1. Introduction
Lakes are considered as important water resources for human livelihood, agriculture, and industrial production and meanwhile are also an important part of the terrestrial water cycle [
1]. The changes of lake water-body (LWB,
Table A1) area are important indicator of climate and environmental changes at the regional and global scale [
2]. Accordingly, the use of wide range, low-cost, and repeated observation satellite imagery has become an important tool for automatically extracting LWBs over large spatial scales [
3]. Automated extraction of LWB is as a prerequisite and key issue for monitoring and protecting lake water resources and ecological environment [
4]. Thus, it has become an important research topic to automatically achieve high-precision extraction of LWB based on remote sensing imagery.
There are three basic methods for the automated extraction of land surface water bodies (LSWBs; e.g., rivers, lakes, reservoirs, etc.) using remote sensing data (
Table 1). The first utilizes spectral information from optical imagery (e.g., MODIS [
4], Landsat [
5], Sentinel-2 [
6], GF-1/2 [
7], etc.), particularly the differences between LSWBs and non-water objects in the visible, near infrared (NIR), and short-wave infrared (SWIR) wavelengths. This method can be further divided into three subcategories: single-band, two-band, and multi-band. Single-band methods commonly set thresholds for the NIR or SWIR bands to extract LSWBs, as these wavelengths are strongly absorbed by water and reflected by vegetation and dry soil [
8]; however, single-band methods often confuse water with other dark materials or shadows. Two-band methods were designed to enhance water identification accuracy through various algebraic operations. For example, the normalized difference water index (NDWI) was proposed [
9]; however, it remained sensitive to built-up land, dark soil, and shadows [
10]. To resolve this issue, Xu (2006) proposed a modified normalized difference water index (MNDWI) by substituting the SWIR band instead of the NIR band to strengthen the spectral difference of built-up land reflectance; however, shadows of mountains and buildings remained an issue [
11,
12].
These methods (≤2 bands) were thus deemed insufficient, and multi-band methods were proposed to increase the differences between the spectral features of water and other land-cover types [
13]. For example, the automated water extraction index (AWEI) was developed to highlight urban LSWBs over shadows and dark surfaces [
14] through two separate indices:
for urban areas without shadows and
for those areas with dramatic shadowing. The multi-band water index (MBWI) uses of green, red, NIR, SWIR1, and SWIR2 bands [
15] and outperforms other indices (such as NDWI and MNDWI) for extracting surface water from low-reflectance surfaces in areas. The above methods can obtain high-precision extraction results of LSWBs using simple calculations between bands. Yet, they mandate the difficult determination and application of threshold values. However, it is difficult to determine an optimal threshold for extracting LSWB from diverse background spectral information of optical imagery that has been widely employed for large-scale LSWB mapping owing to their high spatiotemporal coverage and efficient calculations.
The second method utilizes polarization information of synthetic aperture radar (SAR) imagery. SAR has been used as an effective method for objects-change detection [
16,
17], deformation detection [
18], and water extraction. The method of LSWB extraction uses synthetic aperture radar (SAR) backscatter from single-polarized (e.g., HH, HV, or VV), dual-polarized (HH/HV or VV/VH), or quad-polarized (HH/HV/VV/VH) data, relying on the lower backscatter coefficients of LSWBs than other land-cover types. For example, Tian et al. (2017) constructed a water index (SWI) based on polarization features (VH and VV) using Sentinel-1 data to monitor the dynamic changes in Poyang Lake (the largest natural lake in China); however, lake boundary misclassifications occurred under the influence of these factors (such as aquatic vegetation and muddy waters) [
19]. Zhang et al. (2019) automatically extracted the LSWBs of the Tibetan Plateau from Sentinel-1 data using a proposed support vector machine (SVM) learning algorithm adapted to identify LSWBs by mapping the input feature vectors (backscatter, Grey Level Co-occurrence Matrix (GLCM), and the polarization ratio) in high-dimensional feature space; however, this method was easily affected by mountain shadows due to the imaging mode of the radar sensor [
20]. Valdiviezo-Navarro et al. (2019) proposed an unsupervised methodology based on a local Moran index of spatial association in combination with morphological closing operations for LSWB extraction in complex topographies from SAR images to address false classification results of small water bodies in montane areas; however, this method was ineffective at extracting water-body boundaries under vegetation cover [
21].
