1. Introduction
Extremely heavy rainfall phenomena frequently occur in China. According to data released by China’s Ministry of Emergency Management, floods affected a total of 59.01 million people in 2021. These floods caused direct economic losses of 245.89 billion yuan, accounting for 73.62 percent of natural disasters’ total direct economic losses. After the flooding disaster, quickly and effectively identifying the storm inundation area and the damage to the features is an essential guide for the relevant departments to grasp the disaster data and improve the emergency rescue and disaster prevention and mitigation capabilities.
In recent years, the rapid development of remote sensing technology has made it widely used in pre-disaster prediction and post-disaster rescue [
1]. The methods to achieve water body extraction by remote sensing flood dynamic monitoring can be classified into three types: threshold, object-oriented, and deep learning [
2]. The threshold method, also known as model classification, classifies the images by analyzing the spectral characteristic curves of water bodies. Select a suitable band to construct a model and an appropriate classification threshold to obtain a binary map of water bodies and non-water bodies [
3]. The main methods of image segmentation based on thresholding are the histogram bimodal method, Otsu maximum interclass variance method, EM algorithm (Expectation-Maximum), water body index method, etc [
4]. Mason et al. monitored flooding in urban and rural areas in the UK using a threshold method and change monitoring [
5]. Martinis et al. fully automated near real-time flood detection by fuzzy logic and threshold methods for Germany, Thailand, Albania, and Montenegro [
6]. Mcfeeters used TM images in green and near-infrared bands to construct the Normalized Difference Water Index (NDWI) to extract water body information within the city limits, which can better suppress vegetation information [
7]. The NDWI can hide the vegetation information and suppress the influence of soil, buildings, and shadows to highlight the water body information. This method is simple, fast, efficient, and widely used in water body extraction. In addition, it can achieve better classification results in plain areas with slight variations in topographic relief [
8,
9].
The object-oriented method takes the image object as the primary processing unit. It makes homogeneous image elements into objects of different sizes by various segmentation algorithms, thus realizing the extraction of image information by an object as a unit [
10]. With the increasing spatial resolution of remote sensing images, the classification methods to achieve target feature extraction using the images’ rich spectral and complex texture features are becoming increasingly sophisticated [
11]. Yu et al. introduced multiple SAR image texture features for water body information extraction and compared them with the traditional classification method to improve water body extraction accuracy [
12]. Foroughnia uses supervised and unsupervised classification methods, combined with multispectral and SAR data, to assess the precision and accuracy of flood extraction [
13].
Deep learning has been widely used in remote sensing water extraction with its unique advantages, such as powerful feature representation and automatic feature learning from data through deep neural network structures [
14], commonly used deep learning image segmentation algorithms FCN (full convolutional networks) [
15] and U-Net networks [
16]. Li et al. combined multi-temporal TerraSAR-X data and interferometric coherence as training samples to propose an active self-learning time-integrated convolutional neural network framework (A-SL CNN) [
17]. Xu presents a novel Synthetic Aperture Radar (SAR) image change detection method that integrates effective image preprocessing and Convolutional Neural Network (CNN) classification [
18]. Results show that the proposed method has higher accuracy in comparison with traditional change-detection methods. Wang et al. applied the constructed full convolutional neural network model for water body extraction experiments. The results showed that the fully convolutional neural network model is more automated, better applicable, and has higher extraction accuracy than the traditional threshold method for water body extraction [
19].
Since floods are mostly accompanied by various types of cloudy and rainy weather extremes, traditional optical remote sensing images cannot obtain ground information in a timely and accurate manner. SAR satellites can acquire ground information around the clock and in all-weather due to their wavelength characteristics [
20]. More and more methods are based on SAR images to monitor the changes in water bodies, which has become an indispensable data source in flood emergency monitoring [
21]. However, most current studies mainly focus on improving the accuracy of water body extent extraction, and relatively little research has been conducted on the damage to road networks in flooded areas. The urban road network is dense in form and complex in structure, carrying various types of traffic flows, which is the key to maintaining the city’s regular operation. Therefore, it is essential to identify the inundated location and damage to the road network under extreme weather conditions for post-disaster relief. In the event of flooding, due to the number of satellite constellations and revisit cycles, a specific type of satellite often does not meet the longtime series monitoring requirements for flooding, which also becomes an important problem for flood rescue and post-disaster assessment.
