from 1984 to 2018. However, the precipitation trend varied much in the YRB (Figure 13). The SYRB and LYRB experienced the increasing precipitation, which suggested the water bodies increased. However, the MYRB suffered from the decreasing precipitation, especially in the Dongting Lake Basin, which indicated the water bodies decreased ( Figure S7). Moreover, there is no doubt that precipitation was not the only climate factor. For example, many studies have indicated that the increasing glacier meltwater has led to the expansion of the lakes in the Tibetan Plateau (including the SYRB) due to the warming temperature [50][51][52]. Therefore, both the seasonal water bodies and permanent water bodies in the SYRB increased from 1984 to 2018. precipitation, especially in the Dongting Lake Basin, which indicated the water bodies decreased ( Figure S7). Moreover, there is no doubt that precipitation was not the only climate factor. For example, many studies have indicated that the increasing glacier meltwater has led to the expansion of the lakes in the Tibetan Plateau (including the SYRB) due to the warming temperature [50][51][52]. Therefore, both the seasonal water bodies and permanent water bodies in the SYRB increased from 1984 to 2018. In this study, we observed that the correlation coefficient between maximum water bodies and annual precipitation decreased from 1984 to 2018. The possible reason was the fact that the influence of human activities in the YRB has increased over the past years. Many studies have demonstrated the influence of human activities on the surface water bodies [37,53,54]. YRB experienced intense human activities over the past years, which led to a complicated influence on the surface water bodies. On the one hand, the surface water bodies were encroached on by other land use types because of urban expansion, agricultural production, and so on. For instance, Deng et al. [55] found the urban expansions caused the decrease in the lake area in the Wuhan urban agglomeration. The negative impacts of human activities on the surface water bodies were clearly observed in Wuhan city, in the MYRB (Figure 14). Whereas, on the other hand, human activities can also contribute to more surface water bodies. For example, in the UYRB, the permanent water bodies experienced substantial increases. One important factor was the Three Gorges Dam (TGD). We compared the surface water bodies in the Three Gorges Reservoir Area before (1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002) and after (2003-2018) the TGD and found the permanent water bodies increased by 307.33 km 2 , which demonstrated the huge influences of human activities on the increases of surface water bodies in the UYRB ( Figure 15). Moreover, with more and more attention to water protection for key lakes and reservoirs from China during the recent ten years, the water bodies were expected to increase [18].  In this study, we observed that the correlation coefficient between maximum water bodies and annual precipitation decreased from 1984 to 2018. The possible reason was the fact that the influence of human activities in the YRB has increased over the past years. Many studies have demonstrated the influence of human activities on the surface water bodies [37,53,54]. YRB experienced intense human activities over the past years, which led to a complicated influence on the surface water bodies. On the one hand, the surface water bodies were encroached on by other land use types because of urban expansion, agricultural production, and so on. For instance, Deng et al. [55] found the urban expansions caused the decrease in the lake area in the Wuhan urban agglomeration. The negative impacts of human activities on the surface water bodies were clearly observed in Wuhan city, in the MYRB ( Figure 14). Whereas, on the other hand, human activities can also contribute to more surface water bodies. For example, in the UYRB, the permanent water bodies experienced substantial increases. One important factor was the Three Gorges Dam (TGD). We compared the surface water bodies in the Three Gorges Reservoir Area before (1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002) and after (2003-2018) the TGD and found the permanent water bodies increased by 307.33 km 2 , which demonstrated the huge influences of human activities on the increases of surface water bodies in the UYRB ( Figure 15). Moreover, with more and more attention to water protection for key lakes and reservoirs from China during the recent ten years, the water bodies were expected to increase [18]. precipitation, especially in the Dongting Lake Basin, which indicated the water bodies decreased ( Figure S7). Moreover, there is no doubt that precipitation was not the only climate factor. For example, many studies have indicated that the increasing glacier meltwater has led to the expansion of the lakes in the Tibetan Plateau (including the SYRB) due to the warming temperature [50][51][52]. Therefore, both the seasonal water bodies and permanent water bodies in the SYRB increased from 1984 to 2018. In this study, we observed that the correlation coefficient between maximum water bodies and annual precipitation decreased from 1984 to 2018. The possible reason was the fact that the influence of human activities in the YRB has increased over the past years. Many studies have demonstrated the influence of human activities on the surface water bodies [37,53,54]. YRB experienced intense human activities over the past years, which led to a complicated influence on the surface water bodies. On the one hand, the surface water bodies were encroached on by other land use types because of urban expansion, agricultural production, and so on. For instance, Deng et al. [55] found the urban expansions caused the decrease in the lake area in the Wuhan urban agglomeration. The negative impacts of human activities on the surface water bodies were clearly observed in Wuhan city, in the MYRB ( Figure 14). Whereas, on the other hand, human activities can also contribute to more surface water bodies. For example, in the UYRB, the permanent water bodies experienced substantial increases. One important factor was the Three Gorges Dam (TGD). We compared the surface water bodies in the Three Gorges Reservoir Area before (1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002) and after (2003-2018) the TGD and found the permanent water bodies increased by 307.33 km 2 , which demonstrated the huge influences of human activities on the increases of surface water bodies in the UYRB ( Figure 15). Moreover, with more and more attention to water protection for key lakes and reservoirs from China during the recent ten years, the water bodies were expected to increase [18].  This study did not examine more detailed reasons for the long-term changes of open-surface water bodies in the YRB. In the next work, we will focus on the important regions, including two large lakes (Dongting Lake and Poyang Lake) and Wuhan urban agglomeration. We will investigate the changes and explore reasons by using more economic and social data.

