1 Introduction

Coastal zones are ecologically essential and exceptionally dynamic. Monitoring these regions is essential for coastal environmental protection and development. The waterline, also called shoreline or coastline in the coastal zones, is defined as contact between land and the water body. It plays an essential role in analyzing land/water resources, monitoring coastal erosion [3], as well as global sea-level rise.

Clouds easily contaminate optical remote sensing waterline detection. Waterline extraction from synthetic aperture radar (SAR) imagery is becoming more common due to the radar’s all-weather and all-day capability. However, distinguishing the waterline in SAR images is not as simple a procedure for visible-band sensors. The wind-roughed and wave-modulated water return can frequently equal or exceed the return from a nearby land area, resulting in an inadequate contrast for unambiguous land-sea separation. In addition, affected by the moisture of the sandy sediments [6], this phenomenon is more evident in some tidal flat areas. Besides, the speckle noise generated by the coherent signal-scattering complicates the waterline extraction problem for SAR images.

Since the remote sensing data has been growing exponentially and the manual delineation is labor-intensive and subjective, several automatic or semi-automatic waterline extraction methods for SAR images have been proposed based on two conventional approaches: edge detection [10, 13, 16, 19, 28] and image segmentation [9, 14, 22, 24]. However, no matter which one they are based on, these methods more or less require preprocessing and postprocessing for an accurate extraction result from SAR images [9, 24].

Recently, deep convolutional neural networks (DCNN) have widely been employed to extract information from remote sensing images [11]. Several machine-learning-based methods have been proposed for waterline or coastline extraction from SAR images, which all show far better results than the conventional edge detectors [1, 8, 29]. However, unlike regular land or ice regions, the SAR imaging of tidal flat areas shows dramatic brightness changes under different sea conditions.

In this chapter, a modified U-Net has been used to create a framework for automatic waterline extraction from Sentinel-1 SAR images of a large-scale tidal flat at Subei Bank in the Southern Yellow Sea. The extracted waterlines are continued to be applied to construct the digital elevation model (DEM) series in different years for evolution analysis of tidal flat using the waterline method. In this chapter, we first describe our study area, the unique palm-like Radial Sand Ridges along the Jiangsu coast, and the various sandbanks’ SAR imaging features under different sea conditions. Afterward, we introduce our input data and the DCNN-based method. Finally, after testing the trained model’s performance, we developed a processing chain for constructing the tidal flats DEM with the automatically extracted waterlines and an assimilative ocean tide model.

2 Study Area and Data

The Jiangsu coast is located in the western part of the South Yellow Sea, and its offshore area is characterized by palm-shaped radial sand ridges (RSRs). The RSRs consist of more than ten prominent submarine sand ridges and have a unique radial palm shape with the central apex near Jianggang. This giant system is well-developed owing to the active tidal processes and abundant sediment supply from the river runoff [4]. It has a length of 200 km in the north-south direction and a width of 90 km in the east-west direction, with the water depth ranging from 0 to 25 m [27]. The complex hydrodynamic system [20, 30] makes the area’s topography changeable. As shown in Fig. 1b, there are several large-scale tidal flats distributed in the study area.

Compared to optical imaging systems, the active microwave sensor acquires data independent from night and cloud cover, ensuring continuous study area acquisitions. The Sentinel-1 mission comprises a constellation of two polar-orbiting satellites, operating day and night performing C-band SAR imaging, enabling them to acquire imagery regardless of the weather [26]. The two satellites, Sentinel-1A (launch on 3 April 2014) and Sentinel-1B (launch on 25 April 2016), complement each other allowing six days revisit times or even less (in polar regions). With the support of Google Earth Engine [7], we collect 140 pre-processed Ground Range Detected (GRD) IW (interferometric wide-swath) mode with dual-band cross-polarization (VV and VH) and 10 m spatial resolution Sentinel-1 SAR imagery from 2015 to 2019 for the waterline extraction analysis in this chapter. Among the 140 images, 52 acquired in 2019 are used for training and testing our DCNN model and the remains for constructing the large-scale tidal flats’ DEM in Subei Bank.

