Optimization of DG-LRG Water Extraction Algorithm Considering Polarization and Texture Information

Abstract

Flooding is one of the most frequent natural disasters at present, and can pose a serious threat to transmission towers. In response to the accuracy and timeliness requirements of flood emergency monitoring, a local region growth algorithm combining polarization and texture information is proposed for Synthetic Aperture Radar (SAR) image water recognition. Morphological methods and external geographic information are used to optimize the results, allowing for rapid extraction of the flood range. The method is validated using Gaofen-3 (GF-3) Fine Strip Imaging Mode II (FSII) SAR images covering Fangshan District in Beijing, China. The experimental results indicate that this method can obtain more effective water information compared to traditional threshold segmentation methods, and can also reduce the effects of noise and mountain shadows. It has good applicability and timeliness with respect to large-scale flood emergency disaster monitoring, and can help to rapidly and accurately obtain detailed information of flood-affected areas, thus providing reference for emergency rescue and disaster relief services.

1. Introduction

Flooding is one of the most frequent natural disasters which, due to its wide impact range and significant losses, often poses a threat to the economy and the safety of lives and property, as well as to the normal operation of transmission towers. The main task of flood monitoring is to quickly and accurately identify the range of flood water bodies [1]. Traditional flood monitoring requires on-site investigation and manual surveying, which is time-consuming, labor-intensive, and makes it difficult to meet the needs of large-scale monitoring [2]. Satellite-borne Synthetic Aperture Radar (SAR) microwave remote sensing, as a newly developed technology, has become the main technology used for remote sensing flood monitoring, due to its advantages of not being affected by temporal, spatial, or weather limitations [3,4].

At present, SAR image flood monitoring methods mainly include threshold and classifier methods [5]. A threshold method operates by selecting appropriate grayscale values in the image to distinguish between water and non-water bodies. Generally, this is based on the back-scatter coefficients of ground objects, selecting appropriate parameters based on histogram information to maximize the differentiation between water and non-water bodies. At present, commonly used methods include the maximum inter-class variance method [6], water index extraction method [7], and regional growth algorithm [8]. With the rapid development of machine learning in recent years, relevant approaches have gradually become widely used for water body recognition, including convolutional neural networks (CNNs) [9,10,11,12], random forests (RFs) [13], support vector machines (SVMs) [14,15], and other methods. Although machine learning algorithms have high recognition accuracy, their results are limited by the selection of features, and the process of training samples and models is associated with high time and cost requirements, making it difficult to meet the needs of large-scale flood emergency monitoring. The advantage of threshold methods is that they are simple, fast, and can adequately meet the requirements of emergency response. Despite this efficiency, threshold methods suffer from two unresolved challenges: (1) limited discriminative power: on-water features (e.g., mountain shadows, paved roads) with backscatter coefficients overlapping water bodies are frequently misclassified; (2) global threshold bias: a single threshold fails to account for local variations caused by wind-roughened water surfaces or submerged vegetation. Thus, it is necessary to utilize other methods in combination, such as contour modeling, morphology, and texture information, in order to improve monitoring accuracy [16].

Based on this, we propose a Dual Polarized and GLCM in Local Region Growth (DG–LRG) algorithm that takes into account the polarization mode and SAR image texture features, combining traditional region growth algorithms with the polarization mode, texture features, optimization algorithms, morphological methods, and external geographic information. The primary objective is to enable rapid, high-precision flood mapping without requiring training data. Unlike threshold methods relying solely on backscatter intensity, DG–LRG takes dual-polarized SAR data (HH/HV) and multi-scale texture features as primary inputs. The polarization modes enhance separability between water and specular surfaces, while texture statistics (e.g., homogeneity, contrast) extracted from GLCM suppress false alarms in heterogeneous regions. In this way, the flood range of Fangshan District is extracted, which was the most severely affected area in the “7.31” flood disaster in Beijing.

