Nothing Special   »   [go: up one dir, main page]
Next Article in Journal
Nonlinear Seismic Signal Denoising Using Template Matching with Time Difference Detection Method
Previous Article in Journal
Deep Learning-Based Feature Matching Algorithm for Multi-Beam and Side-Scan Images
Previous Article in Special Issue
Distribution and Structure of China–ASEAN’s Intertidal Ecosystems: Insights from High-Precision, Satellite-Based Mapping
You seem to have javascript disabled. Please note that many of the page functionalities won't work as expected without javascript enabled.
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 4
10.3390/rs17040676
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Deep Learning Extraction of Tidal Creeks in the Yellow River Delta Using GF-2 Imagery

by
Bojie Chen
1,
Qianran Zhang
1,*,
Na Yang
1,
Xiukun Wang
1,
Xiaobo Zhang
1,
Yilan Chen
2 and
Shengli Wang
1
1
College of Ocean Science and Engineering, Shandong University of Science and Technology, Qingdao 266590, China
2
First Institute of Oceanography, Ministry of Natural Resources (MNR), Qingdao 266061, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(4), 676; https://doi.org/10.3390/rs17040676
Submission received: 17 December 2024 / Revised: 12 February 2025 / Accepted: 14 February 2025 / Published: 16 February 2025
(This article belongs to the Special Issue Remote Sensing of Coastal, Wetland, and Intertidal Zones)
Browse Figures
Figure 1
<p>Location of the study area: GF-2 images of the Yellow River Delta (RGB: 3, 2, 1 bands) (<b>a</b>), (<b>b</b>) mudflat creek area, and (<b>c</b>) salt marsh creek area.</p> "> Figure 2
<p>The ODU-Net model structure.</p> "> Figure 3
<p>CA-ASPP module structure.</p> "> Figure 4
<p>Coordinate attention module structure.</p> "> Figure 5
<p>Comparison of results of ablation experiments on the mudflat creek test set.</p> "> Figure 6
<p>Comparison of results of ablation experiments on the salt marsh creek test set.</p> "> Figure 7
<p>(<b>a</b>) Larger spatial mudflat area; (<b>b</b>) prediction results for mudflat creeks; (<b>c</b>) larger spatial mudflat area; (<b>d</b>) prediction results for salt marsh creeks.</p> "> Figure 8
<p>Comparison of the edge detection results (1–3 for mudflat regions, 4–6 for salt marsh regions).</p> "> Figure 9
<p>Semantic segmentation results of different models on the mudflat creek test set.</p> "> Figure 10
<p>Semantic segmentation results of different models on the salt marsh creek test set.</p> ">
Versions Notes

Abstract

:
Tidal creeks are vital geomorphological features of tidal flats, and their spatial and temporal variations contribute significantly to the preservation of ecological diversity and the spatial evolution of coastal wetlands. Traditional methods, such as manual annotation and machine learning, remain common for tidal creek extraction, but they are slow and inefficient. With increasing data volumes, accurately analyzing tidal creeks over large spatial and temporal scales has become a significant challenge. This study proposes a residual U-Net model that utilizes full-dimensional dynamic convolution to segment tidal creeks in the Yellow River Delta, employing Gaofen-2 satellite images with a resolution of 4 m. The model replaces the traditional convolutions in the residual blocks of the encoder with Omni-dimensional Dynamic Convolution (ODConv), mitigating the loss of fine details and improving segmentation for small targets. Adding coordinate attention (CA) to the Atrous Spatial Pyramid Pooling (ASPP) module improves target classification and localization in remote sensing images. Including dice coefficients in the focal loss function improves the model’s gradient and tackles class imbalance within the dataset. Furthermore, the inclusion of dice coefficients in the focal loss function improves the gradient of the model and tackles the dataset’s class inequality. The study results indicate that the model attains an F1 score and kappa coefficient exceeding 80% for both mud and salt marsh regions. Comparisons with several semantic segmentation models on the mud marsh tidal creek dataset show that ODU-Net significantly enhances tidal creek segmentation, resolves class imbalance issues, and delivers superior extraction accuracy and stability.