Fig. 1
figure 1

a Overview of the study area. The elevation data are from the ETOPO1 (NOAA National Geophysical Data Center, 2009). b Sentinel-1B SAR image of the study area at low tide, imaged at Greenwich Mean Time (GMT) 09:54, 26 November 2019

Full size image

Besides the speckle noise, the accuracy and efficiency of the automatic extraction of waterlines in the study area are mainly interfered with by two other factors: the rapid local brightness changes in seawater and tidal flats. The SAR image represents a two-dimensional radar backscatter map of the ocean surface roughness. Therefore, some related processes (such as winds, internal solitary waves, currents, underwater topography, oil spill, rainfall, and eddies) that cause local roughness changes will drive apparent brightness or darkness in imaging. According to Zhang et al. [31], affected by wind and tidal currents, the imaging features of shallow water topography in our study area can often be captured by SAR. As shown in the northeast corner of Fig. 1b, the three underwater sand ridges are shown as narrow bright stripes (1 km wide) in this SAR image. The non-uniform SAR imaging of the sea surface is more evident in Fig. 2. These four sub-images are acquired under different sea conditions and show a considerable imaging difference from each other both on seawater and the tidal flats. The uncertain changes bring great difficulties to the automatic extraction of waterlines for these large-scale tidal flats.

Fig. 2
figure 2

Four typical Sentinel-1 VV-polarized SAR image examples acquired at the different tidal level: a Sentinel-1B image acquired at GMT 09:54, 19 March 2019; b Sentinel-1A image acquired at GMT 09:55, 26 December 2019; c Sentinel-1A image acquired at GMT 09:54, 23 November 2016; d Sentinel-1A image acquired at GMT 09:55, 21 July 2017

Full size image

3 Methodology

3.1 U-Net

The DCNN to extract pivotal information from remote sensing images has been successfully applied in oceanography. Recently, Li et al. [11] established an improved U-Net network to efficiently and automatically extract different ocean process signatures in optical and radar images. The U-Net [23] is a modified fully convolutional network [15] initially developed for biomedical image segmentation. The network is based on the Fully Convolutional Network but extended to work with fewer training images to yield more precise segmentation. The network consists of a contracting path and an expansive path, giving it a U-shaped architecture. As shown in Fig. 3, the left contracting path is a typical convolutional network that consists of repeated application of convolutions, which are followed by a rectified linear unit (ReLU) and a max-pooling operation. During the contraction processing, the spatial information is reduced while the image feature is increased. The right expansive pathway combines a sequence of up-convolutions and concatenations with high-resolution features from the contracting path. One crucial modification in U-Net is that there are many feature channels in the upsampling part, allowing this network to propagate context information to higher resolution layers. Consequently, the expansive path is more or less symmetric to the contracting path, yielding a U-shaped architecture of this network. The main idea is to supplement a usual contracting network by successive layers, where upsampling operators replace pooling operations. Hence these layers increase the resolution of the output. A successive convolutional layer is able to learn to assemble precise output based on this information.

Fig. 3
figure 3

The U-Net architecture specially tailored for this chapter

Full size image

As shown in Fig. 3, the U-Net’s last layer is 1\(\times \)1 convolution with the Sigmoid activation. Traditionally, the loss function of the original U-Net is the cross-entropy. However, in the task of waterline extraction, the samples are highly unbalanced, i.e., the background samples’ numbers are much higher than those of waterline samples (less than 1% points in whole SAR images). Motivated by Lin et al. [12], we adopt the \(\upalpha \)-balanced cross-entropy in this task.

3.2 Data Preparation

The original spatial resolution of the dual-polarized Sentinel-1 SAR imagery downloaded from Google Earth Engine is 10 m. After statistical analysis, we found that the boundary lines between land and water on the VV polarization images are more apparent than those of the VH images. To save computing resources and training time, we only use the VV polarization images and downsample them to a resolution of 50 m. Finally, a full SAR image of the study area is 2229 pixels high and 2005 pixels wide. We further crop the images and their corresponding ground truth into sub-images with 256\(\times \)256 pixels size to keep memory consumption low during training (the edge is filled with black when it is less than 256 pixels). In the end, we acquired a total of 3024 pairs of images for training the U-Net. In addition, before training the network, data augmentations are performed to compensate for a limited number of images in the training dataset. Data augmentation is a technique to increase the amount of data by adding slightly modified copies of already existing data, including random contrast, brightness change, image rotation/cropping, noise injection, etc. It may help the network learn more tidal flat waterline features in the SAR imagery with protean brightness and shapes.

Ground truth labels are necessary when we train a machine learning classifier. Since there is no corresponding waterline product and a method that can automatically extract these edges, we use manual drawing to obtain the ground truth value of the waterline of the 52 Sentinel-1 SAR images acquired in 2019 (the depicted result is shown as the output in Fig. 3). In practice, we use a stylus and touch screen to represent the position of the waterlines accurately. We first randomly select 1/5 of 52 pairs of images, that is, ten pairs as the testing set, to examine the accuracy of extracting the waterlines in the independent data of the trained model. The remaining 42 SAR images with their labels are used for U-Net model training.