The third method of LSWB extraction involves fusing SAR and optical data to exploit their respective advantages. Previous research has improved LSWB extraction through such fusion methods (Bioresita, Slinski, Michael). Saghafi et al. (2021) tried to fuse optical data and SAR data to extract surface water bodies using an approach based on water indices, supervised classifications, and decision fusion. However, they did not discuss the different types of water bodies. [
22] Overall, LSWB extraction by fusing optical and SAR data addresses some residual defects, such as speckle noise on SAR data and environment noise on optical data; however, the above fusion methods were only used to extract single types of LWB (e.g., fresh, saltwater, and natural lakes) or various LWB types (e.g., polluted, urban, and montane lakes) in complex environments (e.g., shadowed, forested, and built-up areas).
Table 1.
The reviewed works of LWB extraction.
Table 1.
The reviewed works of LWB extraction.
Methods | Subcategories | Literature | Characteristics |
---|
Only optical | Single-band | Work et al., 1976 [8] | This method is simple to calculate, but it is easily affected by shadows of mountains and buildings, and it is difficult to determine an optimal threshold. |
Two-band | Xu et al., 2006 [12] |
Multi-band | Feyisa et al., 2014 [14] |
Only SAR | Single-polarized | Guo et al., 1999 [23] | This method can reduce misclassification caused by the spectral heterogeneity, but it is easily affected by mountains and smooth-material ground objects. |
Dual-polarized | Tian et al., 2017 [19] |
Quad-polarized | Guo et al., 1997 [24] |
Data fusion | SAR-optical data | Saghafi et al., 2021 [22] | This method can suppress the interference of shadows, water quality, and smooth-material ground objects. |
From these studies, we summarize that the methods based only on spectral information of optical imagery are easily affected by water quality of LSWB and low reflectivity. However, they are widely used for LSWB mapping at large scale due to their simplicity and efficient calculation. Methods based on polarization information of SAR imagery can reduce misclassification errors caused by the spectral heterogeneity; these methods are also affected by mountains and smooth material ground objects. Smooth surfaces, such as roads and sand, have low retroreflective coefficients similar to those of water bodies and are often mistakenly classified as water bodies. Recently, there have been a few studies on LSWB extraction by fusing optical and SAR data, which is considered more robust and addresses some defects, such as speckle noise on SAR data and environment noise on optical data [
22,
25,
26,
27]. However, the above methods were only used to extract a single type of LWB (inland fresh, saltwater, and natural lakes), and it is rarely reported for extracting various type of LWB (e.g., polluted, urban, and mountainous lakes) in complex environments (e.g., shadowed, forested, and built-up areas).
Accordingly, the objectives of the present research were: (1) To quantify the improved performance of LWB extraction precision by fusing Sentinel-1/2 data and (2) to verify the adaptability of Sentinel-1/2 fusion data to extracting various LWB lakes (urban, clean, mountainous, and polluted lakes).
The remainder of this paper is organized as follows: The study area and datasets are introduced in
Section 2.1;
Section 2.2 describes the methodology in detail, including data pre-processing, water index calculations, feature analysis and fusion, experimental design, and accuracy assessment;
Section 3 details the extraction results for various LWB types; and a discussion and conclusion are presented in
Section 4 and
Section 5, respectively.
4. Discussion
From the above results, it can be seen that the method of fusing Sentinel 1/2 data to extract LWB has the better performance. Compared with the traditional water index method, it has higher extraction accuracy and lower error. MNDWI is simple to calculate and can quickly generate water maps. However, the determination of the water index threshold will change with time and space, especially in complex environments, such as cities and mountainous areas. A large amount of background information has a strong interference in the determination of the threshold. Therefore, in order to better determine whether the fusion method proposed in this study is suitable for LWB extraction in a complex environment, lake environmental noise, lake water-body type, and computational complexity are discussed.
4.1. Lake Environmental Noise
Extracting water bodies in a complex geographic environment is susceptible to interference from a number of sources (e.g., shadows, buildings, forest vegetation) leading to poor extraction accuracy. However, in LSWBs extraction, such sources of environmental noise present can help categorize lakes into urban lakes and mountain LSWBs.
Urban lakes maintain a complex background environment. In the Sentinel-2 optical imagery, there is interference from building shadows and other low-reflectance features. Methods incorporating radar polarization information overcome the influence of the spectral heterogeneity of ground objects; however, they remain susceptible to smooth ground objects (e.g., bright buildings), creating false positives.