1. Introduction

Coastal tidal flats serve as a transition zone between terrestrial and marine environments, playing a crucial role in the exchange of materials between land and sea [1,2]. Tidal creeks, among the most dynamic microgeomorphic features of tidal flats, are typically arranged in a network with varying widths and depths. Their formation is primarily driven by the combined effects of tidal action and the processes of erosion and deposition caused by seawater [3]. The tidal movement of seawater erodes the surface, gradually carving gullies as the tide rises and falls. These creeks are usually found along coastlines with gentle slopes, significant tidal activity, and minimal wave influence. From an ecological perspective, tidal creeks are vital for the exchange of nutrients and organic substances within mangrove forests. They enable mangrove ecosystems to perform essential functions such as purifying seawater, supporting biodiversity, sequestering and storing carbon, and providing other important ecological services [4,5,6]. The lateral exchange and cyclic transformation of nutrients in mangrove–estuarine systems are essential for regulating blue carbon functions and maintaining marine ecosystem health. Furthermore, tidal creeks are important for studying fish community composition [7,8], macrobenthic communities [9], and watershed development as well as mangrove management [10,11]. As the only excellently preserved and youngest wetland ecosystem in the warm-temperature area of the world, the Yellow River Delta serves as a model for the harmonious integration of nature and humanity. Additionally, it is a crucial area for ecological conservation and economic development. Its health and stability depend on the critical role of tidal creeks in maintaining hydrological connectivity [12]. Substantial research has been devoted to understanding the hydrodynamics of tidal creeks and the changes in sediment [13,14,15]. To better assess the impact of tidal creeks on coastal ecosystems, it is essential to analyze their changes with high spatial and temporal resolution [16]. Given that tidal creeks are significantly influenced by tidal water and undergo rapid changes in space and time, efficiently and accurately extracting their spatial distribution is essential for the conservation and effective management of coastal wetlands [17].
From a geometric perspective, tidal creeks exhibit a meandering linear distribution, with their width varying significantly along their course, ranging from several kilometers to just a few centimeters [18]. This variation in width substantially affects the accuracy of creek detection. Additionally, tidal creeks traverse areas consisting of diverse wetland types, with varying water content, pronounced background heterogeneity, differing sand compositions, and a range of vegetation cover [19]. These factors pose considerable challenges for the accurate segmentation of tidal creeks. Given the dynamic morphology of tidal creeks and their critical role in coastal engineering and mudflat development, there is an urgent need for the development of efficient, high-precision methods for tidal creek identification to enable long-term monitoring of tidal flats and their evolution. Currently, the standard method for high-precision tidal creek identification relies on manual labeling, which is both experience-dependent and relatively inefficient. As the volume of satellite remote sensing data increases, the current level of efficiency is insufficient to meet the demands of monitoring the spatial and temporal dynamics of tidal flats [20]. Traditional modeling methods and machine learning approaches have emerged to address these challenges. In recent years, significant research has focused on tidal creek extraction. The flow path simulation method is a semi-automated technique for tidal creek extraction, which combines adaptive height thresholding and weighted distance transformations with high-precision LiDAR data [21]. However, it is sensitive to faults and cracks and is limited to capturing tidal creeks as single-pixel widths, disregarding width information. The heuristic approach, incorporating the modified fuzzy C-means (MFCM) algorithm [22], effectively addresses the heterogeneity of the intertidal background where tidal creeks are located. The method employs a Gaussian filter to extract fine tidal creeks, analyzes spectral properties using water body indices to enhance information about water bodies, and subsequently segments wider tidal creeks based on maximum interclass variance [23]. From a morphological perspective, Gong et al. used the Normalized Difference Water Body Index (NDWI) and tensor voting algorithms to extract the linear features of tidal creeks and further analyze their morphological features [24]. This method is capable of quantitatively detecting tidal creeks, but its limitation lies in its ability to only extract large-scale tidal creeks, and it struggles with extracting smaller-scale tidal creek features. Chirol et al. proposed an enhanced Managed Realignment (MR) method by integrating a LiDAR dataset, semi-automatic creek extraction algorithms based on elevation and slope thresholds, and a new algorithm for creek morphology analysis. The tidal creek network algorithm was successfully parameterized by integrating the LiDAR dataset, semi-automatic creek extraction algorithms based on elevation and slope thresholding, and a new algorithm for creek morphology analysis [25]. This approach helps to analyze the structural characteristics of tidal creeks, but is less time-efficient and less effective for large sample datasets. In contrast, Hyejin Kim et al. utilized LiDAR data with ground-based filtering techniques, selecting three validated filtering techniques—Adaptive Triangulated Irregular Network (ATIN), gLiDAR, and fabric simulation filtering (CSF)—and optimized their application for tidal creek extraction. They ultimately calculated the depth of the tidal creek and generated a tidal creek map [26]. The improved CSF method requires minimal user intervention and has greater automation in tidal creek extraction, but it is still difficult to accurately distinguish between tidal creeks and other features when the height difference between tidal creeks and their surroundings is small or when the ground height changes rapidly.
Although all of these methods have their advantages in tidal creek extraction, they also have limitations, such as low processing efficiency, poor timeliness, and limited ability to capture detailed features. Therefore, combining modern deep learning methods can effectively address these shortcomings and improve the accuracy and efficiency of tidal creek extraction, particularly in complex environments and with high-resolution remote sensing imagery, thereby further optimizing the analysis and extraction of tidal creek morphological features.
Deep learning, in contrast to earlier machine learning models, can automatically learn high-level abstract features from raw data using numerous layers of nonlinear processing. This reduces the need for manual feature engineering and enables the handling of more complex patterns and datasets. Isikdogan et al. proposed a modified loss function for detecting creek centerlines, producing outputs refined to a single-pixel-width skeletal system, which facilitates the construction of a network representing seashore river morphological structures [27]. Compared to other classical semantic segmentation models, such as DeepLab, PSPNet, and HRNet, U-Net [28] is better suited for scenarios with a limited number of features, requires a relatively shallow feature set, and is particularly well adapted to the characteristics of tidal creeks. Hübinger et al. used U-Net with InSAR imagery to detect hydrological barriers and assess the hydrological connectivity of wetlands, aiding in the effective management of decentralized wetland systems [29]. Deng et al. optimized the ResU-Net model for Sentinel-2 imagery to map coastal wetlands in Vietnam for real-time monitoring of wetland changes [30]. Dttu et al. improved the U-Net model for segmenting tidal creeks off the coast of Georgia, USA, to obtain high-resolution aerial data for assessing tidal impact [31]. U-Net was initially developed for medical picture segmentation but has now been successfully deployed to coastal wetland classification, demonstrating superior performance.
The objective of this study is to develop a tidal creek extraction model for the Yellow River Delta. To accomplish this, we use Gaofen-2 satellite imagery to segment tidal creeks from tidal flats. Various models are compared, and tidal creek extraction modeling is conducted to accurately identify creeks of different scales and morphologies in the Yellow River Delta. Additionally, the selected model is optimized to enable fast and efficient extraction of tidal creeks.