2. Research Area and Data Pre-Processing

2.1. Characteristics of the Research

At the end of July 2023, due to the impact of Typhoon “Dussuri”, significant rainfall occurred in the Beijing area, leading to floods. The severely affected Fangshan area (39°30′–39°55′ N, 115°25′–116°15′ E) is considered as the research area in this article, which is located in the southwest of Beijing. It is adjacent to Mentougou District to the north and Fengtai District to the northeast, faces Daxing District across the Yongding River to the east, and borders Zhuozhou City, Laishui County, and Yixian County in Hebei Province to the south and west, respectively. The terrain of this area is high in the northwest and low in the southeast, with mountainous regions (54% of total area) in the northwest and alluvial plains, depressions, and floodplains in the southeast. The elevation range is 30 m (southeastern plains, e.g., Doudian Town) to 2017 m (northwestern mountainous areas, e.g., Baihua Mountain) above sea level. Land use in 2023 comprised 60% forest/cropland (primarily mountainous oak/pine forests), 23% other areas (bare rock, grasslands), 15% urban/built-up zones (concentrated in Liangxiang and Chengguan subdistricts), and 2% water bodies, notably the Dashi River and reservoirs. Fangshan District is located in a warm temperate semi-humid continental monsoon climate zone, with complex landforms and significant differences in relative height between mountainous areas and plains, resulting in significant climatic differences. The annual precipitation in 2022 amounted to 589.4 mm, but in 29–31 July 2023, extreme rainfall triggered by Typhoon Doksuri reached 744.8 mm at Hebei Town Station (maximum hourly intensity: 111.8 mm/h), the highest recorded in Fangshan since 1951. In 2022, the region had a resident population of 1.31 million (Beijing Statistical Yearbook 2023) and an urbanization rate of 78.6%, below the Beijing municipal average of 87.6%. The elevation map of Fangshan District is shown in Figure 1.

2.2. Data Source and Pre-Processing

2.2.1. Data Introduction

GF-3 (Gaofen-3) is the first high-resolution multi-polar SAR satellite in China, providing support for the development of multi-polar SAR remote sensing, solving the problem of lack of SAR image data sources, and providing important data support for flood disaster research. We selected GF-3 data covering the research area with an imaging time of 6–7 am on 1 August 2023 as the main data for extracting the flood range after the disaster. The data comes from the China Resources Satellite Center [17]. The data type is Single Look Complex (SLC) format, and the polarization method is dual polarization with a resolution of 10 m. The detailed parameters are provided in the

Table 1. Basic parameters of GF-3.

Parameters	Basic Information
Imaging Mode	FSII
Polarization mode	Dual polarization (HH + HV)
Data type	SLC
Resolution/m	10
Imaging width/km	100

.

2.2.2. Data Pre-Processing

The obtained GF-3 image could not be directly used for water extraction, and required image pre-processing, such as radiometric correction, multi-look transformation, filtering, and geocoding.

• Radiometric Correction:

The radar sensor measures back-scatter, which is the ratio of the transmitted pulse to the intensity of the received signal. Due to the influence of multi-source errors, radiation error is present in the SLC image. In order to accurately reflect the echo characteristics of ground objects, it is necessary to convert the input signal into radar back-scatter coefficients. After radiometric correction and normalization of SAR images to a unified standard, the back-scatter intensity information obtained is not affected by the observation geometry of the SAR data, which is helpful for comparison and analysis.

2.

Multi-look:

The SLC image of the GF-3 satellite overlaps with the radar echo signal scattered by a single pixel, causing noise in the data. To improve the performance of the GF-3 image and improve the back-scatter estimation accuracy for each pixel, multi-look processing is performed. The SLC data after multi-look processing has improved resolution data, but spatial resolution is decreased. In this experiment, multi-look averaging was applied after radiometric calibration but before geocoding, using the ENVI SARscape5.6.2 toolbox’s Multi-looking operator with default settings. We applied four looks in the azimuth and range directions. This is consistent with the GF-3 FSII mode’s typical configuration.

3.

Filtering:

According to the characteristics of coherent noise in the image, the Lee filtering method is used to filter and process the image. The smoothing effect of this method is weak, it can suppress the noise of the GF-3 image, and it can better preserve the texture information of the image, which is conducive to water extraction.

In this experiment, we employed the ENVI SARscape toolbox’s Refined Lee Filtering operator. A 7 × 7 sliding window was chosen to balance noise suppression and edge preservation. Larger windows (e.g., 9 × 9) over-smooth textures, while smaller windows (5 × 5) retain noise.

4.

Geocoding.

Geocoding refers to the conversion of SAR data from slant-range coordinate systems to geographic coordinate systems. The process of geocoding involves three steps: the first is initial assumption of geometric transformation; the second is improving the geometric transformations; and the third is re-sampling one coordinate system to another. We selected ALOS DEM data with a resolution of 12.5 m for geocoding. The pre-processing result is shown in Figure 2

3. Optimized Local Region Growth Algorithm Considering Polarization and Texture Information

This article proposes an optimized local region growth algorithm that takes into account the polarization mode and texture features of SAR images for fast water extraction, denoted as DG–LRG. The main process of this method can be divided into four steps: the first step is to perform rough water extraction based on threshold segmentation on pre-processed dual polarization GF-3 images; the second step is to select seed points for regional growth based on the rough water extraction (see Section 3.2 for details); the third step is to combine polarization methods and texture features to locally optimize the region growth using certain criteria (see Section 3.3 for details); and the fourth step is optimizing the image segmentation results of regional growth by combining morphological theory and external geographic information data to obtain the final water extraction results. The implementation process is detailed in Figure 3.