3.3 Training

The cropped sub-images from 42 Sentinel-1 SAR imagery are divided into 80% for training and 20% for validation in the training process. The training and testing of the network are implemented by Keras/Tensorflow framework (on NVIDIA Tesla V100 GPU, 32 GB). As mentioned above, we adopt the \(\upalpha \)-balanced cross-entropy as the loss function (\(\upalpha \) is set to 0.99) and the classification accuracy as the performance metric. Furthermore, the batch size is set to 16, and the number of epochs is 4000. Finally, the classification accuracy of the 20% validation images is 94.45% after nearly ten hours of training.

4 Results

4.1 Model Performance

The binary classification accuracy is estimated by calculating the precision and recall of the automatically extracted waterlines to manual ones. The mean precision and recall of the ten testing images are 0.92 and 0.77, respectively (see Table 1 for details).

Table 1 Details of ten testing Sentinel-1 SAR images
Full size table

Four examples of the ten testing results under different sea conditions are shown in Fig. 4. We use three-color lines to compare the difference between the model results and the true values. Yellow represents the waterlines accurately extracted by our DCNN-based model. Red indicates the missing parts of the model, while blue means the false detected lines that shouldn’t be there. The fluctuation of the tides causes drastic changes in the shape and distribution of the waterlines. Figures 4a-c show the results under three typical tidal levels: high, medium, and low, which can also be judged from the exposed area of the tidal flats. What’s interesting here is that Fig. 4c captures a small amount of Enteromorpha information shown as the little bright spot in the northern sea. As shown by the yellow lines in Fig. 4, most of the obtained extraction results from the DCNN-based model correspond well to the manually annotated ground truth waterlines.

4.2 Automatic Topographic Mapping of Tidal Flats

Knowledge of a waterline’s orientation, position, and outline is essential in sea autonomous pilot, verification of coastal platform’s attitude and place, the geolocation of ships, geographic mapping, etc. It also has a specific application for constructing a digital elevation model (DEM) of an intertidal zone by the waterline method [16]. This method is first introduced by Mason et al. [17]. The waterline can be regarded as a quasi-contour line of the topography. This method was proved to be one of the best methods that provide an excellent trade-off between accuracy and cost-effectiveness for the DEM generation of tidal flats [18, 24, 32].

Fig. 4
figure 4

Four examples of the ten testing images overlaid with their corresponding trained model extraction results and ground truth waterlines: a Sentinel-1A image acquired at GMT 09:55, 11 July 2019; b Sentinel-1B at 09:54, 29 July 2019; c Sentinel-1A at 09:55, 17 June 2019; d Sentinel-1A at 09:55, 05 June 2019. The mean precision and recall of the ten testing images are 0.92 and 0.77, respectively

Full size image

This study further attempts to establish a method for automatic topographic mapping of tidal flats based on the waterline method and the DCNN-based waterline extraction model for SAR images. The flowchart of this method is shown in Fig. 5. The elevation generation process can be divided into four steps:

  1. 1.

    Gaining the waterline information in a series of Sentinel-1 SAR images showing different tidal levels automatically by the trained DCNN-based model;

  2. 2.

    Discreting the lines into points and estimating their Lon/Lat position from original images;

  3. 3.

    Evaluating the water level of each point by the ocean tidal prediction model at the SAR imaging time;

  4. 4.

    At last, interpolating the resulting grid of quasi-contour lines to a DEM map.

Fig. 5
figure 5

The flowchart of the method for automatic topographic mapping of the tidal flats developed in this study

Full size image

According to the previous subsection, the DCNN-based model performs well, with little or no postprocessing required to obtain accurate waterlines, even for large-scale tidal flats like the Subei Bank. In addition, our method has extremely high extraction efficiency, with an average of two seconds per SAR image (2229 \(\times \) 2005 pixels, based on the NVIDIA Tesla V100 32GB GPU). According to Zhang et al. [31], the TPXO tide model [5] perform well in the tidal phase in our study area. However, this tidal model presents a systematic underestimation of tidal amplitude. Then, the in-situ water level data from two tidal gauge stations in our study area were used to calibrate this tide model (see [31] for details). The corrected TPXO tide models with Tidal Model Driver software are employed as the ocean tidal prediction model to evaluate the tidal level for each point of each waterline on this method.