Long-Term and Large-Scale Mapping of Open-Surface Water Bodies by GEE Platform
For the long-term and large-scale classification study, the collection of samples is a primary work. The samples should contain different years, seasons, and regions to reduce the possible sample bias. For this purpose, previous studies often chose the images and collected the samples from different years, seasons, and regions. For example, Zou et al. [28] collected the training samples from the 16 sampling plots of different years and regions. However, for each sampling plot, it just contains one year. Therefore, possible sample bias would occur. Moreover, this process is also relatively timeconsuming. In this study, with the help of the high performance on the image processing in the GEE platform, we used the cloud-free percentile composite images as the sampling plot and collected the water and non-water samples from the sampling plot. This process is easy to operate in the GEE and can obtain more representative samples. We also thought the cloud-free percentile composite images can be used to collect samples for classifying other land-use types.
Moreover, this study proposed a new water detection rule called MIWDR based on the existed vegetation indices (NDVI and EVI) and water indices (AWEInsh and AWEIsh). Compared with the water detection rule proposed by Zou et al. [28], the MIWDR in the study can alleviate effectively the impacts of urban building, although it leads to some narrow urban rivers were misclassified. However, we thought the MIWDR is suitable for long-term and large-scale water bodies mapping in the urban regions with a lower threshold for the condition (AWEInsh-AWEIsh) ( Figure 16). The MIWDR had relatively low accuracy for these water bodies in the plateau region of the YRB ( Figure  S3). Therefore, the study adopted a strategy using the different methods to extract water bodies from This study did not examine more detailed reasons for the long-term changes of open-surface water bodies in the YRB. In the next work, we will focus on the important regions, including two large lakes (Dongting Lake and Poyang Lake) and Wuhan urban agglomeration. We will investigate the changes and explore reasons by using more economic and social data.

Long-Term and Large-Scale Mapping of Open-Surface Water Bodies by GEE Platform
For the long-term and large-scale classification study, the collection of samples is a primary work. The samples should contain different years, seasons, and regions to reduce the possible sample bias. For this purpose, previous studies often chose the images and collected the samples from different years, seasons, and regions. For example, Zou et al. [28] collected the training samples from the 16 sampling plots of different years and regions. However, for each sampling plot, it just contains one year. Therefore, possible sample bias would occur. Moreover, this process is also relatively time-consuming. In this study, with the help of the high performance on the image processing in the GEE platform, we used the cloud-free percentile composite images as the sampling plot and collected the water and non-water samples from the sampling plot. This process is easy to operate in the GEE and can obtain more representative samples. We also thought the cloud-free percentile composite images can be used to collect samples for classifying other land-use types.
Moreover, this study proposed a new water detection rule called MIWDR based on the existed vegetation indices (NDVI and EVI) and water indices (AWEI nsh and AWEI sh ). Compared with the water detection rule proposed by Zou et al. [28], the MIWDR in the study can alleviate effectively the impacts of urban building, although it leads to some narrow urban rivers were misclassified. However, we thought the MIWDR is suitable for long-term and large-scale water bodies mapping in the urban regions with a lower threshold for the condition (AWEI nsh -AWEI sh ) ( Figure 16). The MIWDR had relatively low accuracy for these water bodies in the plateau region of the YRB ( Figure S3). Therefore, the study adopted a strategy using the different methods to extract water bodies from different areas in the YRB. In conclusion, for the long-term mapping of surface water bodies at more large scale, such as continent-scale, even global scale, the situation will be more complicated, there is still much work to do.
