1. Introduction
The Tibetan Plateau (TP) is known as the earth’s third pole [
1,
2] and the water tower of Asia, and is an important factor greatly affecting the earth’s climate system [
3,
4]. As the headwater area of ten major rivers in Asia, the TP supplies water resources to the billions of inhabitants in the plateau and surrounding regions [
5]. Surface water such as lakes, rivers, and artificial reservoirs in the TP, which have been greatly affected by climate change [
6] and human disturbance [
7,
8], are playing an essential role in climate balance, hydrological cycle, and ecosystem balance [
9,
10]. Therefore, long-term knowledge of the state of water bodies on the TP can improve understanding of the impact of climate change on the cryosphere and hydrosphere.
The algorithms and techniques proposed for mapping water bodies can be categorized into two major types: (1) manual compilation method based on remote sensing images or geographic information system (GIS) database, and (2) automated mapping method based on remote sensing images.
There are several global and regional surface water data sets developed with the compilation method. The Global Lake and Wetland Database (GLWD) was compiled [
11] based on seven GIS databases gathered from the 1990s. Wan et al. [
12] published a data set of lakes over the TP for three periods: 1960s, 2005, and 2014. The manual interpretation and digitization approaches were applied to the ground survey data and the high spatial resolution satellite data mainly from the China-Brazil Earth Resources Satellite (CBERS) and the GaoFen-1 satellite (GF-1). For the compilation method, digitization through visualization with prior knowledge can ensure relatively high accuracy of water boundary extraction and provide databases with strict quality control [
13,
14]. However, the use of these data sets is constrained by their static characteristics, which require higher temporal resolution water maps to reveal long-term changes of water bodies.
Many water mapping algorithms have been developed to extract and delineate open water features by utilizing various satellite images with different spatial, temporal, and spectral resolutions, among which high spatial resolution sensors such as Landsat satellites are the most widely used. The techniques mainly include the thresholding method [
15,
16,
17,
18,
19], spectral water index method [
20,
21], linear spectral unmixing model [
22], and data exploration techniques [
23]. However, due to the long revisit intervals, the main limitation of large-scale use of Landsat data is the availability of cloudless ground observations [
24]. For example, multi-year time series Landsat images are required to compile global cloudless images [
15].
Moderate Resolution Imaging Spectroradiometer (MODIS) data with daily repeat time and medium spatial resolution has the advantage of monitoring detailed water area variation and has been applied to detect seasonal and interannual changes of water bodies regionally and globally [
25,
26,
27]. Using the MODIS 8-day reflectance product, the support vector machine classification [
28], and a threshold segmentation algorithm [
29] were applied to map the spatial distribution of large lakes over the TP. However, the low spatial resolution of MODIS data will affect the purity of pixel and the clarity of water boundary, thus limiting the accuracy of the extracted feature [
30].
Measuring long-term TP surface water changes with both high temporal resolution (~10 days) and high accuracy remains a challenge. Multiple endmember spectral mixture analysis (MESMA) is commonly used to map subpixel land cover distribution from coarse resolution images [
31,
32,
33,
34,
35]. However, because environmental factors on the TP (such as clouds, snow, glaciers, and shadows) complicate the classification process, there are few or no examples of applying MESMA to generate water fraction maps in an automated and operational manner.
In this paper, we present a comprehensive method that is well-suited for mountain environments to map water bodies over the TP. A multi-index threshold endmember extraction method was applied to automatically select the endmembers from MODIS images. To incorporate endmember variability, a combination called typical and adjacent endmembers is used in MESMA endmember selection strategy.
Section 4 details the accuracy assessment of the algorithm at both the local scale and whole regional scale. The cross-comparison was performed with the Landsat 8 Operational Land Imager (OLI) images and four public data sets, namely, GLWD, MODIS Land Water Mask (MOD44W), the TP lake data set, and the newly published European Space Agency (ESA) surface water data.
Section 5 discuss the advantages, limitations, and implementation of the water mapping algorithm.