To quantitatively monitor the road damage in the “7-20” rainstorm in Zhengzhou, this paper designs a scheme to extract the road network damage in the flood disaster by combining the heterogeneous SAR data and the road network data before, during, and after the rainstorm. At the same time, a time-series monitoring of the disaster process of the “7-20” mega rainstorm is carried out. The length of flooded roads was extracted by the SAR image threshold segmentation method and GIS spatial analysis method to provide data support for quantitative road damage assessment.
2. Study Area and Data
2.1. Study Area
Zhengzhou is a megacity and one of the major economic centers in central China, the capital city of Henan Province, and an important railroad and highway transportation hub in China. 12.6 million people live in Zhengzhou City, the first in the province, according to the seventh national census in 2020. Zhengzhou is bordered by the Yellow River to the north and has 124 large and small rivers in its territory, spanning two significant basins: the Yellow River and the Huai River. The Yellow River basin includes parts of Gongyi and Shangdi, with an area of 2011.8 square kilometers, accounting for 27% of the city’s total area. The Huai River basin includes all of Xinzheng, Zhongyuan District, Erqi District, Guancheng District, and parts of Xinmi, Jinshui District, and Huiji District, with an area of 5499.5 square kilometers, accounting for 73% of the city’s total area. The city has 124 rivers of various sizes, with 29 rivers with larger watershed areas (≥100 square kilometers), including 6 in the Yellow River basin and 23 in the Huai River basin. The rivers crossing the border are the Yellow River and the Ilo River.
In July 2021, under the guidance of the airflow of the Pacific subtropical high pressure, a large amount of water vapor was continuously transported from the sea to the land, which was influenced by the topography to collect rain within Henan Province. From 18:00 on 18 July to 0:00 on 21 July, Zhengzhou received heavy rainfall, with a cumulative average precipitation of 449 mm. The single-day downpour broke the 60-year historical record since the establishment of the Zhengzhou weather station in 1951. This exceeded the regional flood control and drainage capacity [
22]. The Investigation Report of the “720” Very Heavy Rainstorm Disaster in Zhengzhou states that the heavy rainstorm caused extensive flooding in urban and rural areas, severe flooding in urban streets and depressions, short-lived flooding in rivers and reservoirs, and direct economic losses of 40.9 billion yuan. This study selected the road network in the main urban area of Zhengzhou City (Jinshui District, Huiji District, Zhongyuan District, Erqi District, and Guancheng Huizu District) for monitoring, and the specific study area is shown in
Figure 1.
2.2. Different SAR Datasets
Sentinel-1A is the first satellite developed by the European Commission and the European Space Agency for the Copernicus Global Earth Observation Project, which was launched in April 2014 with C-band imaging. Sentinel-1A contains four operating modes: SM, IW, EW, and WV, including an interferometric wide-field mode with a resolution of 5 m × 20 m and an amplitude of 250 km and a revisit period of 12 days. Sentinel-1A SAR data are freely available, providing rich data support for global scholars to monitor the global land and coastal zone.
The GF-3 satellite, the first Chinese high-resolution SAR remote sensing satellite with 1 m spatial resolution, was launched on 10 August 2016, using C-band imaging and containing 12 imaging modes. Depending on the imaging modes, GF-3 can provide multi-polarized SAR images with 1 m to 500 m resolution and 10 km to 650 km width to achieve global monitoring of ocean and land resources [
23]. The imaging mode used in this paper is fine striping with HV polarization and 10 m azimuthal resolution of the image.
Due to the satellite revisit cycle at the time of the “7-20” rainstorm disaster, this paper combines two kinds of heterogeneous SAR satellites, Sentinel-1A and GF-3, as well as the observation data of the “7-20” rainstorm in Zhengzhou. The Sentinel-1A VH SAR image was imaged on 15 July 2021. The two views of GF-3 SAR images were imaged on 20 and 22 July 2021, respectively, and HV polarization was used. The specific parameters of the heterogeneous SAR images with different phases are detailed in
Table 1. Each SAR image is shown in
Figure 2.
