Digital Elevation Model-Driven River Channel Boundary Monitoring Using the Natural Breaks (Jenks) Method

Abstract

River channels are fundamental geomorphological and hydrological features that play a critical role in regulating the Earth’s water cycle and ecosystems and influencing human activities. This study utilized Digital Elevation Model (DEM) data and multi-source remote sensing imagery (including GF-1 WFV, Sentinel-1, and Sentinel-2) to determine river channel dimensions. River water masks were obtained from multiple remote sensing imagery sources and processed through triangulation and segmentation to generate river reach results. Based on these segmented river reaches, buffer analysis was conducted. The buffer analysis results were then used to refine and clip the 5 m DEM and 12.5 m DEM datasets. Finally, river channels were extracted from the clipped DEM data using the natural breaks classification method. The classification accuracy was assessed using a confusion matrix. Experimental results demonstrate a high overall classification accuracy, reaching or exceeding 0.985, with classification consistency (Kappa coefficient) ranging from 0.78 to 0.81. The 5 m resolution DEM exhibited superior performance compared to the 12.5 m resolution DEM in river channel extraction, especially regarding the classification consistency (Kappa coefficient), with the 5 m resolution model outperforming the latter. This approach effectively delineates the river channel boundaries, transcends the constraints of a singular data source, enhances the precision and resilience of river extraction, and possesses several practical applications. The extracted data can support analyses of river evolution, facilitate hydrological modeling at the basin scale, improve flood disaster monitoring, and contribute to various other research domains.

1. Introduction

In physical geography and hydrology, a channel is a terrain in which a relatively narrow body of water, such as a river, a river delta, or a channel, forms the physical boundary of a river (canal system) with a bed and bank. River channels are major geomorphic and hydrological features that are essential to the Earth’s water cycle and have a significant impact on human society and ecosystems. River channel boundaries are very significant in river evolution analysis, watershed hydrological modeling (for example, river runoff), and flood disaster monitoring [1,2,3].

With the advancement of satellite remote sensing technology, hydrology and other earth sciences have experienced significant progress, enabling more precise observations, enhanced data acquisition, and improved analytical capabilities [4,5,6]. Remote sensing technology can acquire extensive, multi-temporal, and large-scale fundamental hydrological data (examples include rainfall data, river width data, evapotranspiration data, etc.) of rivers promptly and consistently [7]. These distinctive characteristics enhance the origin of primary data in hydrology [8,9,10], establish a novel technical and scientific framework for river monitoring, and validate the evolution and hydrological conditions of rivers [11,12,13,14]. Hydrological stations utilizing satellite remote sensing can measure long-term alterations in the channel width, depth, and bed elevation, while also revealing the underlying driving processes [15,16,17,18,19]. With the continuous development of remote sensing technology, it is possible to determine the boundaries of river channels by using the characteristics of rivers and the topographical features of river channels.

The extraction of river channel features based on remote sensing imagery has become a crucial research direction in hydrology. Existing methods can be broadly categorized into three main approaches: traditional image-processing algorithms, machine learning and deep learning techniques, and multi-source remote sensing data fusion. Traditional methods primarily rely on edge detection, morphological processing, and region segmentation techniques. Edge detection methods delineate river channel boundaries by detecting gradient changes in the pixel intensity or color within the image. These methods are particularly effective for high-resolution imagery with clearly defined river contours, but they are highly sensitive to noise and susceptible to interference from complex backgrounds, as demonstrated by techniques, such as Canny edge detection [20] and Sobel operators [21]. Morphological methods extract river channel features by applying operations, such as dilation, erosion, and opening-closing transformations, which enhance the binarized imagery. For example, Zhu et al. [22] proposed a river extraction approach that combines grayscale and morphological features, while Yousefi et al. [23] further improved detection accuracy by integrating mathematical morphology, a Bayesian classifier, and a dynamic change filter. However, these methods often encounter limitations in extraction accuracy when applied to complex hydrological networks or images with substantial background interference. Additionally, the choice of parameters, such as the size and shape of structural elements, significantly affects the accuracy of the extraction results.

With the advancement of deep learning techniques, convolutional neural networks (CNNs) and fully convolutional networks (FCNs) have made significant strides in river channel extraction [24], with notable architectures, including the VGG series, ResNet, and DenseNet. For instance, Li et al. [25] proposed a pixel-based CNN method to extract water surface information within regions of interest (ROIs) from Landsat imagery. This approach integrates spatial texture and spectral features, enhancing computational efficiency while reducing the number of trainable parameters. Tang et al. [26] further optimized the U-Net architecture to enhance river channel extraction in cold and arid regions, resulting in better boundary detection accuracy. These deep-learning-based methods enable the automatic extraction of river channels from remote sensing imagery and can be applied to various data types, including high-resolution satellite images and UAV imagery. However, deep learning methods heavily rely on large volumes of high-quality training data and face challenges related to sample annotation and spatiotemporal consistency. Additionally, the limitations of single-source data pose difficulties in extracting river channels in complex scenarios, highlighting the increasing need for multi-source remote sensing data fusion techniques. For example, Hughes et al. [27] combined Landsat 8 imagery with deep learning techniques to achieve high-precision river extraction, effectively mitigating cloud cover issues. Further studies have shown that integrating optical remote sensing with synthetic aperture radar (SAR) data ensures stable river detection under varying weather conditions. Moreover, the use of Google Earth Engine (GEE) for multi-temporal and multi-spectral imagery analysis enhances dynamic river monitoring, supporting large-scale hydrographic mapping.

