Monitoring monthly surface water dynamics of Dongting Lake using Sentinel-1 data at 10 m

Study area

Dongting Lake (28°30′ to 30°20′N, 110°40′∼113°10′E) is the second-largest freshwater lake in China, located at the middle reaches of the Yangtze River Basin, with an extensive catchment area (Fig. 1) (Ding & Li, 2011). The lake has the Xiangjiang River, the Zi River, the Yuan River, and the Lishui River as its south and west tributaries; the Miluo River and the Xinqiang River as its east tributaries; the Taiping estuary and the Ouchi estuary as the bleeders of Yangtze River water flow, and the lake water discharged back into the Yangtze River at the northeast outlet near Chenglingji (Li et al., 2013; Zhang et al., 2006). In addition, Dongting Lake is a typical passing reservoir, which adjusts both the inflow from and outflow to the mid-reaches of the Yangtze River. During flood season between April and September, the lake provides storage for river flood waters. During the dry season between October and March, it provides water to the river to allow river transportation to continue without significant interruption (Zhang et al., 2006). Because of its important geographic location, Dongting Lake includes three national wetland nature reserves: the East Dongting wetland, the South Dongting wetland, and the West Dongting wetland, which were included in the List of Ramsar Sites in 1992 and 2002 (http://www.ramsar.org/wetland/china). It also supports shipping and fisheries, provides irrigation water, and regulates the local climate. Therefore, monitoring the short term dynamic change in the surface water area of Dongting Lake has direct and significant effects on preventing floods, the stability of the wetland ecosystem and its biodiversity.

Datasets

Data preprocessing

When the radar electromagnetic wave is specularly reflected by calm surface waters, water pixels present lower backscattering than in bare soil or vegetation pixels. So, when the backscattering coefficient is the lowest, the soil moisture content is highest. In our study, the backscattering coefficient for VH and VV polarizations were used to extract the surface water area of Dongting Lake.

Sentinel application platform software (https://sentinel.esa.int/web/sentinel/toolboxes/sentinel-1) provided by the ESA was utilized to preprocess the Sentinel-1A imagery. The workflow included six steps: (1) Sentinel-1A images were allocated to orbit files. (2) σ₀ bands were created from Sentinel-1A images through radiometric calibration. (3) The Range Doppler Terrain Correction was used to orthorectify σ₀ bands with SRTM 3 s as digital elevation model. (4) The backscattering coefficients (in dB) was acquired from the orthorectified σ₀ band by the equation 10 × log10 (σ₀) (Li, Shao & Zhang, 2011). (5) Speckle noise was removed by using a single product speckle filter method with a window size of 5 × 5 pixels (Gupta & Gupta, 2007; Schmidt & Skidmore, 2004; Tian et al., 2015). (6) The median value of all pixels at the same location for each month was used to extract the monthly surface water extent.

Threshold-based classification

During the past decade, many methods for water mapping with SAR images have been developed, such as visual interpretation (Brivio et al., 2002; MacIntosh & Profeti, 1995), image texture analysis (Schumann et al., 2005), histogram thresholding (Matgen et al., 2004), the edge detection approach (Schumann, Di Baldassarre & Bates, 2009), and image statistic-based active contour models (Horritt, Mason & Luckman, 2001; Schumann et al., 2005). In some studies, the thresholding technique has been viewed as an efficient and simple water mapping method (Schumann, Di Baldassarre & Bates, 2009). In general, the key point is determining the threshold (Bazi, Bruzzone & Melgani, 2007). Empirical methods have been most commonly used for determining thresholds (Gstaiger et al., 2012). Empirically defined gray-level thresholds for classification depend much on the operator’s ability to distinguish different gray-scale tones and their knowledge of water body image characteristics, which may lead to variable accuracy due to the subjective evaluation. Other methods for selecting the threshold include the Otsu method (Otsu, 1979), the maximum entropy method (Kapur, Sahoo & Wong, 1985) and the minimum interclass variance method (Radhika, Sekhar & Venkatesha, 2009), all of which are computationally complex and time-consuming. In this study, we used a simple and efficient method to select the threshold.