3. Methods
This paper used VH-polarized Sentinel-1A images on 15 July 2021, and HV-polarized GF-3 images on 20 and 22 July. We selected the dB values of typical water bodies in different SAR images through image alignment, multi-viewing, filtering, geocoding, radiometric calibration, and other processing. Then, we applied the threshold segmentation method to extract water bodies in different SAR images to generate before, during, and after the rainstorm. The binary maps of water bodies and non-water bodies are generated before, during, and after the rainstorm. After acquiring the changes in water bodies, this paper combines the vector data of the Gaudet road network to count the length of affected roads using the GIS spatial analysis method.
Figure 3 shows the road damage monitoring flow chart in Zhengzhou City during the “7-20” rainstorm using heterogeneous SAR images.
3.1. SAR Dataset Processing
The pre-processing of the heterogeneous SAR images of Sentinel-1A and GF-3 mainly includes the steps of image alignment, multi-view, filtering, radiation correction, etc.
Figure 4 shows the processing of the GF-3 image on 20 July.
- (1)
Multi-looking processing: The Sentinel-1A SLC image is multi-view processed with a view number ratio of 4 × 1, and the GF-3 SLC image is multi-view processed with a view number ratio of 2 × 2 to attenuate the influence of coherent speckle noise information on the image quality.
- (2)
SAR image filtering: In this paper, the enhanced Lee filtering method is used for SAR image filtering, which effectively overcomes the shortcomings of the traditional Lee filtering method for non-homogeneous areas with poor filtering effects and adopts the process of distinguishing the target distribution in the image area and divides the SAR image area into the homogeneous area, the non-homogeneous area, and the separation point target area. In this paper, the enhanced Lee filtering method with a window size of 5 × 5 is selected to filter SAR images, which can effectively maintain the edge information of radar images while suppressing noise information.
- (3)
Geocoding and radiometric calibration: Geometric and radiometric corrections are performed to eliminate distortions and obtain the backscatter coefficients of the images by using the 30 m resolution digital elevation data NASA SRTM released.
- (4)
Image registration: By selecting control points on the image, the three SAR images taken before, during, and after the rainstorm are registered and geometrically corrected.
The SAR system can obtain the power ratio of the measured emission and return pulses, and this ratio (backscattering) is projected to the slant range geometry to better compare SAR images’ geometric and radiation characteristics under different SAR sensors and receiving modes. It is necessary to perform geometric and radiation calibration on the slant range SAR data and convert it into a geographic coordinate projection.
The range of geometric distortion in the SAR image is large, mainly caused by the change in terrain. Based on the given digital elevation model, the relationship between the three-dimensional coordinates of the ground point and the two-dimensional coordinates of the slant distance image is established by the distance-Doppler geometric model and the backward projection algorithm. The range-Doppler model is expressed as [
20,
21]:
where
is the tilt range,
and
are the sensor and backscatter unit positions,
and
are the sensor and backscatter unit velocities,
is the carrier frequency,
is the speed of light, and
is the processed Doppler frequency.
For better comparison of heterogeneous SAR image data, the radar data need to be radiometrically calibrated using the radar equation to achieve calibration of the radar backscatter coefficient. The radar backscatter coefficient refers to the radar reflectivity per unit area of the target in the incident direction, and the backscatter coefficient can also be regarded as a combination of three elements: unit cross-section, reflectivity, and directionality. The general form of the radar equation is:
where,
is the transmitting power of the radar transmitter,
is the effective scattering transceiver area,
is the slope distance from the antenna phase center to the target point,
is the electromagnetic wave wavelength, and
is the radar scattering cross section.
3.2. Road Network Processing
To monitor the damage to roads in the main urban area caused by the July 20 rainstorm, the author obtained the road network data of the main urban area of Zhengzhou in 2021 using the Python web crawler tool. Road information includes ten categories of urban first-grade roads, urban second-grade roads, urban third-grade roads, urban fourth-grade roads, expressways, national roads, provincial roads, railways, county roads, and township roads in the main urban area of Zhengzhou. The processing of road network data mainly includes:
- (1)
Culling small roads. The rich variety of roads in the Gaode road network data contains some redundant roads with too short lengths. Fine roads less than 5 m were excluded from this study to reduce data redundancy.
where
is the set of road elements and
L is the length of the road.