3.1. Basic Principles of the Region Growth Algorithm

The region growth algorithm is an image segmentation method, the main idea of which is to gather adjacent pixels with high similarity for a certain aspect of the image to form a special region. In particular, the process involves gradually growing a pixel or sub-region outward in each region according to a specific growth rule, ultimately obtaining a complete closed region with the same features [18]. The number of original seed points determines the distribution and range of the segmentation region, such that the selection of the best seed points should be conducted in the central region and have a concentrated distribution. According to the similarity criterion, the pixels near the seed point with similar attributes are classified into one category. The similarity criterion can be formulated in terms of local features such as image grayscale values, spatial properties, and target texture features. The specific selection can be based on practical applications to ensure the accuracy of local segmentation.

3.2. Automatic Selection of Seed Points Based on DWI

3.2.1. Rough Extraction of DWI Water Based on Otsu

The key to region growth algorithms lies in the determination of seed points (i.e., the initially selected pixels). Traditional region growth algorithms require the manual selection of seed points, which cannot meet automation requirements. However, in the process of automatically selecting seed points, if the seed points are selected in a non-water body area, the final segmentation result may be significantly impacted. Therefore, we first use the Otsu method to extract the rough water bodies in the study area, which is then taken as the selection range for seed points.

Jia, inspired by NDVI, proposed the SDWI (Sentinel-1 Dual Polarized Water Index) based on the VV and VH dual polarization data of Sentinel-1 satellite. This index combines the db values of the two polarization methods to expand the differences between water bodies and other features [7]. The GF-3 satellite image used in our experiment also includes two polarization methods. According to the idea of SDWI, the Dual Polarized Water Index (DWI) is used for coarse extraction of water bodies in the study area. The formula for the DWI is

$$\mathrm{D}\mathrm{W}\mathrm{I}=\ln (10\times \mathrm{H}\mathrm{H}\times \mathrm{H}\mathrm{V}).$$

(1)

The Otsu maximum between-class variance method (also known as the Otsu algorithm) is currently considered to be the best method for automatic threshold selection. This method introduces between-class variance to describe the difference between the target terrain and the background. When the between-class variance between the two is maximal, it is considered that the feature difference between the two is the greatest. The formula is expressed as follows [19]:

$$e^2\left(K\right)=P_A{\left(\mathsf{\alpha }-{\mathsf{\alpha }}_{\mathrm{A}}\right)}^2+P_B{\left(\mathsf{\alpha }-{\mathsf{\alpha }}_{\mathrm{B}}\right)}^2,$$

(2)

where α denotes the mean of the image DWI, and αA and αB are the mean values for the target A and background B, respectively. When the value of e²(K) is maximal, K is the optimal threshold. When the variance difference between the region of interest image and the background image is maximized, theoretically speaking, the difference in elements between the two types of images is the greatest, the confusion between categories is the smallest, and the misclassification error of the segmentation results is also the smallest. The Otsu method can quickly extract the approximate range of water bodies and remove a large portion of non-water areas; however, its results often contain noise and non-water surface features with the same back-scattering characteristics as water bodies (e.g., airport runways, smooth roads, etc.). Therefore, in this experiment, the Otsu results were only used as the selection range for seed points.

3.2.2. Automatic Selection of Seed Points

Due to the presence of noise in the results of rough extraction of water bodies mentioned above, in order to ensure that seed points are selected in non-noisy areas, we use gradient amplitude and DWI variance as constraints, considering the nearest mean point to select seed points. The specific method is as follows: first, the initial water extraction result is extended by “a” unit from top, bottom, left, and right, then a square window with a side length of 2a + 1 is set, sliding in steps of (2a + 1)/2; next, the DWI variance and center pixel gradient amplitude within the window is calculated, as follows:

$$\mathsf{\sigma }={\displaystyle \frac{1}{{(2\mathrm{a}+1)}^2}}\sum {\left(\mathrm{d}\mathrm{w}\mathrm{i}\right(\mathrm{i},\mathrm{j})-\mathrm{u})}^2,$$

(3)

$$\mathrm{G}(\mathrm{x},\mathrm{y})=\sqrt{{(\mathrm{d}\mathrm{l}/\mathrm{d}\mathrm{x})}^2+{(\mathrm{d}\mathrm{l}/\mathrm{d}\mathrm{y})}^2},$$

(4)

where σ is the variance, dwi(i,j) is the DWI value of the pixel, u is the mean of the DWI within the region, G(x,y) is the gradient amplitude of the pixel, dl/dx is the rate of change of the pixel value in the x direction, and dl/dy is the rate of change of the pixel value in the y direction. The variance reflects the overall degree of change in pixel values within the window, while the gradient amplitude reflects the rate of change of the center pixel compared to surrounding pixels. If the gradient amplitude is large, it indicates that a pixel may be located in the edge area of an object. Generally, seed points should be selected in areas with smaller variance and gradient amplitude. Finally, pixels within the window with the closest pixel value to the mean are selected as candidate seed points. The candidate seed points outside the coarse water range are then removed to generate the final seed point set.