The method was based on the normal probability density function of the training sample of water and non-water. The Shapiro–Wilk test in SPSS release 25.0 was performed on four sets of water and non-water training samples. Normality was determined by checking the backscattering coefficient for VH or VV polarizations. For all four sets of training samples data, all tests have a p-value 0.05, which indicates a normal distribution of data. The plots of the normal distribution probability density functions of the water and non-water training samples presented two obvious single peaks (Fig. 3). The water and non-water each corresponded to a peak (Fig. 3). There was a trough between the two peaks. The probabilities of the water and non-water were both lower in the trough region. The value in the trough could be considered as the threshold that separates water and non-water. Thus, we generated the normal probability density function using the training samples, and then used the x-axis value of the intersection of the two probability density function plots as the threshold for identifying surface water. Meanwhile, as water and non-water have the same probability at this threshold, the accuracy of both the water and non-water results is guaranteed to be at a maximum. In addition, this method can also prevent the case where the accuracy of one class is ensured while the accuracy of others is ignored.

Accuracy evaluation

Accuracy assessment of the surface water distribution map is an important part in the image classification procedure. In this study, overall accuracy (OA) (Eq. (1)), user’s accuracy (UA) (Eq. (2)), producer’s accuracy (PA) (Eq. (3)), and Kappa coefficient (Eq. (4)) were used to evaluate the classification accuracy of Sentinel-1A (Congalton, 1991). The series of accuracy metrics can be derived from a confusion matrix which is the most frequently used method in accuracy evaluation (Foody, 2002): (1) $$p_{\text{o}}={\displaystyle \sum _{i=1}^r}\frac{p_{ii}}{N}$$ (2) $$p_{u_i}=\frac{p_{ii}}{p_{i+}}$$ (3) $$p_{{\text{A}}_i}=\frac{p_{ii}}{p_{+i}}$$ (4) $$\text{Kappa}=\frac{N*{{\displaystyle \sum }}_{i=1}^r{p}_{ii}-{{\displaystyle \sum }}_{i=1}^r\left(p_{i+}*P_{+i}\right)}{N^2-{{\displaystyle \sum }}_{i=1}^r\left(p_{i+}*p_{+i}\right)}$$ where p_o is the OA; $p_{u_i}$ is the UA; $p_{{\text{A}}_i}$ is the PA; p_ii is the number of correctly classified samples of class i; r is the number of classes; N is the number of training samples; p_i+ is the number of classified samples of class i; and p_+i is the number of training samples of class i.

Dynamic change analysis method

To understand the water dynamics change characteristics and influences at Dongting Lake, the submergence ratio was defined as the probability of a pixel being inundated by water for one year (Eq. (5)): (5) $$P_{\text{ratio}}=\frac{n}{N}*100\%$$ where P_ratio represents the submergence ratio; n is the inundation number of each pixel observed by Sentinel-1A in a year; and N is the observation times of Sentinel-1A (equal to 12 in this study).

Classification accuracy

Validation samples were used to access the surface water mapping accuracies. The results (Table 3) showed that both VH and VV backscattering coefficients achieved high mapping accuracies, and the water and non-water types were well-distinguished based on the threshold-based classification method. The OA and kappa coefficients of all classification results were above 94.50% and 0.88, respectively. The producer’s and user’s accuracies for both water and non-water types were above 91.50% and 91.67%. It seems that the errors in the results were mainly omissions or commissions of water pixels around the edges of the lakes. To prove the reliability of results, training samples were obtained for another eight months and a support vector machine (SVM) was used to map the water of Dongting Lake. The classification accuracies based on SVM demonstrated that the OA and kappa coefficients of all classification results were above 95.00% and 0.89 (Table 4). The producer’s and user’s accuracies for both water and non-water types were above 92.00% and 92.65% (Table 4). The results indicate that threshold-based classification accuracies were close to the accuracies based on SVM (only 2.24% lower). Thus, the classification errors cannot due to using of a simple algorithm. In contrast, the threshold-based classification method was universal and effective.