2. Materials and Methods

2.1. Study Area

In this research, we selected the Yellow River Delta as the study area, as illustrated in Figure 1. The Yellow River Delta is located on the northeastern coast of Shandong Province, bordered by Laizhou Bay to the south and Bohai Bay to the north. It covers an area of approximately 1530 square kilometers, between 37°34.76′–38°12.31′N and 118°32.98′–119°20.45′E. The region is a deltaic alluvial plain characterized by flat topography and a standard altitude of less than 10 m. The distance from the apex of the delta to the coastline is about 100 km. The hydrology of the region is influenced by the flow dynamics of the Yellow River, where periodic variations in water velocity and discharge lead to corresponding changes in the morphology and distribution of tidal creeks. In recent years, conservation efforts have included clearing 131,000 mu of Spartina alterniflora, returning 72,500 mu of farmland to wetlands and tidal flats, and restoring 52,000 mu of saline–alkali soil, seagrass beds, and other native vegetation. These initiatives have led to the restoration of 188 square kilometers of wetlands. Currently, the natural vegetation coverage in the protected area stands at 55.1%, with tidal creeks spanning vast areas. These creeks experience substantial spatial and temporal fluctuations, which significantly affect the long-term development and use of coastal tidal flats in the surrounding regions.
The waters near the mouth of the Yellow River are influenced by the tidal hydrodynamics of both the Yellow River and the Bohai Sea, resulting in a typical irregular semidiurnal tide. The intertidal zone in this region is vast, featuring extensive chalky, sandy, and silty tidal flats, with well-developed high-, medium-, and low-tide flats, along with numerous tidal creeks [32]. Tidal creeks exhibit significant variations in width and can extend in length from a few meters to over ten kilometers, with depths ranging from 1 to 2 m. The creeks contain silty deposits, predominantly composed of thick floating mud at their bottoms. Tidal creeks are classified based on their location on the beach surface into two main types: salt marsh tidal creeks and mudflat tidal creeks [33]. Salt marsh tidal creeks are characterized by salt marsh vegetation, displaying braided and radial patterns, wide channels, and seawater-filled beds. In contrast, mudflat tidal creeks are mostly found to the east and west of the former Yellow River estuary, characterized by dendritic, densely packed networks with sparse salt marsh vegetation.

2.2. Dataset

In this study, multispectral data from the domestically produced Gaofen-2 (GF-2) satellite were utilized. The satellite is equipped with two high-resolution cameras that enable continuous imaging, featuring a five-day revisit cycle. Each image captured by the Gaofen-2 satellite includes four multispectral bands along with a single panchromatic band. Table 1 presents the specific wavelength ranges for each of these bands. This paper outlines a methodology for preprocessing multispectral data, including orthometric correction, radiometric calibration, and atmospheric correction. These steps are essential for minimizing the effects of atmospheric conditions and illumination on surface reflectance, thereby restoring true surface reflectance. The preprocessing is performed using ENVI 5.3 software. Radiometric calibration is carried out with the Radiometric Calibration tool, while atmospheric correction employs the FLASSH model based on the MODTRAN algorithm. The spatial and spectral resolutions of the sensor are inversely related, with higher spatial resolution corresponding to lower spectral resolution. To address this limitation, the NNDiffuse Pan Sharpening tool was employed to integrate multispectral and panchromatic data, resulting in a final spatial resolution of 4 m.
To assure the integrity and completeness of the tidal stream morphology data, we selected low or ebb tide periods during the period from 2020 to 2023, when the tidal flats in the Yellow River Delta were fully exposed, and used multiple sets of image data to compile the dataset under minimal cloud cover. The selection of these images ensured a clear representation of tidal creek features. Mudflat creeks and salt marsh creeks exhibit distinct morphological characteristics. Consequently, the data were divided into two datasets for model training: mudflat tidal creeks and salt marsh tidal creeks. Large-format images were cropped into smaller input samples with dimensions of 512 × 512 × 3. To further enhance the dataset, input images were augmented through vertical and horizontal flipping, as well as random color adjustments.

2.3. Methodology

In order to more accurately capture positional and geographical details in remote sensing data and to achieve precise segmentation of tidal creeks, we propose a dynamic convolution-based residual U-Net structure for the semantic segmentation of tidal creeks. The encoder and decoder are essential components of the U-Net architecture. The encoder is tasked with gradually extracting features from the input image and reducing the spatial resolution. The decoder employs upsampling techniques to reconstruct the feature maps, ensuring they match the size of the original input image, which leads to incremental segmentation results. The U-Net technique incorporates skip connections in the decoder, which connect the encoder’s feature maps to their corresponding feature maps in the decoder. The detailed structure is presented in Figure 2.
The incorporation of skip connections enables the decoder to more effectively leverage diverse levels of feature information, thereby enhancing the precision and fidelity of image segmentation. In the encoder section, a residual block based on dynamic convolutional blocks is utilized as the foundation for each encoder block. This is followed by a maximum pooling layer, which enhances the scope for the identification of global features. Subsequently, the features derived from the encoders are conveyed to the ASPP module, which incorporates a coordinate focusing mechanism. To mitigate the challenge of feature classification imbalance, dice coefficients are integrated into the focal loss function, thereby improving the optimization of the gradient during training.