3.3. Local Region Growth Criteria Based on Polarization and Texture Information

3.3.1. Extraction of Texture Features

The Gray-level co-occurrence matrix (GLCM) is a method for extracting texture features from grayscale images. The co-occurrence matrix is obtained by calculating the grayscale image, following which some feature values of the matrix are obtained by calculating the co-occurrence matrix to represent certain texture features of the image. This method defines a total of 14 texture features [20,21]. Due to the differences in back-scatter information between SAR images with different polarization methods for the same type of ground object, for this experiment, the db values of the GF-3 satellite were used for GLCM texture feature extraction. In our experiment, eight commonly used statistics—the Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Angular Second Moment, and Correlation—were selected to form a feature matrix to characterize the texture of SAR images. Due to slight differences in back-scatter information and different texture features in SAR images with different polarization methods, in order to reduce redundant information and improve efficiency, while also considering the texture differences of polarization methods, principal component analysis was performed on the feature matrices of the two polarization methods and the first three principal components were extracted for the final statistics. Through experiments, three types of statistical measurements were extracted for the HH polarization GF-3 images: Homogeneity, Entropy, and Contrast. Meanwhile, three statistical measurements were also extracted for the HV polarization GF-3 images: Homogeneity, Dissimilarity, and Angular Second Moments. The expressions and descriptions for these feature statistics are provided in the

Table 2. Texture feature expressions and descriptions.

Texture Features	Expression	Description
Homogeneity	$\mathrm{H}\mathrm{o}\mathrm{m}={\displaystyle \sum _{\mathrm{i}}}{\displaystyle \sum _{\mathrm{j}}}{\displaystyle \frac{1}{1+{(\mathrm{i}-\mathrm{j})}^2}}{\mathrm{P}}_{\mathrm{i},\mathrm{j}}$	Used to measure the local variation of image texture; water blocks are highly homogeneous areas in the image.
Entropy	$\mathrm{E}\mathrm{n}\mathrm{t}=-{\displaystyle \sum _{\mathrm{i}}}{\displaystyle \sum _{\mathrm{j}}}{\mathrm{P}}_{\mathrm{i},\mathrm{j}}\times \mathrm{l}\mathrm{n}{\mathrm{P}}_{\mathrm{i},\mathrm{j}}$	Reflects the rate of change in image grayscale; the entropy value in the center of the water body is small, while the entropy value at the shore is large.
Contrast	$\mathrm{C}\mathrm{o}\mathrm{n}={\displaystyle \sum _{\mathrm{i}}}{\displaystyle \sum _{\mathrm{j}}}{\mathrm{P}}_{\mathrm{i},\mathrm{j}}\times {(\mathrm{i}-\mathrm{j})}^2$	Reflects the clarity of the image and the depth of the texture grooves.
Dissimilarity	$\mathrm{D}\mathrm{i}\mathrm{s}={\displaystyle \sum _{\mathrm{i}}}{\displaystyle \sum _{\mathrm{j}}}{\mathrm{P}}_{\mathrm{i},\mathrm{j}}\times \left|\mathrm{i}-\mathrm{j}\right|$	Linear correlation with contrast.
Angular Second Moment	$\mathrm{A}\mathrm{s}\mathrm{m}={\displaystyle \sum _{\mathrm{i}}}{\displaystyle \sum _{\mathrm{j}}}{{\mathrm{P}}_{\mathrm{i},\mathrm{j}}}^2$	Reflects the thickness and fineness of image texture, with a focus on the low uniformity of internal texture in large surface water bodies.

. In the table, i and j represent the grayscale progression of the image, while Pi,j represents the probability of pixels with grayscale j appearing from pixels with grayscale i in the image. In our experiment, we divided the db values of the SAR images into 64 levels and moved them in four directions: 0°, 45°, 90°, and 135°. For texture feature extraction, the selected window was 7 × 7.