By comparison, the results further indicated the superiority and higher performance of the VH backscattering coefficient as compared with the VV backscattering coefficient. Accordingly, the classification results based on the VH backscattering coefficient were used to monitor the monthly changes and the submergence ratio of Dongting Lake in 2016.

MODIS data have high temporal resolution, the revisit time is one day. It is widely used in monitoring the dynamics of surface water area. However, due to the problem of mixed pixels, the accuracies of these researches are lower. The map producer and map user accuracies of surface water from MODIS are always below 90%, which means that part of the surface water area is not detected by MODIS (Li et al., 2009). It may be that other advanced methods can improve the quantity accuracy; however, MODIS data cannot identify details of the surface water and small water bodies precisely at this time. The advantage of our research is that data used in this study exhibit a high spatial resolution, which can reduce misclassification, so the accuracy of our study is higher than MODIS and the results of classification are more reliable. Landsat data have 30 m of spatial resolution. The classification accuracies of water surface maps based on Landsat data are always higher than 90% (Zhou et al., 2017). However, Landsat data cannot provide monthly cloud-free images, especially in regions with high precipitation, thus such images cannot be used for monitoring inner-annual water dynamics.

Water identification threshold

The thresholds for the VH and VV backscattering coefficients were −21.56 and −15.82 dB. Zeng et al. (2017) selected the value −16.35 dB as a threshold for the VV backscattering coefficient to extract the surface water area of Poyang Lake using Sentinel-1 SAR imagery. Their thresholds for surface water identification were similar to the thresholds used in our research. We also obtained monthly VH and VV backscattering coefficient thresholds using the method proposed in this paper and corresponding monthly training samples. The results are shown in Table 5. The monthly thresholds for VH backscattering coefficient were between −22.59 and −20.46 dB. The monthly thresholds for VV backscattering coefficient were between −16.91 and −14.30 dB. The thresholds either for VH or VV backscattering coefficient varied slightly over time. The thresholds in this paper for the VH and VV backscattering coefficients were −21.56 and −15.82 dB, which was close to the monthly thresholds. The key factor influencing the threshold is the status of water surface. Wind is an important factor that disturbs the water surface. Monthly longitudinal (u) and lateral (v) wind speeds data at 10 m above the water surface of Dongting Lake were acquired from the ERA-Interim data (http://app.ecmwf.int/datasets/). The v and u wind speeds were between 0.47 m/s and 4.58 m/s. Dongting Lake experiences gentle breeze with grade-3 wind-forces throughout the year. Wind speeds vary slightly during the year. Figure 4 is an illustration of a fitting curve for monthly wind speeds and thresholds for VH or VV backscattering coefficient. The two fitted linear functions (R² = 0.01 and R² = 0.12) demonstrated that the threshold was not obviously related to the wind speed. Therefore, seasonal changes of the effect of wind speed on water surface over the year can be ignored. In addition, the accuracy assessment showed that all the results reached a very high accuracy (91.50–98.48%), which indicated that thresholds in this study were feasible.

By comparison, the accuracies of water and non-water classification based on the VH backscattering coefficient were slightly higher than that based on the VV backscattering coefficient. This finding confirmed that the VH polarization was more sensitive to the presence of surface water than the VV polarization, which was consistent with the conclusion found by Pham-Duc, Prigent & Aires (2017). In addition, wind-induced surface roughness over open water can influence the backscattering coefficients, especially for the VV polarization. The average annual breeze speed at Dongting Lake was with wind-force of three grades (Liu et al., 2009). The wind may affect VV backscatter coefficients more and further lead to more misclassifications between water and non-water than VH-polarized data. In addition, the water surface identification method solely uses Sentinel-1 data, which is not affected by clouds. This method is therefore valuable for the water surface monitoring work in the region affected by high cloud coverage and precipitation.