4. Results
In this study, the accuracy assessment of the water fraction unmixing algorithm was conducted at both the local scale and regional scale. At local scale, Landsat 8 OLI data were used as accurate reference maps of the water cover to evaluate the subpixel water fraction estimates. The Landsat satellites images are the most widely used optical satellite images to extract and analyze surface water information. Two tests were conducted: one was with the MOD09A1 data; the other was with the simulated MODIS reflectance data from the OLI image. At regional scale, the water map of the Tibetan Plateau was compared with four public data sets: the TP lake data set, the GLWD, the ESA surface water data, and MOD44W.
4.1. Validation of the Algorithm with Pairs of OLI/MODIS Images
To evaluate the accuracy of the water maps, eight pairs of OLI/MODIS image subsets (see
Table 1) were selected, which cover the three largest lakes in the TP, namely Qinghai Lake, Selin Co, and Nam Co, as well as five other large lakes. The acquisition dates of the OLI and MODIS data are within eight days. The OLI water maps were produced from the maximum likelihood classification with the Landsat 8 OLI data, followed by manual removal of the shaded area caused by mountains. The cloud and the associated shadow areas were detected using the OLI BQA data and excluded from the comparison. The OLI binary water maps with a resolution of 30 m were resampled to 25 m resolution and aggregated to 500 m, 1000, m and 2000 m resolution using an averaging method to generate the reference maps.
Figure 3 shows examples of the false color OLI map and comparison of the water fraction map from MODIS (raster maps) and OLI data (vector maps) for Qinghai Lake, Selin Co, and Nam Co in UTM projection. The OLI water boundary vector features were generated by an automatic vectorization tool (ArcScan) in ArcGIS 10.2.
The MODIS and the reference OLI water maps were compared using: (1) mapped water area; (2) pixel-based correlation analyses (RMSE, bias, coefficient of determination (R
2), linear regression).
Table 4 shows the comparison results between Landsat and MODIS derived subpixel water fraction maps for each OLI/MODIS image pair listed in
Table 1.
Table 4 reveals that the algorithm estimates the water area well: the error is within 5% for all eight OLI/MODIS pairs, with a total error of −1.4%. The underestimation is mainly attributed to: (1) the small water bodies with widths less than 500 m are not mapped in the MODIS Water Map; (2) although MOD09A1 provides the best possible 8-day composite Level 2 Gridded (L2G) observation, even with the absence of cloud and shadow, the cloud effect still exists in some pixels. By comparing with the QA data of MOD09A1, we found that in some areas (e.g., the red rectangle in
Figure 3d), the surface water cannot be mapped due to the clouds in the MODIS image.
Pixel-based correlation analyses were used to analyze the difference between the reference Landsat water fractions (x) and the MODIS water fractions (y). At 500 m resolution, the overall RMSE is 7.86%. The RMSE of image pair no. 2 is the biggest (10.1%) partially because there are many small puddles which are not mapped in the MODIS image. As the resolution became coarser (from M500 to G2000), pixel-based correlation analysis shows the R2 increased by between 0.02 and 0.05. The biases at three resolutions (500 m, 1000 m, and 2000 m) are the same due to the linear aggregation resampling of the pairs of maps. The total bias is −0.2%, which reveals a small underestimation of the water fraction in the spectral unmixing algorithm, which is mainly because the small water bodies were not considered in the algorithm.
4.2. Performance of the Water Mapping Algorithm at the Boundary of Water Bodies
High precision co-registration of different resolution satellite images is difficult to accomplish and may cause inaccuracy in the test. In order to focus on the accuracy of the algorithm at the boundary of water bodies, the algorithm was tested using the simulated MODIS image from Landsat 8 OLI image, which can provide perfect co-registration accuracy. The OLI sensor has bandwidths corresponding to the MODIS sensor, but its bandwidths are wider than that of MODIS (see
Table 5). The consistency between the OLI data and the MODIS data has been demonstrated [
63,
64].