3.3.2. Determination of Objective Function

To determine the growth criteria for region growth, it is necessary to select a discrimination threshold. In our experiment, the Euclidean distance of the feature vector was selected as the judgment basis, and the extracted texture features were combined with the image coordinate matrix (x,y) direction to form a texture–space five-dimensional matrix. In order to avoid a certain component in the vector having a significant impact on clustering, before calculating the Euclidean distances between pixels, the distances of each component of the five-dimensional vector were normalized first, following which the normalized distances were combined into a composite distance, which was used as a discriminant criterion. The texture feature distance dp and spatial distance ds of elements m and n in the matrix are calculated as follows:

$${\mathrm{d}}_{\mathrm{p}}=\sqrt{{({\mathrm{p}}_{1\mathrm{m}}-{\mathrm{p}}_{1\mathrm{n}})}^2+{({\mathrm{p}}_{2\mathrm{m}}-{\mathrm{p}}_{2\mathrm{n}})}^2+{({\mathrm{p}}_{3\mathrm{m}}-{\mathrm{p}}_{3\mathrm{n}})}^2},$$

(5)

$$d_s=\sqrt{{({\mathrm{x}}_{\mathrm{m}}-{\mathrm{x}}_{\mathrm{n}})}^2+{({\mathrm{y}}_{\mathrm{m}}-{\mathrm{y}}_{\mathrm{n}})}^2},$$

(6)

We compute the Euclidean distance between the first three principal components (PCs) of GLCM features. PCs reduce redundancy while preserving 95% variance, the texture–space feature vector for each pixel is defined as [p1, p2, p3, x, y]where p1, p2, and p3 are the first, second, and third principal components of the texture feature information of the elements in the matrix, respectively, and x and y are the row and column coordinates of an element, respectively. The Euclidean distance between pixel coordinates ensures spatial continuity during region growth. The composite distance D can be expressed as follows:

$$\mathrm{D}=\sqrt{{({\mathrm{d}}_{\mathrm{p}}/{\mathrm{d}}_{\mathrm{p}\mathrm{m}})}^2+{({\mathrm{d}}_{\mathrm{s}}/{\mathrm{d}}_{\mathrm{s}\mathrm{m}})}^2},$$

(7)

where d_pm and d_sm represent the maximum values of the texture feature distance and spatial distance, respectively, representing the texture and spatial differences between pixels. The composite distance D integrates texture and spatial distances to guide region growth. We normalized dp and ds, divided by their maximum values within each local sub-region. This normalization eliminates dimensional differences between texture features (unitless) and spatial coordinates (pixels), allowing equal contributions from both components. The normalized Euclidean norm ensures that texture and spatial terms contribute equally to the similarity measure. This approach is widely adopted in multi-feature fusion [22].

The composite distances for the two polarization methods are denoted as D_HH and D_HV, respectively.

Due to the uneven degree of back-scatter during SAR imaging, as well as the presence of noise and homo-spectral anomalies in the intensity map, it is difficult to accurately extract water bodies within the study area using the global region growth algorithm. Based on this, we propose a local optimization region growth algorithm, which determines an objective function and divides the study into N × N local regions. The value of N can be roughly set according to the size of the influence and terrain; in this experiment, N = 2. The choice of N = 2 aligns with the topographic diversity of Fangshan District. The northwestern mountainous areas and southeastern plains (Figure 1) exhibit distinct backscatter patterns. Dividing the image into 2 × 2 sub-regions allows the optimization algorithm to separately optimize thresholds for mountain-shadowed and open-water areas, whereas a global approach (N = 1) would over-smooth these transitions. However, an excessively large value of N may significantly compromise computational efficiency, thereby failing to meet the requirements for emergency flood monitoring. Therefore, the image is divided into 2 × 2 sub-regions to balance local adaptation and computational efficiency. The optimal solution of the objective function is obtained within each local region, which not only ensures the scientific nature of the growth criteria but also helps to highlight local features.

To ensure that the most suitable growth criteria can be obtained for different local regions, the idea of Shannon entropy is introduced into the determination of the objective function, combined with the two polar recombination distances. The entropy value of a certain local region can be expressed as

$$\mathrm{H}=-({\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime }\times \mathrm{l}\mathrm{n}{\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime }+{\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime }\times \mathrm{l}\mathrm{n}{\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime }),$$

(8)

where D_HH’ and D_HV’ denote D_HH/e and D_HV/e, respectively. To further ensure the segmentation effect, extreme measures are introduced, and the final expression of the objective function H is

$$\begin{gathered} \mathrm{H}=-\left(\right({\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime }\times \mathrm{l}\mathrm{n}{\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime }+{\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime }\times \mathrm{l}\mathrm{n}{\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime })+\mathrm{m}\mathrm{a}\mathrm{x}{\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime }\mathrm{l}\mathrm{n}{(\mathrm{m}\mathrm{a}\mathrm{x}\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime })+ \\ \mathrm{m}\mathrm{a}\mathrm{x}{\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime }\mathrm{l}\mathrm{n}\left({\mathrm{m}\mathrm{a}\mathrm{x}\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime }\right)+\mathrm{m}\mathrm{i}\mathrm{n}{\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime }\mathrm{l}\mathrm{n}\left({\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{D}}_{\mathrm{H}\mathrm{H}}^{\prime }\right)+\mathrm{m}\mathrm{i}\mathrm{n}{\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime }\mathrm{l}\mathrm{n}\left({\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{D}}_{\mathrm{H}\mathrm{V}}^{\prime }\right)). \end{gathered}$$

(9)

When the objective function H takes its maximum value, it indicates that the composite distance difference between the two polarization methods is maximal, and this distance is the maximum threshold that the seed point can grow to.