In addition to optical and radar data, topographic information plays a critical role as an auxiliary data source for river channel extraction. The Digital Elevation Model (DEM), which provides fundamental terrain elevation data, is widely applied in hydrological research for delineating watershed boundaries [28]. The topographic information contained within DEMs allows for the extraction of various geomorphological features, such as the slope, aspect, and flow direction, making it an indispensable tool for characterizing hydrographic networks [29,30,31,32]. Hiatt et al. [33] explored a DEM-based method for extracting the river network topology and applied it to estuaries and braided river systems. Cao et al. [34] utilized GIS and DEM-based drainage and hydrology extraction techniques to enhance the accuracy of drainage delineation. By leveraging DEM data, it is possible to extract not only river systems but also define watershed boundaries. Beyond river network extraction, DEMs are also valuable for delineating watershed boundaries. Colombo et al. [35] employed medium-resolution (250 m) DEM data to map river networks and watershed boundaries across Europe, developing a novel mathematical morphology-based algorithm to achieve landscape stratification of the drainage density. With advancements in high-resolution DEM acquisition, their application in small-scale hydrographic net-work extraction has become increasingly widespread. Dong et al. [36] applied local minimum searching techniques for hydrographic identification and used Bresenham’s line algorithm alongside mathematical morphology operations to delineate river channels. Muthusamy et al. [37] conducted flood modeling for urban rivers using DEMs of varying resolutions. Li et al. [38] successfully extracted small open-surface rivers in the upper Yellow River by integrating Sentinel-2 optical imagery (10 m resolution) with DEM data (90 m resolution). Xue et al. [39] demonstrated that high-resolution DEMs effectively mitigate interference from mountainous shadows, thereby ensuring accurate river width estimations. Collectively, these studies underscore the importance of DEMs, not only for river network extraction but also for refining river channel boundary detection, thus playing a vital role in watershed hydrological modeling and hydrodynamic analysis.

In river channel extraction research, DEM is primarily used to mitigate influencing factors and support hydrographic network delineation, rather than directly extracting river channel boundaries. With advancements in remote sensing satellite technology, high-resolution DEM data, with resolutions as fine as 5 m, are now available from sources, such as GF-7 and ZY-3. This study seeks to integrate multi-source remote sensing data to extract river channel boundaries from DEM data, thereby advancing the application of high-precision DEM in river feature extraction. The main objective is to develop a methodology suitable for large-scale river channel delineation by combining high-resolution DEM data with multi-source remote sensing imagery. By leveraging multi-source remote sensing data for water mask extraction as a basis for river localization within extensive DEM datasets, this study aims to effectively merge water mask information with DEM data. Using the natural breaks algorithm—a data classification method that optimally partitions values into distinct groups by minimizing within-class variance and maximizing between-class variance, thus revealing inherent patterns in continuous data—an accurate and efficient approach to large-scale river boundary extraction is proposed. This approach addresses the limitations of existing methods, which are typically confined to local-scale applications.

2. Data and Method

2.1. Study Area

The Dawen River (Figure 1) (116°E to 118°E, 35.7°N to 36.6°N) is a significant tributary basin in the lower reaches of the Yellow River in northern China. It is located in the central Tai Lai Plain of Shandong Province and flows from east to west through the counties and cities of Jinan, Xintai, Tai’an, Feicheng, Ningyang, Wenshang, and Dongping. The main course of the Dawen River extends approximately 208 km to its confluence with the lake, encompassing a catchment area of 9069 square kilometers. The river exhibits a total elevation drop of 362 m from its source to its terminus [40]. This region is part of the Taishan Regional Mountains, Waters, Forests, Farmlands, Lakes, and Grasslands Ecological Protection and Restoration Project [41]. The findings of this study provide valuable insights for future research on river channel formation and morphological analysis within the Dawen River catchment.

2.2. Datas

The datasets utilized in this study include Gaofen-1 (GF-1, WFV), Sentinel-1, Sentinel-2, and Digital Elevation Model (DEM) data, as detailed in

Table 1. Remote sensing data sources and information.

Datas	Resolution	Sources	Purpose
Low and medium precision DEM	5 m (2021), 12.5 m (2021)	Geospatial Data Cloud	River channel extraction
GF-WFV Multispectral data	16 m (June–September 2022)	China Centre for Resources Satellite Data and Application	River surface extraction
Sentinel-1	10 m (June–September 2022)	GEE
Sentinel-2	10 m (June–September 2022)	GEE

. GF-1, Sentinel-1, and Sentinel-2 data were employed for river water mask extraction, while DEM data were used to delineate river channels. DEM datasets with different spatial resolutions (12.5 m and 5 m) were obtained from the Geospatial Data Cloud (gscloud.cn (accessed on 10 January 2023)). GF-1 WFV optical multispectral images (see

Table 2. GF-1WFV data parameter information.

Parameter	16 m Resolution Multispectral Camera
Spectral range	0.45–0.52
	0.52–0.59
	0.63–0.69
	0.77–0.89
Width	800 km (4 cameras)

for parameters) were acquired from the China Centre for Resource Satellite Data and Application (http://www.cresda.com/CN/ (accessed on 10 January 2023)). Sentinel-1 (

Table 3. Sentinel-1 data parameter information.

Parameter	Basic Setup
Mode of work	Interferometric wide mode (IW)
Mode of polarization	Dual polarization (VV, VH)
Data types	Ground distance multi-view (GRD)
Revisit cycle	12 days
Pixel size	10 × 10
Imaging orbit	Rail up mode

) and Sentinel-2 (

Table 4. Sentinel-2 data parameter information.

Number of Wave Band	Width of Band/nm	Central Wavelength/nm	Spatial Resolution/m
B2	458–523	490	10
B3	543–578	560	10
B4	650–680	665	10
B8	785–900	842	10

) datasets were accessed through the Google Earth Engine platform (https://code.earthengine.google.com/ (accessed on 10 January 2023)). In this study, the Sentinel-1 dataset used is Level-1 Ground Range Detected High-Resolution (GRDH) data acquired in Interferometric Wide Swath (IW) mode for land observations, with a spatial resolution of 10 m × 10 m. The data include two polarization modes: vertical–vertical (VV) and vertical–horizontal (VH). Additionally, four spectral bands were utilized (blue (B2), green (B3), red (B4), and near-infrared (B8)), with a 10-day repeat observation cycle and a consistent observation angle.