2.3.1. Omni-Dimensional Dynamic Convolution (ODConv)

To improve the segmentation performance of U-Net, we opted to use Omni-dimensional Dynamic Convolution (ODConv) [34] in place of conventional convolution within the encoder’s residual block. ODConv is a dynamic convolutional algorithm that enhances the performance of convolutional neural networks (CNNs) by adaptively adjusting the shape and size of the convolutional kernel based on the features of the input data during the convolution process. This method enables ODConv to effectively accommodate a wide variety of input data. It incorporates a learnable deformation module that modifies the convolution kernel in real time, considering not only the spatial dimensions and input/output channel dimensions but also the kernel’s shape and size. Consequently, ODConv enhances the feature extraction capabilities of CNNs, allowing them to better handle diverse characteristics of input data.
ODConv, in contrast to other dynamic convolutional methods, employs a single convolutional kernel, significantly reducing the number of parameters. Despite this reduction, ODConv preserves high accuracy and efficiency, demonstrating robust generalization capabilities that are particularly well suited for detecting tidal creeks.
Unlike other dynamic convolutional models, ODConv employs a parallel technique to implement a multidimensional attention mechanism. This approach enables the model to achieve greater versatility in attention across multiple dimensions of the convolutional kernel space. The attention mechanism in dynamic convolution is crucial, and an efficient design can significantly improve the accuracy of lightweight CNNs while maintaining rapid inference. The formula is as follows:
y = ( α w 1 ⊙ α f 1 ⊙ α c 1 ⊙ α s 1 ⊙ W 1 + … + α w n ⊙ α f n ⊙ α c n ⊙ α s n ⊙ W n ) * x
where the scalars α w i , α f i , α c i , α s i , and W i represent the entire convolution kernel, the output channel, the input channel, the convolution kernel space, and the convolution kernel itself, respectively.
In the U-Net encoder block, ODConv replaces the traditional convolution operation. ODConv takes into account not only the spatial, input, and output channel dimensions but also the shape and size of the convolution kernel. This adaptive approach enhances the segmentation accuracy of tidal creeks by tailoring the convolution process to the characteristics of different input data.

2.3.2. Atrous Spatial Pyramid Pooling Utilizing a Coordinate Attention Mechanism (CA-ASPP)

Atrous Spatial Pyramid Pooling (ASPP) is a deep learning technique designed for extracting multiscale features, commonly applied in semantic segmentation tasks. ASPP was initially introduced in DeepLabV2 [35], with a later improvement in DeepLabV3+ [36]. This improvement recommended substituting the standard convolution in ASPP with depthwise convolution to minimize the number of parameters and accelerate computations. ASPP consists of several key components, including 1 × 1 convolution, atrous convolution, and a pooling pyramid, which captures features from various receptive fields using atrous convolution with varied dilation rates. The model employs global pooling and 1 × 1 convolution to aggregate and refine features, effectively capturing global contextual information. The combination and integration of these multiscale elements enhance the model’s expressive capacity.
We incorporated coordinate attention into the original ASPP module to enhance the model’s capability to identify orientation-specific and location-sensitive features in images. This method retains spatial dependency and precise localization information across vertical (H) and horizontal (W) dimensions. The coordinate attention (CA) algorithm utilizes global average pooling to compute channel attention weights while encoding global spatial information, considering both channel and location relationships. This enables the model to effectively capture correlation information between distant targets. Below is a schematic diagram illustrating the multiscale module utilizing a coordinate attention mechanism (CA-ASPP) (Figure 3).
For a given input, the step equation for the n-th channel is expressed as follows:
y i = 1 H × W ∑ m = 1 H ∑ n = 1 W x i ( m , n )
where y i is the output of channel i, and x i ( m , n ) represents the value of x at coordinates (m,n).
To preserve the spatial information associated with global pooling, the CA module, after acquiring the vectors in both spatial directions of the input, splits the channel attention into two concurrent one-dimensional feature coding operations to average the pooling of the vectors in each direction. The input features in (2) are combined in both vertical and horizontal orientations to create two distinct, direction-aware feature maps using two 1D pooling processes. For a specified input, each channel is encoded using two pooling kernels along the horizontal and vertical axes, namely (H, 1) for the horizontal axis and (1, W) for the vertical axis, with each kernel addressing different spatial ranges. Equations (3) and (4) represent the outputs.
y i h ( h ) = 1 W ∑ 0 ≤ m < W x i ( h , m )
y i w ( w ) = 1 H ∑ 0 ≤ n < H x i ( n , w )
Spatial dimension concatenation can be applied to the average pooled result, and then the channel dimensions can be reduced using convolution. The expression is given in (5):
f = δ ( F 1 ( [ y h , y w ] ) )
where δ denotes a nonlinear activation function. The function f ∈ R C r × ( H + W ) redistributes the complete feature vector into two directional vectors, labeled f h and f w , where r is a reduction ratio utilized to limit the block size. Then, the number of channels of the eigenvectors in both directions is rescaled by applying two convolutional layers, denoted as F h and F w , respectively, to yield g h and g w . These are finally weighted in both directions with the original input information, as illustrated in (6).
z i ( m , n ) = x i ( m , n ) × g i h ( m ) × g i w ( n )
In this study, the optimized features of the ASPP module are provided as input to the CA module to enhance orientation and position recognition within the image. Additionally, the remotely sensed image preserves contextual information from a distance. Figure 4 illustrates the formation of coordinate attention.

2.3.3. Loss Function

In tidal creek segmentation, the background is significantly overshadowed by the tidal creek. However, the cross-entropy loss function, commonly used in semantic segmentation, treats each category equally and is susceptible to issues related to category imbalance. To mitigate the issue of imbalanced positive and negative samples, we use a hybrid loss function that incorporates dice loss [37] with focal loss [38]. Dice loss helps to address sample imbalance by learning the category distribution, while focal loss focuses the model’s attention on underrepresented categories in datasets with skewed class distributions.
Focal loss builds upon the standard cross-entropy loss by incorporating an adjustment factor, ( 1 − p n ( c ) ) 2 , which reduces the influence of easily classifiable samples and increases the importance of harder-to-classify ones. This strategy solves category imbalance by giving greater weight to more difficult samples, which improves the model’s performance. The formula for focal loss is as follows:
L F ocal = ∑ c = 0 C − 1 ∑ n = 1 N g n ( c ) ( 1 − p n ( c ) ) 2 l o g ( p n ( c ) )
where g n c is the target value, p n c is the predicted probability, and ( 1 − p n ( n ) ) 2 is the adjustment factor that reduces the weight of easily classifiable samples to focus more on difficult ones.
In this study, the focal loss function is paired with the dice loss function to address the class imbalance issue during model training. Dice loss helps to learn the class distribution to mitigate pixel imbalance, while focal loss emphasizes learning from misclassified pixels. The total loss function is given by (8):
L = L D i c e + λ L F o c a l = C − ∑ c = 0 C − 1 T P p ( c ) T P n ( c ) + α F N p ( c ) + β F P p ( c ) − λ 1 N ∑ c = 0 C − 1 ∑ n = 1 N g n ( c ) ( 1 − p n ( c ) ) 2 l o g ( p n ( c ) )
where T P n ( c ) , F N p ( c ) , and F P p ( c ) denote the true positives, false negatives, and false positives of class c, respectively, and C signifies the total number of classes. The parameters α and β balance the impact of false negatives and false positives on the loss function. The dice coefficient emphasizes the overall consistency of the prediction results, while focal loss addresses the uneven distribution of positive and negative samples by adjusting the weights.