3.3.3. Local Optimization Based on PSO

As the above objective function does not have an analytical solution, we use the Particle Swarm Optimization (PSO) algorithm to search for local regions and determine the optimal solution. Particle swarm optimization is an optimization algorithm based on swarm intelligence, which abstracts the solution of optimization problems into the optimal state of particle motion. For this algorithm, we need to set the initial particles size M, the speed of the initial particles are

$$\mathrm{V}\left(0\right)=\left({\mathrm{V}}_1\right(0),\dots \dots ,{\mathrm{V}}_{\mathrm{M}}(0\left)\right),$$

(10)

and the initial positions are

$$\mathrm{X}\left(0\right)=\left({\mathrm{X}}_1\left(0\right),\dots \dots ,{\mathrm{X}}_{\mathrm{M}}\left(0\right)\right),$$

(11)

where V_i(0) and X_i(0) represent the initial velocity and initial position of the ith particle, respectively. The updated formulas for the velocity and position vectors of the particles are

$${\mathrm{V}}_{\mathrm{i}}\left(\mathrm{t}\right)=\mathrm{k}\left[\begin{array}{c}{\mathrm{V}}_{\mathrm{i}}\left(\mathrm{t}-1\right)+{\mathrm{C}}_{1}rand\left({\mathrm{P}}_{\mathrm{i}}\left(\mathrm{t}-1\right)-{\mathrm{X}}_{\mathrm{i}}\left(\mathrm{t}-1\right)\right)\\ +{\mathrm{C}}_{2}rand\left(\mathrm{G}\left(\mathrm{t}-1\right)-{\mathrm{X}}_{\mathrm{i}}\left(\mathrm{t}-1\right)\right)\end{array}\right],$$

(12)

$$X_i\left(t\right)=\left[{\mathrm{X}}_{\mathrm{i}}\left(\mathrm{t}-1\right)+{\mathrm{V}}_{\mathrm{i}}\left(\mathrm{t}\right)\right],$$

(13)

where [*] denotes the rounding function, “rand” represents a random value within [0, 1], V_i(t) is the update speed of particle i in the tth iteration, which is required to be in the range (−Vmax, Vmax), and Vmax is the maximum update speed (set in advance). Furthermore, the parameter k is the contraction factor (in this experiment, k = 0.742), and C₁ and C₂ are learning factors. C₁ controls the ability of particles to learn from themselves, while C₂ controls the ability of particles to learn from groups. The initial state X(0) and initial velocity V(0) of the particle swarm can be randomly selected, but it is required that X(0) is within the solution space and V(0) is within (−Vmax, Vmax). P_i(t) represents the optimal position that particle i passed through during the tth iteration, while G(t) represents the optimal position that all particles passed through during the tth iteration in the process of finding the optimal solution. The optimal solution obtained within each local region is used as the growth criterion for water extraction, and all growth regions are merged to obtain the final water extraction result.

3.4. Optimization of Image Segmentation Results

The results obtained above take into account the characteristics of GF-3 polarization and water texture information, which can effectively reduce the impact of noise on the results and also reduce false extraction situations (e.g., misclassification of smooth road surfaces, mountain shadows, and airport runways). However, in flooded areas, the water environment is relatively complex. Extreme weather is often accompanied by high wind speeds, and the submerged vegetation, crops, and turbidity in a water body can increase its surface roughness, affecting the back-scatter coefficients of radar signals. This can lead to isolated pixels in the water extraction results, resulting in burrs and voids [23]. In addition, although the DG–LRG method proposed in this article can reduce the impact of mountain shadows to some extent, it is difficult to distinguish between these shadows and water bodies located in a mountainous area, especially when they are connected in an image. Based on this, morphological methods and external DEM data were introduced in the experiment, in order to optimize the water extraction results.

Morphology defines four basic operations: expansion, corrosion, opening, and closing. Assuming that the structural element is A and the target image is B, common edge detection operators are defined as follows:

$$\mathrm{A}○\mathrm{B}=\left(\mathrm{A}\ominus \mathrm{B}\right)\oplus \mathrm{B},$$

(14)

$$\mathrm{A}\mathrm{●}\mathrm{B}=\left(\mathrm{A}\oplus \mathrm{B}\right)\ominus \mathrm{B}$$

(15)

where A○B is an open operation, A●B is a closed operation, ⊕ is an expansion operation, and ⊖ is a corrosion operation. In order to remove isolated pixels from the water extraction results and fill in voids, we used an open-then-close operation to optimize the results.