2.3. Method

2.3.1. Technical Route

In this study, multi-source remote sensing satellite data were used to extract river mask data, followed by the use of 5 m DEM and 12.5 m DEM data to delineate the river channel boundaries. The overall technical workflow is illustrated in Figure 2. First, the water body index method was applied to extract the river water mask (see Section 2.3.3 for details). To ensure data continuity, the initially extracted water mask was reconstructed to generate a complete river water mask dataset. The reconstructed river water mask data were then processed using Delaunay triangulation and segmented into river reaches. Based on the reach segmentation results, a buffer analysis was conducted to determine the computational extent of the DEM. The DEM raster data were clipped according to the buffer zone. Finally, the natural breaks algorithm was applied to the clipped DEM to delineate the river channel extent. The formulas used in this study, along with their details, are provided in

Table 5. Formula details.

Num	Formula	Notation Description	Description
1	$\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{\left(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}-\mathrm{N}\mathrm{I}\mathrm{R}\right)}{\left(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}+\mathrm{N}\mathrm{I}\mathrm{R}\right)}}$	GREEN represents the reflectance in the green wavelength band, and NIR represents the reflectance in the near-infrared wavelength band.	Normalized Water Index
2	${\mathrm{K}}_{\mathrm{S}\mathrm{D}\mathrm{W}\mathrm{I}}=\mathrm{l}\mathrm{n}(10\ast \mathrm{V}\mathrm{V}\ast \mathrm{V}\mathrm{H})$	VV and VH indicate the dual-polarization data from Sentinel-1.	The SDWI is derived with reference to the water index NDWI and employs band calculations between the dual-polarization Sentinel-1 data to enhance water body features, achieving effective water information extraction.
3	$\mathrm{W}=2\ast {\displaystyle \frac{\mathrm{S}}{\mathrm{L}}}$	w: river reach width; S: river reach area; L: river reach perimeter.	The width of the disconnected river reach.
4	$\mathrm{b}\mathrm{u}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{r} \mathrm{d}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}=\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{m} \mathrm{w}\mathrm{i}\mathrm{d}\mathrm{t}\mathrm{h}\ast \mathrm{c}\mathrm{o}\mathrm{e}\mathrm{f}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t}$	The coefficient is set to a default value of 2.	Buffer distance.
5	$\mathrm{G}\mathrm{V}\mathrm{F}={\displaystyle \frac{\mathrm{S}\mathrm{D}\mathrm{A}\mathrm{M}-\mathrm{S}\mathrm{D}\mathrm{C}\mathrm{M}}{\mathrm{S}\mathrm{D}\mathrm{A}\mathrm{M}}}$	GVF: goodness of variance fit; SDAM: the squared deviation from the array mean; SDCM: the squared deviation from the class mean.	Goodness of variance fit.
6	$\mathrm{S}\mathrm{D}\mathrm{A}\mathrm{M}=\sum {\left({\mathrm{x}}_{\mathrm{i}}-\mathrm{X}\right)}^2$	X = array average; Z = the average of the classes.	The squared deviation from the array mean.
7	$\mathrm{S}\mathrm{D}\mathrm{C}\mathrm{M}=\sum \sum {\left({\mathrm{x}}_{\mathrm{i}}-\mathrm{Z}\right)}^2$	X = array average; Z = the average of the classes.	The squared deviation from the class mean.
8	$\mathrm{O}\mathrm{A}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{X}}_{\mathrm{i}\mathrm{i}}}{\mathrm{M}}}$	M is the total number of sampling points, and Xii is the number of class i ground objects on the diagonal that are verified to be correct.	The percentage of the number of correctly extracted inspection points in each terrain category over the total number of inspection samples.
9	$\mathrm{U}\mathrm{A}={\displaystyle \frac{{\mathrm{X}}_{\mathrm{j}\mathrm{j}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{X}}_{\mathrm{j}\mathrm{i}}}}$	Xjj is the number of features of class j on the diagonal that are verified correctly, and Xji represents the number of features extracted as class j that are actually class i features at the time of verification.	The ratio of the validation points in a certain class that are consistent with the reference data to the total number of validation points in that class.
10	$\mathrm{P}\mathrm{A}={\displaystyle \frac{{\mathrm{X}}_{\mathrm{j}\mathrm{j}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{X}}_{\mathrm{i}\mathrm{j}}}}$	Where Xjj is the number of features of class j on the diagonal that are verified correctly, and Xij represents the number of features extracted as class i that are actually class j at the time of verification.	On the diagonal of the confusion matrix, the ratio of the number of correct ground feature classes extracted to the sum of the columns of that class in the confusion matrix.
11	$\mathrm{C}\mathrm{E}=1-\mathrm{U}\mathrm{A}$	UA: User Accuracy; CE: Commission Error.	Refers to the ratio of the sample points of a certain type of ground object misclassified as other land classes to the total number of sample points of the same ground class during verification.
12	$\mathrm{O}\mathrm{E}=1-\mathrm{P}\mathrm{A}$	PA: Producer Accuracy; OE: Omission Error.	Refers to the ratio of sample points that should belong to a certain class but are not classified as such in the verification to the total number of sample points in that class, and the sum of the mapping accuracy is 1.
13	$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}={\displaystyle \frac{{\mathrm{P}}_{\mathrm{o}}-{\mathrm{P}}_{\mathrm{e}}}{1-{\mathrm{P}}_{\mathrm{e}}}}$	Where Po is the number of all correct classifications divided by the total number of samples, which is OA. Pe is the product of the true number and the total number of classified pixels in the category divided by the total number of samples.	The Kappa coefficient, which is calculated on the basis of the confusion matrix, takes into account the pixels of all classifications and thus enables a more objective evaluation of the classification results.