Inner-annual water area dynamics

The classification results were obtained using the universal threshold to distinguish water and non-water from the VH backscattering coefficient images. Figure 5 shows the monthly surface water acreage of Dongting Lake in 2016. Surface water distributions showed significant seasonal variation. During the rainy season, the lake was in a vast expanse of water, but in the dry season, the lake lost water. The surface water acreage of Dongting Lake changed from 738.89 km² in December to 2040.33 km² in July. During the rainy season (April to August), the lake area was larger than 1,000 km², with an average of 1572.59 km². During the dry season (September to March), the lake area was less than 887.38 km², with an average of 809.75 km². The annual surface water area of Dongting Lake extracted from MODIS images ranged from 440 to 1,900 km² from 2005 to 2009 (Huang, Li & Xu, 2012). The difference between the minimum values is greater than about 300 km². The main reason is that the spatial resolution of MODIS is too coarse to allow extraction of the small water bodies, or to identify the water detail in certain regions of Dongting Lake that are narrow, especially during the dry season. In addition, the change characteristics of the water area at Dongting Lake showed obvious regional differences. The seasonal change of water area happened mainly in the East Dongting wetland and the South Dongting wetland. The spatial extent of the water area of the West Dongting wetland remained relatively stable.

In 2016, the water area of Dongting Lake began to increase in April and reached a peak in July, then started to decline in September and reached a low in December (Fig. 6). The maximum loss in surface area was 803.23 km² with a variation rate of 47.51% from August to September. The maximum increase in area was 609.67 km² with a variation rate of 60.81% from April to May. The average changed acreage was 268.50 km². The variation rate in each monthly interval showed that there were eight changes greater than 10%. The average variation rate was 21.38%.

The precipitation, evaporation, runoff, and infiltration were all major influences on Dongting Lake’s surface water areas. However, only monthly precipitation and evaporation data for 2016 were acquired from the ERA-Interim data, which were downloaded from the European Centre for Medium-Range Weather Forecasts (http://app.ecmwf.int/datasets/). In our study, effective monthly precipitation equals the monthly precipitation minus the monthly evaporation. Figure 6 shows the correlation between the surface water area of Dongting Lake and the effective precipitation over time. The tendencies of both are consistent. The surface water area of Dongting Lake increased as the effective precipitation increased, and deceased as the effective precipitation decreased. Both the surface water area and the effective precipitation reached a maximum value in July. Therefore, the effective precipitation was one of the most important factors that influenced the dynamic changes of surface water area of Dongting Lake.

Submergence ratio analysis

The submergence ratio shows the spatiotemporal dynamics change of surface water area at Dongting Lake in 2016 (Fig. 7). The overall change characteristics of the value of submergence ratio presented a decrease from the lake center to the edges. The acreage of the permanent water that was submerged the entire year in 2016 (submergence ratio = 1) was 556.35 km², accounting for 19.65% of the total area of Dongting Lake. The permanent water was mainly distributed in the central district of Dongting Lake. Approximately 83% of permanent water was located in the south Dongting Lake (51%) and the East Dongting Lake (32%) (Fig. 8). The area of seasonal water was 1525.21 km², accounting for 53.87% of the total area of Dongting Lake, and these regions should receive more attention due to their higher risk of flooding. The seasonal water was generally distributed around permanent water. The seasonal water (0.75 < submergence ratio < 1, and 0 < submergence ratio ≤ 0.25) were all mainly located in the South Dongting Lake (49%, 56%). The seasonal water (0.25 < submergence ratio ≤ 0.50) was mainly located in the East Dongting Lake (60%). The acreage of seasonal water (0.50 < submergence ratio ≤ 0.75) was equal in both the East Dongting Lake (45%) and the South Dongting Lake (45%). Regardless of whether the water was permanent, water or seasonal water were less than 20% in the West Dongting Lake. In this study area, there were still 749.93 km² of surface area that was not submerged at any time during the year (defined as non-water land, with a submergence ratio of 0), and 88% of the non-water land was located in the East Dongting Lake and the South Dongting Lake.

As a whole, the permanent surface water (submergence ratio = 1) of the East Dongting Lake, the South Dongting Lake, and the West Dongting Lake accounted for 20% of the subset lake area. Meanwhile, there were still some differences in the spatial characteristics of these three subset lakes. In the West Dongting wetland, the non-water (submergence ratio = 0) accounted for the largest proportion of the total surface area (48%). However, the permanent water accounted for the largest surface area in both the East Dongting Lake (72%) and the South Dongting Lake (49%) (Fig. 9).