3. Results

3.1. Evaluation Indicators

Overall accuracy is a crucial metric for evaluating the performance of a classification model. It is defined as the ratio of the number of correctly classified samples to the total number of samples. This metric reflects the model’s overall performance across all categories. This indicator can be misleading in situations of class imbalance, as high accuracy may mask poor performance in specific categories. Consequently, we chose recall, the F1 score, and the kappa coefficient as evaluation metrics to be used in conjunction for a more comprehensive assessment of the model.
A c c = T P + T N T P + T N + F P + F N
Recall measures the proportion of successfully identified positive samples, indicating the number of true positive predictions made from all actual positive instances in the collection. Sensitivity is calculated in the same manner as recall.
R e c a l l = T P T P + F N
Precision measures the accuracy of forecasts by representing the proportion of samples categorized as positive that are genuinely positive. The F1 score, which is particularly relevant in cases of class imbalance, reflects the overall performance of the model in identifying positive class samples.
P r e = T P T P + F P
F 1 = 2 × s e n × p r e s e n + p r e
The kappa coefficient is a metric that assesses the classification accuracy by quantifying the level of agreement between predicted and actual values [39].
K a p p a = 2 × ( T P × T N − F P × F N ) ( T P + F P ) × ( F P × T N ) + ( T P + F N ) × ( F N + T N )
TP and TN represent the correctly classified tidal creeks and backgrounds, respectively. FP indicates the misclassification of tidal creeks as backgrounds, while FN refers to the misclassification of backgrounds as tidal creeks. Collectively, these metrics—accuracy, recall, F1 score, and kappa coefficient—offer a thorough assessment of the model’s segmentation performance.

3.2. Experiment Results

U-Net, a widely adopted base architecture for semantic segmentation, is frequently employed to map and analyze coastal marshes. To evaluate the efficiency of the ODU-Net modules, we ran separate ablation experiments on the mudflat creek and salt marsh creek datasets. The models’ performance was assessed using a variety of metrics on the test data. The comparative analysis on the bog and salt marsh datasets is shown in Table 2 and Table 3.
In this experiment, U-Net was used as the benchmark model, with ODConv replacing traditional convolution to construct the encoding block. The multiscale module incorporated ASPP and employed the focal–dice loss function. Building on this foundation, a CA-ASPP module was gradually implemented. Each module within ODU-Net made a significant contribution to the model’s performance. After integrating the improved ResBlock, the kappa coefficients for the salt marsh and mudflat tidal creeks increased by 1.45% and 3.49%, respectively, indicating that ResBlock effectively enhances segmentation consistency. As shown in Figure 5, the vanilla U-Net model effectively captured the overall morphological structure of the tidal creeks; however, it displayed voids in the wider tidal creeks and breaks in the narrower ones. After integrating the enhanced ResBlock, the continuity of the tidal creeks post-segmentation improved, resulting in a reduction in voids in the wider creeks. Nevertheless, this enhancement was accompanied by an increase in the segmentation of background regions as tidal creeks.
Additionally, incorporating the coordinate attention module resulted in improvements in recall of 3.78% and 3.33%. This indicates that the mechanism allows the model to concentrate on essential contextual and location-sensitive features, thereby enhancing tidal creek segmentation. The results demonstrate that the CA-ASPP module enhances the model’s ability to prioritize targets of interest. Furthermore, the modified loss function increased the kappa coefficient of ODU-Net by 2.30% and 2.23%, and the F1 score by 2.10% and 2.19%, while also boosting the classification accuracy for small objects. This indicates that incorporating the dice coefficient into the focal loss function effectively addresses class imbalance issues. As illustrated in Figure 6, the incorporation of the CA-ASSP module enhanced the continuity of wide tidal creeks and improved the detection of smaller tidal creeks. Although the segmentation performance for small tidal creeks showed improvement, some interruptions persisted, and the misclassification of background pixels as wide tidal creeks remained a challenge. However, after implementing the enhanced loss function, the model placed greater emphasis on finer tidal creeks, resulting in a significant reduction in interruptions. Additionally, the improved loss function substantially decreased the misclassification rate.
The comparison of the tidal creek segmentation results across the two regions demonstrates that ODU-Net effectively segments tidal creeks in high-quality images with a resolution of 4 m. Both the F1 score and kappa coefficient exceeded 0.8, indicating the model’s ability to accurately capture features relevant to the classification task. The model’s overall accuracy exceeded 97%, with recall rates of 83.94% and 92.10% for the two regions, respectively. These results confirm the model’s effectiveness in accurate recognition and its ability to capture essential features for segmentation.
Figure 7 illustrates the dendritic pattern of mud marsh tidal creeks and the braided, radial pattern of salt marsh tidal creeks. The model effectively captures the overall morphological spatial characteristics of the tidal creeks, demonstrating its ability to identify features as small as 5 pixels at a 4 m resolution. This study underscores the exceptional performance of ODU-Net in correctly detecting and segmenting tidal creek features. Additionally, ODU-Net operates effectively in diverse environments, enabling it to detect tidal streams at fine spatial scales.