The OLI image over Ngoring Lake [the worldwide reference system (WRS): 143/036] obtained on 16/08/2014 (
Figure 4a) was used as the reference data. By aggregation of the OLI data to 480 m resolution using the average resampling method, the MODIS reflectance (
Figure 4b) was simulated. The 480 m coarse resolution data can achieve a perfect co-registration accuracy with the OLI data which the MODIS 500 m product cannot. The size of the simulated MODIS image is 100 × 100 pixels (48 km × 48 km). Since band 5 of MODIS data does not correspond to any band of OLI data, only six other bands were used in the test. By comparing the OLI water fraction reference map (
Figure 4c) with the MODIS water fraction map (
Figure 4d), it can be seen that the algorithm can delineate the lake shoreline well. The total area of water body estimated from Landsat is 687.54 km
2, while the total water area from the simulated MODIS data is 685.82 km
2. The main reason for the underestimations (0.25%) is because the water mapping algorithm ignores small water bodies, for example, the southwest bottomland of Ngoring Lake (the red rectangle in
Figure 4c).
To evaluate the performance of the algorithm at marginal water pixels, the scatter plot of the fractional water body retrieved from simulated MODIS data versus the Landsat 8 OLI water maps is shown in
Figure 5. The small isolated water bodies were removed from the comparison. The linear regression is good with RMSE of 14.7% and R
2 of 0.78. Since the comparison is performed only at the mixed pixels of the lake boundary, when all pixels of the scene participated in the comparison, the RMSE and R
2 become 5.81% and 0.98. Therefore, the result shows that the algorithm performs well on the boundary of the lakes and better for the whole scene.
4.3. Validation with the Public Data Sets
Seven MOD09A1 tiles (h23v05, h24v05, h24v06, h25v05, h25v06, h26v05, h26v06) on 2014/10/08 were used to generate the water fraction map of the Tibetan Plateau, which is shown in
Figure 6. To assess the performance of the subpixel water mapping algorithm at regional scale, firstly the TP water fraction map in this work was compared with two public data sets: the TP lake data set and the GLWD. The three water maps were projected into Albers Equivalent Conical Projection, from which the latitudinal and longitudinal water area distributions were aggregated at steps of 1° and compared with each other (see
Figure 6). The MODIS water fraction map in this study matches very well with the TP lake data sets by Wan et al. [
12] in both latitudinal and longitudinal distributions. However, a small difference was found between these mentioned two and the GLWD, because they reflect lake areas at different time periods and some small lakes are typically omitted in GLWD maps [
12,
15]. In terms of the longitudinal distribution between 93.5° and 95.5° East, the lake area in this study is greater than that found by Wan et al. [
12], although it shows a similar distribution as the GLWD. This is mainly because human-exploited salt lakes (especially in Qaidam Basin) are not included in the TP lake data set.
To further evaluate the accuracy of the algorithm at the basin scale, the lake area of this study, the MOD44W in 2014, and ESA surface water data in Oct 2014 were aggregated by basins and compared with the TP lake data set, which was used as reference data since the manual compilation method can ensure relatively high accuracy of water boundary extraction.