.

2.3.2. River Water Extraction and Reconstruction Methods

• Normalized water index method

The Normalized Difference Water Index (NDWI) was calculated using GF-1 and Sentinel-2 imagery, followed by threshold segmentation based on the Otsu method (maximum between-class variance algorithm) to extract water mask data. The NDWI is commonly used to enhance the visibility of open water in remote sensing imagery by leveraging near-infrared (NIR) and visible light bands [42]. This approach highlights water bodies while minimizing interference from soil and vegetation. The NDWI formula is as follows:

$$\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{\left(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}-\mathrm{N}\mathrm{I}\mathrm{R}\right)}{\left(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}+\mathrm{N}\mathrm{I}\mathrm{R}\right)}}$$

(1)

where GREEN denotes the reflectance in the green wavelength band, and NIR represents the reflectance in the near-infrared wavelength band.

2.

SDWI

The Sentinel-1 Dual-Polarized Water Index (SDWI) is derived from Sentinel-1 data and utilized for water body extraction. The calculation is based on HV and HH backscattering coefficients, followed by threshold segmentation using the Otsu method (maximum between-class variance algorithm). Originally proposed by Jia Shichao et al. [43], SDWI serves as an effective approach for water body identification. The formula is as follows:

$${\mathrm{K}}_{\mathrm{S}\mathrm{D}\mathrm{W}\mathrm{I}}=\mathrm{l}\mathrm{n}(10\ast \mathrm{V}\mathrm{V}\ast \mathrm{V}\mathrm{H})$$

(2)

where KSDWI represents the calculated index value, with VV and VH denoting the dual-polarization data from Sentinel-1. SDWI is derived with reference to the Normalized Difference Water Index (NDWI) and utilizes band calculations between Sentinel-1′s dual-polarization data to enhance water body features, thereby enabling effective water body extraction.

3.

Water mask reconstruction method

Reconstructing the river water mask at discontinuous water reaches involves drawing line segments at the breakpoints, referencing optical remote sensing images. The spatial relationship between the line shape file and river water mask data is then utilized to remove patches and reconstruct the discontinuities in the river water mask. First, a spatial relationship analysis is performed between the line shape file and the river water mask data to identify intersections. Water mask patches that do not intersect with the line shape file are removed, leaving the river water mask data. For discontinuous water masks, the average width at both ends of the discontinuity is calculated. Once the average width is determined, a buffer analysis is applied to the line shape file. After obtaining the average width, a buffer analysis is performed on the line shape file. The resulting buffer zone is then merged with the water mask data to generate the complete river water mask data. The formula used to calculate the average river width is as follows:

$$\mathrm{W}=2\ast {\displaystyle \frac{\mathrm{S}}{\mathrm{L}}}$$

(3)

where w is the river reach width; S is the river reach area; L is the river reach perimeter.

As shown in Figure 3, the dashed section represents the gap segment of the river, where the combined area of the two end segments (S1 + S2) is equal to S and the combined perimeter of the two end segments (L1 + L2) is equal to L. The results of the implementation are presented in Figure 4.

2.3.3. Delaunay Triangulation and River Reach Segmentation

The river water mask data are triangulated to generate skeletal lines and triangular elements. In cartographic mapping, the river’s water mask is treated as a narrow polygonal patch, with the centerline defined as the path of points equidistant from two or more polygon edges. This centerline represents the set of points within the polygon that minimizes the distance to each edge. The nodes of the skeleton lines for any polygon are located at the midpoints of the common edges between neighboring triangles. These midpoints correspond exactly to those of the adjacent triangles resulting from the constrained Delaunay triangulation of the polygon, as illustrated in Figure 5. In Figure 5b, the triangles represent the triangular elements, while the red lines denote the polygonal skeleton lines. The red points, connected by the skeleton lines, represent the skeleton nodes.

Each triangular element obtained from the Delaunay triangulation is defined as a river unit. However, due to their small individual areas, these triangular elements are insufficient to characterize the properties of extended river sections. To enhance computational efficiency, the river mask data are segmented. The river reach segmentation process considers several key factors, including the maximum river length, maximum segment unit count, and minimum unit count. Before segmentation, the maximum segment length is determined based on river characteristics to establish an upper limit based on the segment size. The maximum segment unit count serves as a constraint to regulate the segment length, preventing excessive subdivision at each node, and is adjusted according to the river’s sinuosity. The minimum unit count is applied to remove irregular, jagged triangular units at the edges, ensuring smoother segmentation.

2.3.4. River Channel Extraction

• DEM clipping

River channel extraction is performed through a combination of river reach segmentation and DEM data integration. A buffer analysis is applied to each river reach, and the corresponding buffer zones are used to clip the DEM data. The clipped DEM data are then processed using the natural break algorithm to delineate the channel boundaries for individual river reaches. Finally, the extracted channel boundaries from all reaches are merged to reconstruct the complete river channel

First, the minimum bounding polygon is computed for each river reach. The minimum bounding polygon is a geometric shape that tightly encloses a given geographic feature. By applying this computation to each river reach, the precise boundary of each reach can be delineated. Based on the geometric characteristics of the minimum bounding polygon, the minimum width of the river reach is then calculated. This width serves as a critical parameter for subsequent computations and analyses, as it represents the narrowest horizontal extent of the river reach.

Using the derived minimum width, the buffer distance is determined. The buffer distance is calculated using the following formula:

$$\mathrm{b}\mathrm{u}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{r} \mathrm{d}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}=\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{m} \mathrm{w}\mathrm{i}\mathrm{d}\mathrm{t}\mathrm{h}\ast \mathrm{c}\mathrm{o}\mathrm{e}\mathrm{f}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t}$$

(4)

In this study, the coefficient is set to a default value of 2. Using the computed buffer distance, a buffer analysis is performed for each river reach. Buffer analysis, a widely used spatial analysis technique in Geographic Information Systems (GISs), expands outward from the river segment by a specified distance (i.e., the buffer distance) to generate a buffer zone. This buffer zone serves as a precise spatial boundary for clipping the Digital Elevation Model (DEM). The resulting buffer zone is then applied to clip the DEM, producing segmented DEMs corresponding to each river reach. These segmented DEMs, which focus on the river reaches and their surrounding topography, provide more refined data for subsequent river channel extraction.