Remote Sens. 2019, 11, x FOR PEER REVIEW 17 of 22 different areas in the YRB. In conclusion, for the long-term mapping of surface water bodies at more large scale, such as continent-scale, even global scale, the situation will be more complicated, there is still much work to do. Figure 16. Comparisons of the water identification rule of this study and Zou et al. [28] in urban regions. (a-k) indicates the extracted water bodies according to the method of this study in Shanghai, Nanjing, Suzhou, Hangzhou, Nanchang, Changsha, Wuhan, Chongqing, Chengdu, Guizhou, and Kunming, respectively. (i-v) indicates the resulsts according to the method of Zou et al. [28] in the same cities with (a-k).

Uncertainty and Prospects
This study explored the spatiotemporal changes of open-surface water bodies in the YRB from 1984-2018 using all the available Landsat images. However, the number of image showed the remarkable spatial and temporal variations in the YRB, which resulted in a negative influence on the study results. The number of Landsat images during the period of [1984][1985][1986][1987][1988][1989][1990][1991][1992][1993][1994][1995][1996][1997][1998][1999][2000][2001][2002][2003][2004][2005][2006][2007][2008][2009], and 2010-2018 were 18,437, 29,696, and 27,460, respectively. The total number during the first period was much less than that during the following two periods. Therefore, we observed that the areas of maximum water bodies during the first period were far lower than those for the latter two periods. In 2012, Landsat 5 terminated, so the number of images in 2012 decreased. We calculated the correlation coefficient between maximum water bodies and annual precipitation under two scenarios: one included the images of 2012 and another did not include the images of 2012 (Table 4). The comparison between two scenarios indicated the correlation coefficient increased without the images of 2012. The quantitative results further demonstrated the influence of the number of images.
Moreover, we also investigated the changes in water bodies in the YRB from 1984 to 2018 based on the concept of WIF. Firstly, from the perspective of monthly variations, the amount of Landsat images from May to September is the least, while the area of open-surface water bodies was still the largest. These results indicated the monthly variations of open-surface water bodies in the YRB was reliable since the precipitations from May to October are estimated to be about 70%-90% of the annual total precipitation in the YRB [43]. Regarding the temporal trends, the study found the decreasing lake areas in Poyang Lake and Dongting Lake ( Figure S7) from 1984 to 2018, which was in line with previous studies [18,54]. The reasonable reason is the fact that the total number of images during each month is enough to obtain a relatively reliable WIF. However, the temporal changes of open-surface water bodies in the YRB from 1984 to 2018 still showed the uncertainty, especially for the region with fewer images.
Furthermore, we compared the spatial pattern of permanent and seasonal water bodies of this study and the JRC water product ( Figure 17). In general, the two products have a similar spatial pattern. The total permanent water bodies in the YRB based on JRC product were 30,610.79 km 2 , which was 1534.09 km 2 more than this study. The total seasonal water bodies in the YRB based on JRC product were 27,007.51 km 2 , which was 5481.27.09 km 2 more than this study. The reason is that the WIF of JRC product was manually adjusted, therefore the WIF of some pixels was larger than those of this study. Therefore, using the same threshold (25% and 75%) to classify the water bodies in the YRB would result in more water bodies than this study. maximum water bodies during the first period were far lower than those for the latter two periods. In 2012, Landsat 5 terminated, so the number of images in 2012 decreased. We calculated the correlation coefficient between maximum water bodies and annual precipitation under two scenarios: one included the images of 2012 and another did not include the images of 2012 (Table 4). The comparison between two scenarios indicated the correlation coefficient increased without the images of 2012. The quantitative results further demonstrated the influence of the number of images. Moreover, we also investigated the changes in water bodies in the YRB from 1984 to 2018 based on the concept of WIF. Firstly, from the perspective of monthly variations, the amount of Landsat images from May to September is the least, while the area of open-surface water bodies was still the largest. These results indicated the monthly variations of open-surface water bodies in the YRB was reliable since the precipitations from May to October are estimated to be about 70%-90% of the annual total precipitation in the YRB [43]. Regarding the temporal trends, the study found the decreasing lake areas in Poyang Lake and Dongting Lake ( Figure S7) from 1984 to 2018, which was in line with previous studies [18,54]. The reasonable reason is the fact that the total number of images during each month is enough to obtain a relatively reliable WIF. However, the temporal changes of open-surface water bodies in the YRB from 1984 to 2018 still showed the uncertainty, especially for the region with fewer images.