- (2)
Topology check. The original road network data is intricate and overloaded with details, so it is necessary to perform topological checks on the roads to avoid errors in the subsequent analysis. We check each line element for topological errors such as line segment self-intersection, line overlap, hanging points, and pseudo-nodes. Finally, we obtain the corrected road network data for the study area shown in
Figure 5.
Three feature areas with different spatial distributions were selected to verify whether the processed road network data fit the actual road network. Their processed road network information was confirmed with the corresponding Google images. The three areas have different road network characteristics due to different geomorphological and spatial distribution characteristics. Area A is located near the intersection of Science Avenue and Xushui River Road, where the road network is dominated by Xushui River Road on both sides of the river, and the road network information of Science Avenue and Xushui River Road can be seen from the enlarged map of area A in
Figure 5. The B area is located near the Longhubei subway station. The road network information in this area is relatively simple, and the comparison is also apparent. C area is located at the intersection of West Haoheng Road and West Fourth Ring Road, which is dominated by elevated bridges and has more accurate road network information. In summary, this paper’s processed road network information is reliable after comparing it with Google images.
3.3. Water Body Extraction
Threshold segmentation uses the gray difference between the target and the background in the image to be extracted. It separates the target from the background by setting different thresholds to divide the pixel level into several classes. The general process is to determine whether a pixel point in an image belongs to the target or background region by judging whether each pixel point’s feature attributes meet the threshold requirements, thus converting a grayscale image into a binary image.
where
is the image after thresholding,
denotes the image element value of the point
, the pixel marked as 1 corresponds to the target object, the pixel marked as 0 corresponds to the background, and
is the target object segmentation threshold interval.
Determining the threshold value of water bodies on different images is key to this experiment. In this paper, we use manual empirical selection to superimpose the SAR images to be processed on Google Earth, find the obvious water body sample points such as water bodies and lakes, and identify the DN value of the water body image element. The DN value of the image element is the backscattering coefficient of the image element in dB.
Figure 6 shows the backscattering coefficients of Sentinel-1A and GF-3 for two different SAR images of some typical water samples. Based on the analysis of the statistical characteristics of the sample points, the optimal interval where the water body thresholds are located is determined as [0, 0.001] for the Sentinel-1A image water body extraction threshold interval and [0, 0.009] for the GF-3 image water body extraction threshold interval in this experiment.
6. Conclusions
To solve the problems of difficult imaging and untimely acquisition of disaster information during flooding by traditional optical remote sensing means, this paper designs a time-series monitoring scheme for flooding by combining heterogeneous SAR images with the “7-20” rainstorm in Zhengzhou City as an example. By studying two types of heterogeneous SAR data, Sentinel-1A and GF-3, which were in transit during the “7-20” rainstorm in Zhengzhou City, the scheme successfully realized the dynamic monitoring and information extraction of flooding before and after the mega rainstorm by using the threshold segmentation method. Meanwhile, remote sensing and GIS data are fused to analyze the damage to urban roads under the influence of flooding, demonstrating the capability and application of obtaining disaster spatial and temporal information from multiple perspectives supported by multi-source SAR data. The results show that: (1) Guancheng District was the most affected area in the “7-20” rainstorm disaster in Zhengzhou City, and the total area of water bodies in this district increased by 33,402.2 m2 by 20 July 2021. The recovery of the affected area in Erqi District was slow. As of 22 July 2021, the water body area in Erqi District has been reduced by 12,461.1 m2 compared with that on July 20. This phenomenon is related to the dam failure of Guojiazui Reservoir in Erqi District. (2) In this rainstorm disaster, the total length of the road network in the main urban area of Zhengzhou was affected by 1324.63 km, of which the road network in Zhongyuan District was most seriously affected, reaching 349.17 km.
This scheme adopts two kinds of heterogeneous SAR images to obtain detailed information on flood time sequence changes in the “7-20” rainstorm disaster in Zhengzhou City. It demonstrates the advantage of a joint observation scheme using heterogeneous radar image data in extracting timely and accurate information on surface flooded water bodies in rainstorms and flood disasters.