2.

Natural breaks (Jenks)

The river channel exhibits significant topographical variations, with its cross-section typically forming a parabolic or trapezoidal shape. Consequently, river boundaries can be delineated by analyzing elevation differences within the river basin, utilizing topographic data for segmentation and classification. The histogram of Digital Elevation Model (DEM) data often display a multi-peak distribution, where each peak corresponds to a distinct land feature. By applying the natural breaks algorithm, land features at different elevations are classified into distinct groups. Among these groups, the river channel is identified as the one with the lowest elevation. Each feature is represented as a separate region in the DEM histogram, and the number of distinct features depends on the complexity of the terrain captured in the DEM.

The natural breaks method of classification is a data-clustering technique proposed by George Frederick Jenks [44]. The natural breaks algorithm is crucial for identifying statistically significant turning points and breakpoints within a data series. These breakpoints are inherently determined rather than arbitrarily chosen, enabling the classification of study objects into groups with similar characteristics. This approach facilitates the establishment of effective and meaningful classification boundaries [45]. Additionally, remote sensing data often exhibit natural inflection points and fractures that can be effectively categorized using the natural breaks classification method. While its application in geographic information research remains somewhat limited, this approach is widely used in geographic information science, particularly for generating thematic maps in ArcGIS10.6 software. Chen et al. [46] conducted an experimental study on the classification of geographic environmental units in China, integrating the natural breaks method with other techniques. Their findings demonstrate that the natural breaks method offers strong adaptability and high accuracy in delineating geographic environmental units.

The natural breaks classification method is a data classification approach widely used in cartography and geographic data analysis. It categorizes data based on inherent natural groupings by identifying divisions where values naturally cluster. This method is particularly effective for classifying continuous data, such as the elevation, income distribution, or spectral values in remote sensing. Its primary objective is to minimize within-class variance while maximizing between-class variance, ensuring that each class remains as internally homogeneous as possible while being distinctly separated from others. This optimization helps determine the most meaningful categorization of values across different groups [44]. The goodness of variance fit (GVF) is calculated by dividing the difference between the squared deviation from the array mean (SDAM) and the squared deviation from the class mean (SDCM) by SDAM:

$$\mathrm{G}\mathrm{V}\mathrm{F}={\displaystyle \frac{\mathrm{S}\mathrm{D}\mathrm{A}\mathrm{M}-\mathrm{S}\mathrm{D}\mathrm{C}\mathrm{M}}{\mathrm{S}\mathrm{D}\mathrm{A}\mathrm{M}}}$$

(5)

$$\mathrm{S}\mathrm{D}\mathrm{A}\mathrm{M}=\sum {\left({\mathrm{x}}_{\mathrm{i}}-\mathrm{X}\right)}^2$$

(6)

$$\mathrm{S}\mathrm{D}\mathrm{C}\mathrm{M}=\sum \sum {\left({\mathrm{x}}_{\mathrm{i}}-\mathrm{Z}\right)}^2$$

(7)

X = array average; Z = the average of the classes.

The closer the goodness of variance fit (GVF) is to 1, the better the performance. A GVF of 1 can only be achieved if the within-class variance is zero.

For each segmented DEM, the natural breaks (Jenks) algorithm is applied to classify data based on their intrinsic distribution. This method automatically identifies natural groupings within the dataset, allowing for the precise extraction of river channels. By accounting for terrain continuity and variation, the natural breaks algorithm effectively differentiates the river channel from the surrounding topography, thereby enhancing the extraction accuracy. Finally, the extracted river channels from individual segments are mosaicked together. Mosaicking integrates multiple localized geographic datasets into a continuous whole, ensuring seamless connectivity and forming a coherent representation of the entire river channel.

2.3.5. Accuracy Verification

In the Dawen River Basin, validation points were uniformly sampled and established (Figure 6) and subsequently classified as either river or non-river points. Visual interpretation, based on Google Maps and 2 m high-resolution panchromatic satellite imagery, was conducted to determine whether each validation point fell within the river area. A total of 39,496 validation points were distributed across the basin, with 1498 identified as river points and 37,998 as non-river points. A confusion matrix analysis was performed to evaluate the classification results, calculating the overall accuracy, user accuracy, producer accuracy, and the Kappa coefficient. This analysis facilitated the creation of an accuracy assessment table for river extraction results in the Dawen River plain, incorporating various DEM precision levels in combination with multi-source remote sensing imagery.

• Confusion matrix

The confusion matrix, also known as the error matrix, is a standard framework for evaluating the accuracy of remote sensing image classification. It presents classification results in an n × n matrix format, where rows and columns represent different classes.

Based on the confusion matrix, key accuracy metrics—including the overall accuracy, user accuracy, misjudgment error, producer accuracy, and omission error—can be computed to assess classification performance. The calculation methods for these indices is as follows:

Overall Accuracy (OA): the proportion of correctly classified inspection points across all terrain categories relative to the total number of inspection samples.

The formula is as follows:

$$\mathrm{O}\mathrm{A}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{X}}_{\mathrm{i}\mathrm{i}}}{\mathrm{M}}}$$

(8)

where M is the total number of sampling points and X_ii is the number of class i ground objects on the diagonal that are verified to be correct.

User Accuracy (UA): the proportion of validation points in a given class that match the reference data, relative to the total number of validation points in that class.