3.3. Comparisons and Analysis

To evaluate the efficacy of the proposed method, four deep learning semantic segmentation models and methods for edge detection were selected for comparison. The performance of ODU-Net was evaluated against several classical segmentation models, including PSPNet, DeepLabV3+, PoolFormer [40], and SegNeXt [41]. The segmentation accuracies of these models were assessed on the mudflat creek dataset and salt marsh creek dataset. Table 4 and Table 5 present the results of the comparative experiments.
Edge detection is a useful technique, particularly when the image boundaries are well defined and contrast is high. However, the high background heterogeneity of tidal creeks, often associated with complex or indistinct boundaries, renders edge detection algorithms vulnerable to background variations in tidal creek feature extraction, leading to inaccurate recognition results. In this research, we selected the Canny operator for comparative analysis. Although the Canny operator is effective under certain conditions, it does not perform well in the tidal creek region due to the diversity and complexity of the background. As illustrated in Figure 8, the Canny operator fails to precisely identify the boundaries of tidal creeks in mudflat and salt marsh regions, especially those with high background heterogeneity. This indicates that traditional edge detection methods have significant limitations when dealing with terrain features such as tidal creeks with high background complexity. Therefore, in the automated extraction of tidal creeks, a single edge detection method may not adequately address the challenges in complex environments.
In comparison with four semantic segmentation models, the experimental results demonstrate that both the PSPNet and DeepLabV3+ models exhibit limited performance with small datasets, achieving an F1 score of approximately 63%. This limitation hinders their ability to effectively segment and extract tidal creeks, resulting in poor segmentation performance. In comparison, our proposed ODU-Net improves the F1 score and kappa coefficient by 5.91% and 5.90%, respectively, over PoolFormer, which is based on the Transformer architecture. Furthermore, the SegNeXt model, which uses cross-scale spatial pyramids to merge information from different levels through skip connections, is outperformed by ODU-Net with an increase of approximately 5.24% in both the F1 score and kappa coefficient.
Based on the data presented, the suggested ODU-Net significantly enhances the overall accuracy of tidal creek segmentation in the Yellow River Delta. The integration of the coordinate attention module significantly improves the model’s performance in segmenting tidal creek data. Additionally, adjustments made to the loss function notably enhance the segmentation accuracy for small targets.
To further illustrate the segmentation outcomes, Figure 6 visualizes and analyzes the results from the different models. Tidal creek images of various scales from the test set are used to compare the segmentation results of PSPNet, DeepLabV3+, PoolFormer, SegNeXt, and the proposed ODU-Net. The comparative results are shown in Figure 9.
The figure clearly demonstrates that PoolFormer often misclassifies background areas as wide tidal creeks near the ocean. SegNeXt shows noticeable discontinuities in the segmentation results of small tidal creeks, particularly those less than 5 pixels in width. PSPNet produces fragmented segmentation for fine tidal creeks, failing to effectively identify smaller tidal creeks compared to the other models.
Compared to mudflats, salt marshes exhibit greater background heterogeneity, making tidal creek extraction more challenging. On the salt marsh test set, the F1 scores for PSPNet and DeepLabV3+ are approximately 60%, with kappa coefficients around 61%, highlighting limitations in segmenting tidal creeks in these environments. In comparison to SegNeXt, our proposed ODU-Net improves the F1 score and kappa coefficient by 7.10% and 7.12%, respectively, for salt marsh tidal creek extraction. Among the five models tested, ODU-Net achieved the highest recall rate, indicating superior performance in accurately identifying tidal creeks. Figure 10 presents the extraction results for salt marsh tidal creeks, where marked areas reveal that DeeplabV3+ and SegNeXt struggle with the segmentation of smaller creeks. PSPNet experiences discontinuities in wider creek extraction, while PoolFormer demonstrates significantly better performance in extracting mudflat creeks than salt marsh creeks. When applied to salt marshes, the results reveal considerable fragmentation, with many tidal creeks appearing discontinuous. In contrast, ODU-Net successfully identifies most of the fine tidal creeks, with the shapes in the segmentation results closely aligning with the true features. Overall, ODU-Net’s segmentation performance on tidal creeks surpasses that of other semantic segmentation methods