Figure 10a,b is enlarged partial views of the Donghu Lake composite image according to the extraction results of the five methods. Area a, in the northwest part of the lake, was affected by buildings and shadows, creating low LSWBs extraction accuracy values for Methods I–III. Conversely, Methods IV and V reduced the interference of background objects and more completely extracted the LSWBs. Area b in the northeastern part of Donghu Lake suffered from the interference of buildings, mountain shadows, highly reflective surfaces, and roads. Accordingly, Methods I–III did not perform well, whereas Methods IV and V performed better. Area c in the northern part of Dianchi Lake is close to the urban area and contained complex background features. Affected by building shadows, vegetation, roads, etc., the extraction results of Methods I–III are not good, and there are a great deal of speckle noises around the lake. However, Methods IV and V could effectively distinguish the water and non-water pixels. These methods reduced the impact of ground object disturbance, more completely identifying LSWBs in the complex areas and producing a smooth and continuous water boundary with improved accuracy and reduced error rates.
The extraction process for montane lakes in plateau areas was primarily affected by mountain shadows and vegetation. In view of the limitations of remote sensing spectral information in complex geographic environments, mountain shadows were often incorrectly identified as water bodies. LSWBs concealed under vegetation along the land boundary area were prone to mix. As showed in
Figure 11, areas a and b are located in the montane areas on the west and east sides of Fuxian Lake, respectively.
Figure 11c shows an area along the southwest bank of the plateau of Lake Erhai, close to the mountains. Affected by the shadows of the mountains, there were weak patches that were incorrectly identified as water in Method I, and a large number of background objects were incorrectly identified as water in Method III. Due to the influence of the so-called radar shadow (during the imaging process, the top of the mountain is closer to the sensor than the bottom of the mountain, which will cause the object to appear to “collapse” towards the sensor) of the radar sensor, some misclassifications also occurred in Method II. Fusion Methods IV and V, however, effectively improved extraction results.
4.2. Lake Water Body Types
The LWB extraction was also affected by water quality. To verify the adaptability of the proposed methods, lakes were divided into clean and polluted states. Overall, the water quality of clean lakes Erhai and Fuxian maintained low turbidity, chlorophyll concentrations, and clear lake boundaries. It can be seen from the extraction results that, in Method III, the surrounding mountain shadows are wrongly extracted as water bodies. In methods I and II, there were also a small amount of wrong extraction around the lake. In general, although there were some commission errors, for the extraction of LSWBs, all five methods can be completely extracted.
Alternatively, Chaohu and Taihu lakes were considered polluted.
Figure 12a displays the island-containing area of Chaohu Lake, and the surrounding chlorophyll concentrations were relatively high. Methods I and III produced poor, incomplete extractions due to the influence of cyanobacteria. In Method II, the water body containing algae was more completely extracted. Method V maintained the highest extraction accuracy, overcoming the interference of cyanobacteria and more completely separating the islands. Method IV reduced the influence of algae to a certain extent albeit inferior to Method V. Area b contains small islands in Taihu Lake and roads connecting the islands. It can be seen from
Figure 12b that Methods I and III extracted roads in lakes, but Method II cannot extract roads, and the extracted islands are incomplete and broken. The fusion Methods IV and V clearly extract the complete islands and roads. Area c presents the chlorophyll concentrated boundaries on the south side of Taihu Lake. Methods I and III suffered from interference by algae, leading to incomplete LWB extractions. In Method II, S1 radar data largely avoided interference of algae but were affected by the surrounding urban buildings, creating numerous misclassifications with discontinuous and incompletely extracted LSWBs in the boundary areas. Under Methods IV and V, the LSWBs were more completely extracted, and the effects of LSWBs containing cyanobacterial blooms were effectively eliminated. At the same time, it also reduced the interference of the surrounding environment noise, and the error extraction was reduced.
4.3. Computational Complexity
In this paper, the LWB extraction mainly includes three steps: feature analysis and extraction, feature fusion, and classification. In the feature analysis and extraction stage, the operation is simple, and the computational complexity is low. Feature fusion mainly uses the layer stacking to generate multidimensional data from Sentinel-1/2 data, and this process is also relatively low in computational complexity. Therefore, the computational complexity of the entire process mainly depends on the efficiency of SVM. The complexity of SVM is considered from the following two aspects: the dimensions of features and total number of samples. We set up two sets of experiments to verify: (1) The number of training samples remains unchanged at 1000; the feature dimensions of the input classifier are: 1, 2, 3, 10, and 12 (the same as the feature dimensions of the five methods proposed in this study); and the running time under different feature dimensions is recorded. (2) The feature dimension is 1 and remains unchanged; the numbers of input training samples are: 500, 1000, 2000, 3000, 4000; and the running time under different number of samples is recorded.
As shown in
Figure 13, it can be found that the computational complexity of SVM has nothing to do with the dimensionality of features; rather, it is related to the total number of training samples and depends on the number of support vectors [
39]. Therefore, the computational complexity will not increase as the dimension of features changes. Compared with water index method, the method by fusing Sentinel-1/2 data is still an efficient approach.