The formula is as follows:

$$\mathrm{U}\mathrm{A}={\displaystyle \frac{{\mathrm{X}}_{\mathrm{j}\mathrm{j}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{X}}_{\mathrm{j}\mathrm{i}}}}$$

(9)

where X_jj is the number of features of class j on the diagonal that are verified correctly, and X_ji represents the number of features extracted as class j that are actually class i features at the time of verification.

Producer Accuracy (PA): the ratio of correctly classified ground feature instances (diagonal elements of the confusion matrix) to the total number of instances in the corresponding column of that class. The formula is as follows:

$$\mathrm{P}\mathrm{A}={\displaystyle \frac{{\mathrm{X}}_{\mathrm{j}\mathrm{j}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{X}}_{\mathrm{i}\mathrm{j}}}}$$

(10)

where X_jj is the number of features of class j on the diagonal that are verified correctly, and X_ij represents the number of features extracted as class i that are actually class j at the time of verification.

Commission Error (CE): refers to the ratio of the sample points of a certain type of ground object misclassified as other land classes to the total number of sample points of the same ground class during verification, and the sum of CE and user accuracy is 1. The formula is as follows:

$$\mathrm{C}\mathrm{E}=1-\mathrm{U}\mathrm{A}$$

(11)

Omission Error (OE): refers to the ratio of sample points that should belong to a certain class but are not classified as such in the verification to the total number of sample points in that class, and the sum of the mapping accuracy is 1. The formula is as follows:

$$\mathrm{O}\mathrm{E}=1-\mathrm{P}\mathrm{A}$$

(12)

2.

Kappa coefficient

The Kappa coefficient is a statistical measure used for consistency testing and serves as an indicator of the classification accuracy. Calculated from the confusion matrix, it accounts for all classified pixels, providing a more objective evaluation of classification performance. The formula for calculating the Kappa coefficient is as follows:

$$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}={\displaystyle \frac{{\mathrm{P}}_{\mathrm{o}}-{\mathrm{P}}_{\mathrm{e}}}{1-{\mathrm{P}}_{\mathrm{e}}}}$$

(13)

where P_o is the proportion of correctly classified samples (equivalent to OA), and Pe is the expected agreement by chance, calculated as the product of actual and classified sample counts divided by the total samples. The accuracy criteria for the Kappa coefficient evaluation are shown in

Table 6. The Kappa coefficient accuracy assessment standard.

Kappa	Accuracy
0–0.2	Very Poor
0.21–0.4	Fair
0.41–0.60	Good
0.61–0.80	Very Good
0.81–1	Excellent

.

3. Results

3.1. Water Mask Extraction and Reconstruction Results

The river water mask was extracted using the water body index method. Figure 7 and Figure 8 illustrate the water mask extraction and reconstruction results based on GF-1 imagery (additional results are provided in the Supplementary Materials: Figures S1–S5). The initial hydrological data exhibited fragmentation and discontinuities in river networks. After reconstruction, most fragmented segments were eliminated, and the continuity of river water masks was effectively restored. However, due to the resolution of the remote sensing imagery, some smaller rivers were omitted.

3.2. Results of River Channel Extraction

Figure 9 illustrates the triangular tessellation of the water mask along with the segmentation results, including triangular elements, the river skeleton, and river reach segmentation. Delaunay triangulation was applied to generate computational units for a subsequent analysis. However, this process divided the river reaches into numerous small triangles, leading to excessive segmentation and increased computational complexity. To address this, an additional segmentation step was introduced to balance the computational efficiency and segmentation scale. The optimized river reaches exhibit more uniform lengths and shapes, effectively reducing over-segmentation and preserving river geometry. As a result, the efficiency of subsequent river extraction computations is improved. However, in large water bodies, such as upstream reservoirs, segmentation remains relatively fragmented. To avoid negatively impacting downstream river segmentation, no further adjustments were made. Future research will focus on refining the river reach segmentation method to enhance its effectiveness.

The extraction results indicate that the 5 m DEM outperforms the 12.5 m DEM, confirming that higher DEM resolution enhances the accuracy of river channel delineation. As shown in Figure 10b (additional extraction results are provided in the Supplementary Materials: Figures S6–S8), the 12.5 m DEM produces irregular and coarse boundaries, failing to accurately represent the actual river channel. In contrast, the 5 m DEM results exhibit misclassification, where areas beyond the river channel are erroneously identified as part of it. This misclassification is primarily attributed to the flat and low-lying topography of the plain, where minimal elevation variations make accurate differentiation challenging. While the primary river channel can be readily extracted by identifying natural discontinuities in the DEM, accurately delineating the riverbed’s beach area remains difficult. Additionally, low-lying regions adjacent to the river are often misclassified as part of the channel due to their similar elevation characteristics.

3.3. Accuracy Analysis Results

River water masks derived from multi-source remote sensing images during the flood season were utilized for water mask reconstruction. River channel extraction was performed using a combination of 5 m and 12.5 m DEMs, while 2 m panchromatic high-definition images were incorporated for a visual analysis. A confusion matrix analysis was then conducted to assess the accuracy of the extraction results. The accuracy results radar chart is shown in Figure 11. From the radar chart, it can be preliminarily observed that all OA values exceed 0.9. The highest UA value is achieved by Sentinel-1 5 m DEM, while the highest PA and Kappa values are both observed for Sentinel-1 12.5 m DEM.

The detailed accuracy results are presented in

Table 7. Accuracy evaluation results of GF-1 channel extraction results of the Dawen river.

	Class	Others	River Channel	Sum	UA	CE
GF1-5 m DEM	Others	37,893	454	38,347	0.988	0.012
	River channel	105	1044	1149	0.909	0.091
	Sum	37,998	1498	39,496
	PA	0.997	0.697
	OE	0.003	0.303
	OA	0.986
	Kappa	0.782
GF1-12.5 m DEM	Others	37,911	462	38,373	0.988	0.012
	River channel	87	1036	1123	0.923	0.078
	Sum	37,998	1498	39,496
	PA	0.998	0.692
	OE	0.002	0.308
	OA	0.986
	Kappa	0.784

,

Table 8. Accuracy evaluation results of Sentinel-1 channel extraction results of the Dawen river.