4. Discussions

The experimental results demonstrate that the ODU-Net model performs well across various metrics (e.g., F1 score, kappa coefficient, recall, and overall precision), consistent with findings from other studies reporting similar performance improvements in semantic segmentation tasks using deep learning models on high-resolution remote sensing images (e.g., Yin et al. [42]). The design of the residual convolution block and coordinate attention mechanism based on Omni-dimensional Dynamic Convolution enables the model to effectively utilize detailed information in 4 m resolution remote sensing images, significantly enhancing segmentation accuracy for small targets, similar to approaches described by Fu et al. [43], who emphasized the importance of multiscale feature extraction in improving segmentation precision.
The primary advantage of our method lies in its exceptional performance in boundary identification and segmentation of small targets in the Yellow River Delta tidal creeks, which aligns with the conclusions of Suwanprasit et al. [44], who demonstrated that high-resolution imagery is crucial for distinguishing fine-scale morphological features in complex environments. The use of 4 m resolution images in this study enabled the accurate segmentation of subtle tidal creek boundaries, similar to other studies focusing on fine-scale topographic feature extraction (e.g., F. Marques et al. [45]). In contrast, lower-resolution images often lead to the loss of detailed information, resulting in lower segmentation accuracy, an issue we address by employing high-resolution data to improve operational efficiency.
Tidal creeks in the mud marsh region of the Yellow River Delta typically display complex net-like or tree-like branching patterns, primarily shaped by tidal forces and sediment deposition. Mou et al. [46] similarly observed that tidal creeks in such regions often display intricate branching patterns, which are difficult to accurately segment without advanced models. ODU-Net, by achieving a high degree of accuracy in segmenting these complex structures, supports the findings of Dutt et al. [31], who demonstrated that deep learning models can effectively capture the topological relationships of watercourse networks, particularly in dynamic coastal environments. The model’s ability to accurately segment both mire and salt marsh tidal creeks demonstrates its strong topological characterization capabilities.
Despite the excellent performance of the model, it remains constrained by the resolution of satellite remote sensing images. We observed that in 4 m resolution images, ODU-Net encounters difficulties in segmenting very fine tidal creek features, such as end tidal creeks smaller than 2 pixels. This limitation arises from the challenges that high-resolution imagery faces in capturing extremely fine details, as discussed in Zhu et al. [47]. Future work will explore the use of higher-resolution remote sensing images and advanced image enhancement techniques to improve the segmentation of these fine targets and address these resolution constraints.
In future work, we intend to extend the ODU-Net model to other regional datasets, a strategy successfully employed in other studies (e.g., Elshamli et al. [48]) using transfer learning techniques to adapt the model to new regions. This approach enables the fine-tuning of models to capture regional variations, as suggested by Jin et al. [49]. Additionally, we will explore multimodal data fusion to further enhance the robustness and generalization of ODU-Net across diverse geographical regions.

5. Conclusions

The segmentation of tidal creeks is crucial for differentiating creeks from rivers at high resolutions and in vast geographical areas, significantly contributing to the study of coastal wetlands. In this study, we assess ODU-Net, an innovative deep neural network model designed for tidal creek segmentation, using data with a resolution of 4 m. The incorporation of the ResNet module addresses challenges such as gradient vanishing, while the coordinate attention enhances the model’s ability to capture relationships among various targets, thereby preserving spatial contextual information in imagery. Additionally, the modified loss function improves model accuracy by addressing the imbalance between positive and negative image categories during training. The main conclusions of this study are presented below.
The experimental results demonstrated that the F1 scores for the bog and salt marsh datasets reached 83.20% and 83.19%, respectively, while the kappa coefficients were 81.69% and 82.32%. The overall accuracies for the mudflat creek and salt marsh creek test sets were 97.46% and 98.37%, respectively. These findings indicate that ODU-Net can significantly enhance the accuracy of tidal creek segmentation in remotely sensed imagery.
The proposed model outperformed DeepLabV3+, PSPNet, PoolFormer, and SegNeXt in segmenting tidal creeks. The F1 score, kappa coefficient, and recall rate exhibit significant improvements over the other models, with the final predicted results of the tidal creeks closely aligning with actual conditions.

Author Contributions

Methodology, B.C.; data curation, Q.Z.; writing—original draft, X.W.; investigation, Y.C.; writing—review and editing, N.Y., X.Z. and S.W. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported in part by the Key Research and Development Program of Shandong Province, China, under Grant 2023CXPT054.