Table 6 summarizes the detailed results of the comparison. The human-exploited salt lakes, manmade reservoirs, and big rivers were masked out from both the MODIS water fraction map and the MOD44W like the TP lake data set. It should be noted that the satellite images used in the four water maps are different in both temporal and spatial resolution. The MODIS water fraction map was generated using the best MODIS observations during an 8-day period (October 8–15) at 500 m resolution; the MOD44W provides water maps using all MODIS 16-day composite images for the reported year at 250 m resolution; the ESA surface water map used Landsat Images to produce monthly open water occurrence (Oct 2014) at 30 m resolution; the TP data set utilized the multi-temporal (wet and dry season) images (mainly from the GF-1 wide-field-of-view (WFV) sensor at 16 m resolution) to generate the lake water boundaries. As shown in
Table 6, the total lake areas for the four data sets are 45,125.04 km
2 (this study), 43,616.72 km
2 (MOD44W), 46,551.18 km
2 (ESA), and 46,595.08 km
2 (the TP lake data set), respectively. ESA surface water maps have the highest accuracy (−0.09) due to highest resolution (30 m) and advanced big data exploration and information extraction technology. At the MODIS resolution (500 m and 250 m), the difference of the total area (−1470.04 km
2, −3.15%) between the MODIS water fraction map and the reference TP lake data set is less than that (−2978.36 km
2, −6.39%) between the MOD44W and the TP lake data set. This is mainly because the water area in the Inner F and Inner E basin for the MOD44W water mask is greatly underestimated. There is also an underestimation of the lake area in this study, (e.g., Inner E, Brahmaputra and Yellow Basin), which is attributed to: (1) saline lakes (especially in the Inner E basin) which sometimes have a salt layer above the water surface, which greatly increases the reflectance of the water pixels and causes the pixels to be classified as ‘nonwater’; (2) sometimes ice and snow cover the lakes, which cause the pixels to show reflectance of the snow; (3) although the MOD09A1 is processed using 8-day composite data, cloud effect still exists at some water pixels which causes underestimation; (4) as mentioned in
Section 4.2, the small lakes with widths less than 500 m are not mapped by the algorithm. Using the MOD09A1 image from 17 May 2014, we also compared the MODIS water mapping results (water area 38,533.048 km
2) with the ESA surface water data (water area 39,894.66 km
2) in May 2014. The data are not listed in this paper. The area difference is −3.41%, which means that these two data (this study and ESA surface water data) can consistently reveal the seasonal variation of TP surface water to a certain extent.
Figure 7 illustrates the scatter plots of the basin lake area derived from this study versus the reference TP lake data set and the MOD44W versus the TP lake data set. From
Figure 7, we can see the linear relationship between the MODIS water fraction map in this study and the TP lake data set is slightly stronger than the relationship between the MOD44W and the TP lake data set, which reveals a better water mapping accuracy.
6. Conclusions
Prior works focus on mapping the TP surface water using high-resolution satellite data, which cannot monitor long-term changes with both high temporal resolution (e.g., 10 day) and high accuracy. This study presents a method that is very suitable for mountain areas to map open water in the TP. A multi-index threshold method was implemented to select endmembers. To overcome the confusion between the water and hill shade spectrum, two steps were implemented: (1) only the neighboring pixels around the pure water pixels participate in the unmixing process; (2) a shadow map is generated based on a SRTM DEM and is masked out from the water fraction map. The accuracy of the algorithm was assessed using the Landsat 8 OLI data (local scale) and four data sets (regional scale): the TP lake data set, the MOD44W, and the GLWD. At local scale, the comparison with the Landsat water maps revealed good accuracy. To overcome the co-registration error due to resolution differences between MODIS and OLI data, the performance of the algorithm at the boundary of the lakes was evaluated using a simulated MODIS image generated from the OLI image. The comparison illustrates the unmixed water fraction map from the simulated MODIS image is highly correlated with the water fraction map retrieved from the original OLI image (R2 = 0.78), which reveals that the algorithm performs well at the boundary of lakes. At the regional scale, the water fraction map in 2014 matched well with the TP lake data set and the GLWD in both latitudinal and longitudinal distribution. The lake area estimation is more consistent with the reference TP lake data set (difference of −3.15%) than the MOD44W (difference of −6.39%).
The limitations of the method are worth noting. Due to the coarse resolution of the MODIS images and the mixed water pixel extraction strategy, the algorithm only considers water bodies with a width greater than 500 m. In addition, although MOD09A1 provides the best possible 8-day composite L2G observation in the absence of cloud and shadow, the cloud effect still affects some pixels, which might cause undetected water pixels. Therefore, future work will consider time-series information on MODIS images to mitigate the cloud effect and provide more temporally consistent water maps.