	Class	Others	River Channel	Sum	UA	CE
Sentinel1-5 m DEM	Others	37,949	462	38,411	0.988	0.012
	River channel	49	1036	1085	0.955	0.045
	Sum	37,998	1498	39,496
	PA	0.999	0.692
	OE	0.001	0.308
	OA	0.987
	Kappa	0.796
Sentinel1-12.5 m DEM	Others	37,906	394	38,300	0.990	0.010
	River channel	92	1104	1196	0.923	0.077
	Sum	37,998	1498	39,496
	PA	0.998	0.737
	OE	0.002	0.263
	OA	0.988
	Kappa	0.813

,

Table 9. Accuracy evaluation results of Sentinel-2 channel extraction results of the Dawen river.

	Class	Others	River Channel	Sum	UA	CE
Sentinel2-5 m DEM	Others	37,858	407	38,265	0.989	0.011
	River channel	140	1091	1231	0.886	0.114
	Sum	37,998	1498	39,496
	PA	0.996	0.728
	OE	0.004	0.272
	OA	0.986
	Kappa	0.793
Sentinel2-12.5 m DEM	Others	37,889	459	38,348	0.988	0.012
	River channel	109	1039	1148	0.905	0.095
	Sum	37,998	1498	39,496
	PA	0.997	0.694
	OE	0.003	0.306
	OA	0.986
	Kappa	0.778

and

Table 10. Accuracy evaluation of water surface channel extraction results of the largest river in the Dawen River.

	Class	Others	River Channel	Sum	UA	CE
Maximum water surface-5 m DEM	Others	37,917	411	38,328	0.989	0.011
	River channel	81	1087	1168	0.931	0.069
	Sum	37,998	1498	39,496
	PA	0.998	0.726
	OE	0.002	0.274
	OA	0.988
	Kappa	0.8091
Maximum water surface-12.5 m DEM	Others	37,853	430	38,283	0.989	0.011
	River channel	145	1068	1213	0.881	0.120
	Sum	37,998	1498	39,496
	PA	0.996	0.713
	OE	0.004	0.287
	OA	0.985
	Kappa	0.780

. The extraction of primary river channels in the Dawen River Basin achieves an overall accuracy exceeding 0.98, with a Kappa coefficient above 0.75, indicating substantial agreement based on the Kappa evaluation criteria (0.61–0.8). The commission error for river channel classification is below 0.15, while the omission error remains under 0.31. These results validate the effectiveness of the proposed method, demonstrating that the DEM-based natural discontinuity point approach is both reliable and efficient for river channel extraction.

The results of river channel extraction based on GF-1 (Table 7) indicate high classification accuracy. The producer accuracy (~0.997) and user accuracy (~0.998) for non-channel areas in both DEMs are exceptionally high, with minimal misclassification error. For river classification, the user accuracy is 0.909 for GF1-5 m DEM and 0.923 for GF1-12.5 m DEM, showing only a slight difference. The overall accuracy of GF1-5 m DEM and GF1-12.5 m DEM is both 0.986, demonstrating negligible variation. The Kappa coefficients for the two DEMs are 0.785 and 0.784, respectively, indicating nearly identical classification agreement.

The Sentinel-1 river channel extraction results (Table 8) show that both DEM resolutions achieve exceptionally high overall accuracy, approximately 0.990, indicating strong performance in overall accuracy. However, river region classification remains suboptimal, particularly in terms of the producer accuracy, which is 0.692 for the 5 m DEM and 0.737 for the 12.5 m DEM. The Kappa coefficient further suggests that the 12.5 m DEM exhibits slightly better classification agreement than the 5 m DEM.

The Sentinel-2 extraction results (Table 9) indicate varying classification accuracy for channel and non-channel areas. The classification of river channels performed suboptimally, particularly in terms of the producer accuracy, which was 0.728 for the 5 m resolution model and 0.694 for the 12.5 m resolution model. The Kappa coefficient suggests that the 5 m resolution model exhibits slightly better classification agreement than the 12.5 m resolution model. The increased misclassification rate indicates suboptimal extraction performance, likely due to the limited effectiveness of Sentinel-2 in river water mask extraction, which subsequently affects the accuracy of river channel delineation.

Table 10 presents the river channel extraction results based on the maximum water mask data. The user accuracy of the 5 m DEM extraction is 0.931, indicating the effective identification of channel areas. However, the producer accuracy is 0.726, suggesting that a significant portion of channel areas are misclassified as non-channels. The omission error of 0.274 further highlights the substantial misclassification of channel areas. Despite this, the overall accuracy reaches 0.988, demonstrating strong classification performance. The Kappa coefficient of 0.8091 indicates a high level of consistency in classification, though some misclassifications persist. For the 12.5 m DEM, the user accuracy is 0.881, reflecting the effective recognition of river channel areas. The overall accuracy is 0.985, indicating a strong classification performance. The Kappa coefficient is 0.780, signifying a high level of classification consistency, albeit slightly lower than that of the 5 m resolution model. Both DEM resolutions achieve exceptionally high overall accuracy, approaching 0.990, demonstrating their strong performance in overall classification. The Kappa coefficient further suggests that the 5 m resolution model exhibits slightly better classification consistency than the 12.5 m resolution model.

The overall accuracy of all results is high, approaching or exceeding 0.985, indicating that the classification outcomes are highly commendable. Kappa coefficients typically range from 0.780 to 0.810, reflecting substantial concordance in the categorization results, although some misclassifications are present. The Kappa coefficient for the 5 m resolution DEM is typically slightly higher than that for the 12.5 m resolution DEM, indicating that the 5 m resolution DEM provides improved classification agreement.