For the mountain shadow areas that still exist in the water extraction results, external DEM data are used as auxiliary information for their removal. The specific method is to register the SAR images with DEM data, calculate the radar signal overlay shadow area based on the satellite incidence angle and flight azimuth angle, as well as the slope and aspect generated by the DEM, and define the slope threshold. Areas with slope values greater than 15° are defined as the final mountain shadow areas, Filtering out this portion of pixels from the extracted water pixels yields the optimized water extraction result.

4. Results and Discussion

4.1. Flood Range Extraction Results

To verify the effectiveness of the method proposed in this article, typical water bodies such as lakes, mountain water bodies, and river channels were selected for detailed analysis within the study area. The Otsu threshold segmentation results, DG–LRG method extraction results, and the optimization results were compared using morphological and external DEM data.

Figure 4 shows the extraction results for a lake area. It can be seen that, in the intensity map, some areas within the lake have complex water environments, and there are differences in pixel grayscale compared to other water bodies. Therefore, in the Otsu threshold segmentation results, some water bodies are missing, and there is obvious noise throughout the entire area, leading to many omissions and misclassifications in the Otsu results. In the results of the DG–LRG method proposed in this article, the noise interference was significantly reduced and the water area was relatively complete. However, there were still some isolated pixels forming voids, and the extraction of some small areas was incomplete. In comparison, the optimized results were more complete and smoother, with a significant reduction in internal voids and fewer misclassifications and omissions.

Figure 5 shows the extraction results for a mountain water body. It can be seen that, due to the presence of mountain shadows, there were a large number of misclassifications in the Otsu threshold segmentation results, and the river was not continuous, showing an intermittent state. In the DG–LRG results, the problem of mountain shadows was largely solved; however, many mountain shadows adjacent to rivers were also shown as eligible areas, resulting in misclassification. From the results of the optimized model, it can be seen that, after external DEM optimization, shadows adjacent to water bodies were basically eliminated. However, there was also a certain deviation between the results and the actual situation and, due to differences in registration accuracy and resolution, there may be errors, such as the removal of rivers. However, compared to the results before optimization, the accuracy was significantly improved.

Figure 6 shows the extraction results for a river channel area. Through visual interpretation of the optical image, it can be seen that the part that intersects with the river is a road. In the Otsu threshold segmentation results, some roads were mistakenly classified as water bodies, and the river itself appears discontinuous, with obvious noise effects around the river channel. This noise is significantly reduced in the DG–LRG results, the river channel is more complete, and the situation of roads being misclassified is also improved. The effect of the optimized model was not as obvious in this situation, mainly manifested as a reduction in internal voids in the river, smoother edges, and the result were more in line with human visual characteristics.

4.2. Evaluation of Water Extraction Accuracy

Due to the time required for the flood to recede, a 2 m resolution optical image without cloud cover for Fangshan District, on August 3^rd, was selected for high-precision identification of water bodies. In terms of cloud coverage, the selected high-resolution optical imagery exhibits <5% cloud cover, with cloud-contaminated pixels masked out using 10-m resolution Sentinel-2 data. For water body extraction, the process was conducted by three experienced remote sensing image analysts, achieving 95% inter–annotator agreement. The recognition results are shown in Figure 7.

This result was used as reference for the accuracy evaluation. Compared with the water body results extracted by GF-3, the three indicators of Recall, Precision, and False Alarm Rate (FAR) were used for accuracy evaluation of the DG–LRG method.

In particular, Recall is the proportion of correctly detected positive cases among all actual positive cases, which can measure a model’s ability to recognize positive cases; Precision is the proportion of correctly detected positive cases among all positive cases detected by the model, which can measure the accuracy of the model; and the FAR is the proportion of negative cases detected as positive cases in actual negative cases, which can measure the degree of false alarms in the model [24,25].

Table 3. Statistical comparison of water extraction accuracy.

Methods	Recall/%	Precision/%	FAR/%	F1 Measures/%
Otsu	86.71	77.65	22.35	81.93
DG–LRG	83.24	89.87	10.13	86.43
Optimized DG–LRG	84.13	90.76	9.24	87.32

shows the accuracy evaluation statistical results for the Otsu threshold method, DG–LRG method, and optimized DG–LRG method.

It can be seen that the Otsu threshold method had the highest recall rate and could identify a large number of water bodies. However, due to the influence of noise and mountain shadows in the study area, its precision rate was only 77.65%. Although the recall rate of the DG–LRG method proposed in this article was slightly lower than that of the Otsu threshold method, its precision rate was significantly improved, reaching 89.87%. Comparing the F1 index calculated using the comprehensive recall rate and precision rate, the DG–LRG method showed an improvement of 4.5% when compared to the Otsu threshold method. After morphological and external DEM optimization, the F1 index of DG–LRG further improved.