Fig. 6
figure 6

Assembled tidal level evaluated waterlines extracted manually from all SAR images acquired in 2019

Full size image

We first used the waterlines of 2019 to verify the accuracy of the waterline method in measuring tidal flats elevation in our study area. We eliminated five scenes with wind speed greater than 10 m/s, which have a large offset from their original location caused only by tidal fluctuation [25]. Then the remaining waterlines from 47 SAR images was assigned with the tidal level value using the corrected ocean tidal model (see Fig. 6).

Finally, as shown in Fig. 7, these points/lines were interpolated to obtain gridded DEM of the large-scale tidal flats in our study area. One transect line of measured topographic data, which were acquired by an in-situ survey in May 2019, was used to test the accuracy of the derived DEM. The mean absolute error along this transect line is about 0.3 m (see Fig. 8).

Among the waterline method steps, the most time-consuming one is to extract the waterline, especially for SAR images. With the support of the DCNN-based automatic waterline extraction model, the efficiency of implementing this method can be significantly improved. We took the generation of the tidal flats’ DEM for 2018 as an example. A total of 29 pre-processed Sentinel-1 SAR images throughout the year were collected with Google Earth Engine and used as inputs to the DCNN-based model to obtain the geolocation of waterlines quickly. The final gridded DEM result for 2018 is shown in Fig. 9a after the same subsequent processes such as tidal level evaluation and spatial interpolation. In addition, interannual topographic changes can be analyzed by subtracting these two waterline-derived DEMs. As shown in Fig. 9b, the topography of these large-scale tidal flats changes significantly in two years under the action of strong tidal currents [2]. The erosion-deposition balance showed a net deposition of 0.12 km\(^3\) from 2018 to 2019. It implies our presented methodologies are also suitable for rapid monitoring the morphological and sedimentary changes of large-scale intertidal areas.

Fig. 7
figure 7

The derived DEM result of Subei Bank for 2019 using the waterline method (overlaid on the Sentinel-1A SAR image acquired at GMT 09:55, 20 November 2019)

Full size image
Fig. 8
figure 8

Topography profile comparison between the derived DEM and the in-situ transect (solid black line in Fig. 7)

Full size image

5 Discussions

Because of the frequent lack of consistent, sufficient intensity contrast between land and water regions and the complications of distinguishing waterline from other object boundaries, waterline extraction is harrowing with most general-purpose edge detectors or image segmentation techniques, especially for radar images in the intertidal areas. Previous studies used edge detection methods where a thresholding process was necessary at some point under relatively complex imaging conditions (such as the methods developed by [16, 21], and [9]). In addition, with the unprecedented amount of data containing waterline information available, an automatic extraction method should be prioritized. The DCNN-based method developed in this study performed well for automatic waterline extraction from SAR imagery in large-scale tidal flats area under changeable imaging conditions.

Fig. 9
figure 9

a The derived DEM result of Subei Bank for 2018 using the waterline method based on the waterlines automatically extracted by our DCNN model; b Tidal flats DEM of difference map between 2019 and 2018 (both overlaid on the Sentinel-1A SAR image acquired at GMT 09:55, 20 November 2019)

Full size image

With the support of big data platforms such as Google Earth Engine and the ocean tidal prediction model, we developed a waterline method-based workflow that can quickly obtain relatively accurate DEM of tidal flats after extracting multi-temporal waterlines from SAR images under different tidal levels. This technique provides an efficient method for the rapid analysis of large-scale tidal flat topography evolution, which is of great significance for applying SAR images to monitoring coastal terrains.

6 Conclusions

This chapter proposes a DCNN-based method to extract waterlines automatically from SAR images. Our approach shows a relatively high extraction accuracy for the waterlines in complicated large-scale tidal flats (the mean precision and recall are 0.92 and 0.77, respectively) and efficiency (several seconds per image) simultaneously. This chapter also presents the first attempt for intertidal DEM generation of the Subei Bank using the waterline method by analyzing high spatial resolution SAR images. The DEM results show that, in general, there is a good agreement between the derived elevation and in-situ topographic data, implying that the waterline method based on SAR images can be used for large-scale tidal flats such as the Subei Bank area. Furthermore, based on the waterline extraction model and the waterline method, we developed a novel workflow for automatic topographic mapping of large-scale tidal flats, which has excellent potential for rapid analysis of intertidal topography evolution.