4. Discussion

Line charts depicting the producer accuracy, user accuracy, overall accuracy, Kappa coefficient, commission error, and omission error for DEM data at various resolutions are presented from the perspective of different sensors (Figure 12). The 5 m resolution DEM consistently outperforms the 12.5 m resolution DEM in both non-channel and channel classifications, particularly in the classification agreement, as indicated by the Kappa coefficient. Results from integrating various remote sensing image extractions (Sentinel-1, Sentinel-2, maximum river surface, GF-1) show high classification efficacy in non-channel areas but comparatively lower performance in channel areas, especially in producer accuracy. Despite generally favorable categorization outcomes, there remains room for improvement in the producer accuracy and the reduction of omission errors in the river region. The 5 m resolution DEM typically shows superior extraction efficacy and consistency compared to the 12.5 m resolution DEM; however, its higher resolution captures more ground object details, leading to extraction results that may include features not associated with the river area. The quality of the water mask data influences the extent of the algorithm’s analytical range. With a consistent analysis range coefficient, higher water mask quality is linked to an expanded analysis range and improved accuracy in river channel extraction.

In terms of the extraction performance, river channel delineation in the main stream region yields satisfactory results. However, in tributary areas, the resolution of the remote sensing imagery used for river water mask extraction impacts accuracy, resulting in the failure to capture small tributaries. This limitation leads to the omission of narrow river channels in these regions. Furthermore, the accuracy of water mask extraction directly influences the final delineation of river boundaries. While the natural break algorithm employed in this study mitigates spectral confusion (e.g., same object, different spectra; different objects, same spectra), spectral inconsistencies still affect the accuracy of water mask extraction. To address these challenges, future studies should consider a hierarchical river extraction approach, utilizing different extraction methods for various regions to achieve comprehensive river delineation across the entire basin.

In this study, water mask data reconstruction was performed using manually delineated line shape file data. This approach relies on high-resolution imagery and is effective in regions with minimal water mask discontinuities. However, when high-resolution imagery is unavailable or water mask discontinuities are prevalent, the manual delineation method becomes cumbersome and inefficient. To address this limitation, future research could incorporate watershed river network data to reconstruct interrupted river reaches. Various river network extraction methods have been developed, and future studies may explore the feasibility of using river networks, rather than multiple line segments, to improve both the extraction efficiency and accuracy.

This study successfully delineated river boundaries and integrated water body data, enabling the extraction of river channels across large-scale DEM regions. This advancement enhances the spatial coverage of river extraction; however, several challenges remain. Extraction methods need to be adapted to different river types. For example, the Dawen River, with its relatively simple topography and channel morphology, including both straight and meandering sections, presents fewer challenges for extraction. In contrast, braided and anastomosing rivers, with their complex morphologies, pose significant difficulties for extraction. Therefore, more comprehensive extraction methods that consider additional influencing factors are necessary. Future research should focus on evaluating the applicability of various extraction approaches for different river types and further optimizing methods tailored to specific river morphologies.

Additionally, although this study utilized DEM data for river channel boundary extraction, the accuracy of the DEM data directly influences the reliability of the extracted results. Enhancing the precision of DEMs derived from remote sensing imagery is a crucial research direction for improving river extraction accuracy. Future efforts should focus on advancing high-precision DEM generation techniques to enhance the overall reliability of river channel boundary extraction.

In summary, this research offers a more refined approach to river channel delineation compared to previous studies by focusing on the topographic perspective rather than solely relying on water masks to define river boundaries. The data used in this study are more comprehensive, enabling a detailed representation of the river morphology and providing more accurate river boundary data for studies on river evolution. While river boundary delineation is fundamental for river analysis, investigating the evolution of river physical boundaries using this foundational dataset holds greater significance. Future research should expand beyond river extraction to explore river evolution, hydrological connectivity, and human–environment interactions, fostering a deeper understanding of river system dynamics.

5. Conclusions

This study utilizes the natural breaks classification method, along with GF-1 multispectral data, Sentinel-1 radar data, and Sentinel-2 multispectral data, to extract river channel boundaries from DEM data. Water mask data for the flood season (June-September) in the study area were retrieved and reconstructed using GF-1, Sentinel-1, and Sentinel-2 data. River channel boundaries were delineated by combining 5 m and 12.5 m DEM data, and the results were subsequently assessed and analyzed. The overall accuracy across all resolutions and sensors is high, nearing or exceeding 0.985, indicating strong classification performance. The categorization outcomes for river areas show producer accuracies generally ranging from 0.65 to 0.75, suggesting room for improvement in the precision of river boundary extraction. The 5 m DEM consistently outperformed the 12.5 m DEM, particularly in classification consistency (Kappa coefficient), indicating that higher-resolution DEM data are beneficial for river channel extraction. The study also highlights that the quality of water mask data extraction directly impacts the accuracy of river channel boundary results, with superior water mask extraction correlating with better river channel extraction quality. This underscores a potential area for methodological improvement.

To address the suboptimal classification performance in river channel regions, improvements can be achieved by optimizing classification algorithms and integrating multi-source data fusion to enhance accuracy. Future studies could focus on optimizing resolution selection to balance the classification accuracy with computational efficiency. Data from multiple sensors provide complementary information, and their integration can improve both the classification accuracy and reliability. Advanced multi-source data fusion techniques should be explored to achieve optimal categorization results. Additionally, leveraging cutting-edge deep learning algorithms, along with reinforcement and transfer learning methods, could significantly enhance DEM data classification, particularly in complex terrains and dynamic conditions. The application of these findings in real-world geographic information systems (GISs), environmental monitoring, and disaster alert systems will validate their effectiveness, offering opportunities for future optimization and refinement. The advancements and extensions of this research will provide reliable technical and data support for river evolution analysis, watershed hydrological studies, and related research and applications.