The trade-off between precision and recall in DG–LRG versus Otsu can be attributed to fundamental differences in their operational principles.

The reasons for the Recall reduction in DG–LGR may be that region growth algorithms rely on seed points to initiate segmentation. If seeds are sparsely distributed in complex water bodies (e.g., fragmented lakes or narrow rivers), partial water areas may be omitted, reducing recall, while Otsu’s global threshold captures all potential water pixels (including noise). Meanwhile, the reasons for the Precision improvement in DG–LGR may be that the integration of GLCM texture features (homogeneity, contrast) and dual-polarization (HH/HV) information enables DG-LGR to effectively suppress false alarms from spectrally similar non-water features (e.g., smooth roads, mountain shadows). Closing operations fill voids within water bodies, while DEM-based shadow removal eliminates misclassified pixels, further enhancing precision. Optimized DG–LRG’s precision (90.76%) ensures reliable identification of true flood extents, even at the cost of minor under-detection. In post-disaster rescue, accurate inundation area maps have more decision-making value than overly detected “noise maps”.

4.3. Examples of Flood Emergency Monitoring

4.3.1. Flood Emergency Monitoring in Fangshan District

The optimized DG–LRG method was used to monitor the flood inundation range in the Fangshan area, and the monitoring results are shown in Figure 8. In the figure, the blue area represents the pre-disaster water bodies extracted from optical images, while the red area represents the water body expansion results monitored using GF-3 data on 1 August 2023. Compared with the pre-disaster water bodies, it was found that, by 1 August, having been affected by the rainstorm, the water body area in Fangshan District had expanded by 44.15 km² and the flows of Dashi River (Figure 8A), Xiaoqing River (Figure 8B), and Yongding River (Figure 8C) had soared. The villages around Dashi River in Doudian Town, Shilou Town, and Liulihe Town in Fangshan District were seriously affected, with the water body area expansion rate reaching 3.32%. This example demonstrates that the DG–LRG method can quickly extract the range of water bodies while providing more objective and accurate results, which is helpful for the timely and accurate assessment of disaster situations. Thus, it has good applicability in the context of large-scale flood emergency disaster monitoring.

4.3.2. Monitoring of Flooded Transmission Towers in Fangshan District

Based on the flood inundation area extracted above, monitoring was conducted on the submerged towers in Fangshan District. Affected by the spread of floods and the surrounding environment of the towers, a total of 88 affected towers in Fangshan District were determined based on a 10 m buffer zone in the flooded area. According to on-site survey statistics by inspectors, 83 towers were significantly affected by floods. Accordingly, the proposed method obtained a Precision rate of 94.32% and a FAR of 5.68%, proving that the method proposed in this paper can be effectively applied for the emergency monitoring of towers affected by flood disasters. Based on factors such as the distance between the tower and the flooded area, combined with on-site survey results, the risk level of the affected towers was divided. A total of 0 low-risk towers, 4 medium-risk towers, and 79 high-risk towers were determined in Fangshan District. Figure 9 shows the distribution map of the affected towers in Fangshan District.

5. Conclusions

This article proposed an optimized local region growth algorithm that takes into account the polarization mode and texture features of SAR images for rapid water body extraction. An experiment was conducted with Fangshan District as the research object, and the main conclusions are as follows.

Compared with traditional threshold segmentation methods, the DG–LRG method proposed in this study can obtain more effective water body information, with a recall rate of 83.24% and a precision rate of 89.87% for water bodies in the study area. In terms of the F1 Measure, the DG–LRG method yielded a 4.5% improvement over the Otsu threshold segmentation method.

Optimizing the DG–LRG results based on morphological and external DEM methods allowed for more accurate water body information to be effectively obtained. These advancements highlight their practical value for rapid flood emergency monitoring, as evidenced by the successful identification of 88 transmission towers at risk with 94.32% precision. However, the algorithm’s performance remains constrained by its dependency on seed point selection and the trade-off between computational efficiency and local optimization granularity. For instance, fragmented water bodies in narrow river channels may be partially omitted due to sparse seed distributions. Future research should explore hybrid strategies combining DG–LRG with edge-aware deep learning models to recover omitted water pixels while maintaining high precision.

From the results of emergency flood disaster monitoring examples, the proposed method was shown to be capable of rapidly and accurately extracting the water body range in the experimental area and, therefore, has good applicability and timeliness in the context of large-scale flood emergency monitoring. The proposed method can help to obtain detailed information for the affected area in a timely and accurate manner, providing reference for emergency rescue and disaster relief services, as well as providing support for flood control dispatch and transmission line decision making. Future enhancements, such as optimizing the PSO convergence process and validating the framework with multi-sensor SAR data, are expected to improve its robustness across diverse geographic and climatic conditions. These advancements will further bridge the gap between methodological innovation and operational demands in global flood disaster management.