Furthermore, we compared the spatial pattern of permanent and seasonal water bodies of this study and the JRC water product ( Figure 17). In general, the two products have a similar spatial pattern. The total permanent water bodies in the YRB based on JRC product were 30,610.79 km 2 , which was 1534.09 km 2 more than this study. The total seasonal water bodies in the YRB based on JRC product were 27,007.51 km 2 , which was 5481.27.09 km 2 more than this study. The reason is that the WIF of JRC product was manually adjusted, therefore the WIF of some pixels was larger than those of this study. Therefore, using the same threshold (25% and 75%) to classify the water bodies in the YRB would result in more water bodies than this study. Figure 17. Comparisons of the spatial pattern of permanent and seasonal water bodies of this study and JRC water product. "Agreement" indicated both the water mapping have the same water type. In contrast, "Disagreement" indicated the two water mapping having different water types. (a-f) Figure 17. Comparisons of the spatial pattern of permanent and seasonal water bodies of this study and JRC water product. "Agreement" indicated both the water mapping have the same water type. In contrast, "Disagreement" indicated the two water mapping having different water types. (a-f) indicates the source regions of the YRB, Wuhan city, Taihu Lake, the Three Gorges Dam (TGD), Dongting Lake and Poyang Lake, respectively. GEE platform provided fast access to MODIS, Landsat, and Sentinel images. Merging the multi-source remote sensing data to investigate the changes of open-surface water bodies in the GEE platform can efficiently reduce the uncertainty, and it is also a hot topic.

Conclusions
In this study, a framework for long-term mapping of open-surface water bodies in the GEE platform was proposed. It includes a percentile-based image composite method to collect the training samples and a new water detection rule called MIWDR to quickly extract the open-surface water bodies. The percentile-based image composite method can obtain more representative samples for a classification model and can be used to map other land-use types. The MIWDR is based on the existing water index and vegetation index, and is suitable for the mapping of the long-term and large-scale Landsat water bodies in the urban regions.
The study investigated the long-term changes of open-surface water bodies in the YRB from 1984 to 2018 based on all available Landsat 5/7/8 images in the GEE platform. The open surface water bodies were classified into permanent water bodies and seasonal water bodies according to the inundation frequency. The YRB was dominated by the permanent water bodies, especially in the LYRB. With the increasing precipitation and intense human activities, the permanent water bodies in the YRB increased along with the decreases in the seasonal water bodies from 1984 to 2018. However, the long-term changes of open-surface water bodies in the YRB revealed the obvious spatial heterogeneity. For the permanent water bodies, all the sub-basins (SYRB, UYRB, MYRB, and LYRB) revealed an increasing trend, especially for SYRB and UYRB. Whereas, when it comes to seasonal water bodies, SYRB had an increasing trend, while UYRB, MYRB, and LYRB showed a decreasing trend. Moreover, the most frequent conversion was different among the SYRB (from non-water bodies to seasonal water bodies), UYRB (from non-water bodies to permanent water bodies), MYRB (from seasonal water bodies to non-water bodies), and LYRB (from non-water bodies to seasonal water bodies).

Supplementary Materials:
The following are available online at http://www.mdpi.com/2072-4292/11/19/2213/s1, Figure S1: Spatial distribution of the average annual precipitation in the YRB. Figure S2: Comparisons of the water detection rule of this study and Zou et al. [28] in the non-urban regions. Figure S3: Comparisons of the water detection rule of this study and Zou et al. [28] in the urban regions. Figure S4: Comparisons of the water detection rule of this study and Zou et al. [28] in the plateau region. Figure S5: The scatter density of water samples. (a) and (b) indicates the water samples from the sampling plots with Landsat path ≥ 131. (c) and (d) indicates the water samples from the sampling plot with Landsat path < 131. Table S1: The Landsat and Sentinel image for accuracy assessment. Figure