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Zhang, Z.; Xu, N.; Li, Y.; Li, Y. Sub-Continental-Scale Mapping of Tidal Wetland Composition for East Asia: A Novel Algorithm Integrating Satellite Tide-Level and Phenological Features. Remote Sens. Environ. 2022, 269, 112799. [Google Scholar] [CrossRef]
  2. Murray, N.J.; Phinn, S.R.; DeWitt, M.; Ferrari, R.; Johnston, R.; Lyons, M.B.; Clinton, N.; Thau, D.; Fuller, R.A. The Global Distribution and Trajectory of Tidal Flats. Nature 2019, 565, 222–225. [Google Scholar] [CrossRef] [PubMed]
  3. Shen, Y.-m.; Zhang, R.-s.; Wang, Y.-h. The Tidal Creek Character in Salt Marsh of Spartina Alterniflora Loisel on Strong Tide Coast. Geogr. Res. 2003, 22, 520–527. [Google Scholar] [CrossRef]
  4. Newell, R.I.E.; Marshall, N.; Sasekumar, A.; Chong, V.C. Relative Importance of Benthic Microalgae, Phytoplankton, and Mangroves as Sources of Nutrition for Penaeid Prawns and Other Coastal Invertebrates from Malaysia. Mar. Biol. 1995, 123, 595–606. [Google Scholar] [CrossRef]
  5. Dittmar, T.; Lara, R.J. Driving Forces Behind Nutrient and Organic Matter Dynamics in a Mangrove Tidal Creek in North Brazil. Estuar. Coast. Shelf Sci. 2001, 52, 249–259. [Google Scholar] [CrossRef]
  6. Gleeson, J.; Santos, I.R.; Maher, D.T.; Golsby-Smith, L. Groundwater–Surface Water Exchange in a Mangrove Tidal Creek: Evidence from Natural Geochemical Tracers and Implications for Nutrient Budgets. Mar. Chem. 2013, 156, 27–37. [Google Scholar] [CrossRef]
  7. Mallik, A.; Chakraborty, P.; Ramteke, K.K.; Bhavan, S.G.; Jaiswar, A.K.; Nayak, B.B.; Pal, P.; Bhushan, S. Diet Characteristics of Tidal Creek-Associated Fishes of the Northeastern Arabian Sea with Special Reference to Microplastic Ingestion. Chemosphere 2024, 363, 142886. [Google Scholar] [CrossRef] [PubMed]
  8. Laffaille, P.; Feunteun, E.; Lefeuvre, J.-C. Composition of Fish Communities in a European Macrotidal Salt Marsh (the Mont Saint-Michel Bay, France). Estuar. Coast. Shelf Sci. 2000, 51, 429–438. [Google Scholar] [CrossRef]
  9. Lerberg, S.B.; Holland, A.F.; Sanger, D.M. Responses of Tidal Creek Macrobenthic Communities to the Effects of Watershed Development. Estuaries 2000, 23, 838–853. [Google Scholar] [CrossRef]
  10. Ning, Z.; Li, D.; Chen, C.; Xie, C.; Chen, G.; Xie, T.; Wang, Q.; Bai, J.; Cui, B. The Importance of Structural and Functional Characteristics of Tidal Channels to Smooth Cordgrass Invasion in the Yellow River Delta, China: Implications for Coastal Wetland Management. J. Environ. Manag. 2023, 342, 118297. [Google Scholar] [CrossRef]
  11. Wang, B.; Cheng, W.; Xu, H.; Wang, R.; Song, K.; Bao, A.; Shi, Q. Vegetation Differentiation Characteristics and Control Mechanisms in the Altay Region Based on Topographic Gradients. Ecol. Indic. 2024, 160, 111838. [Google Scholar] [CrossRef]
  12. Kearney, W.S.; Fagherazzi, S. Salt Marsh Vegetation Promotes Efficient Tidal Channel Networks. Nat. Commun. 2016, 7, 12287. [Google Scholar] [CrossRef]
  13. Christiansen, T.; Wiberg, P.L.; Milligan, T.G. Flow and Sediment Transport on a Tidal Salt Marsh Surface. Estuar. Coast. Shelf Sci. 2000, 50, 315–331. [Google Scholar] [CrossRef]
  14. Voulgaris, G.; Meyers, S.T. Temporal Variability of Hydrodynamics, Sediment Concentration and Sediment Settling Velocity in a Tidal Creek. Cont. Shelf Res. 2004, 24, 1659–1683. [Google Scholar] [CrossRef]
  15. Temmerman, S.; Bouma, T.J.; Govers, G.; Lauwaet, D. Flow Paths of Water and Sediment in a Tidal Marsh: Relations with Marsh Developmental Stage and Tidal Inundation Height. Estuaries 2005, 28, 338–352. [Google Scholar] [CrossRef]
  16. Zhao, B.; Liu, Y.; Xu, W.; Liu, Y.; Sun, J.; Wang, L. Morphological Characteristics of Tidal Creeks in the Central Coastal Region of Jiangsu, China, Using LiDAR. Remote Sens. 2019, 11, 2426. [Google Scholar] [CrossRef]
  17. Halls, J.; Costin, K. Submerged and Emergent Land Cover and Bathymetric Mapping of Estuarine Habitats Using WorldView-2 and LiDAR Imagery. Remote Sens. 2016, 8, 718. [Google Scholar] [CrossRef]
  18. Fagherazzi, S.; Hannion, M.; D’Odorico, P. Geomorphic Structure of Tidal Hydrodynamics in Salt Marsh Creeks. Water Resour. Res. 2008, 44, 2007WR006289. [Google Scholar] [CrossRef]
  19. Magolan, J.L.; Halls, J.N. A Multi-Decadal Investigation of Tidal Creek Wetland Changes, Water Level Rise, and Ghost Forests. Remote Sens. 2020, 12, 1141. [Google Scholar] [CrossRef]
  20. Paynter, I.; Schaaf, C.; Bowen, J.L.; Deegan, L.; Peri, F.; Cook, B. Characterizing a New England Saltmarsh with NASA G-LiHT Airborne Lidar. Remote Sens. 2019, 11, 509. [Google Scholar] [CrossRef]
  21. Lohani, B.; Mason, D.C. Application of Airborne Scanning Laser Altimetry to the Study of Tidal Channel Geomorphology. ISPRS J. Photogramm. Remote Sens. 2001, 56, 100–120. [Google Scholar] [CrossRef]
  22. Yang, Y. Landsat 8 OLI Image Based Terrestrial Water Extraction from Heterogeneous Backgrounds Using a Reflectance Homogenization Approach. Remote Sens. Environ. 2015, 171, 14–32. [Google Scholar] [CrossRef]
  23. Gong, Z.; Wang, Q.; Guan, H.; Zhou, D.; Zhang, L.; Jing, R.; Wang, X.; Li, Z. Extracting Tidal Creek Features in a Heterogeneous Background Using Sentinel-2 Imagery: A Case Study in the Yellow River Delta, China. Int. J. Remote Sens. 2020, 41, 3653–3676. [Google Scholar] [CrossRef]
  24. Gong, Z.; Mou, K.; Wang, Q.; Qiu, H.; Zhang, C.; Zhou, D. Parameterizing the Yellow River Delta Tidal Creek Morphology Using Automated Extraction from Remote Sensing Images. Sci. Total Environ. 2021, 769, 144572. [Google Scholar] [CrossRef] [PubMed]
  25. Chirol, C.; Haigh, I.D.; Pontee, N.; Thompson, C.E.; Gallop, S.L. Parametrizing Tidal Creek Morphology in Mature Saltmarshes Using Semi-Automated Extraction from Lidar. Remote Sens. Environ. 2018, 209, 291–311. [Google Scholar] [CrossRef]
  26. Kim, H.; Kim, Y.; Lee, J. Tidal Creek Extraction from Airborne LiDAR Data Using Ground Filtering Techniques. KSCE J. Civ. Eng. 2020, 24, 2767–2783. [Google Scholar] [CrossRef]
  27. Isikdogan, F.; Bovik, A.; Passalacqua, P. Automatic Channel Network Extraction From Remotely Sensed Images by Singularity Analysis. IEEE Geosci. Remote Sens. Lett. 2015, 12, 2218–2221. [Google Scholar] [CrossRef]
  28. Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015; Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F., Eds.; Lecture Notes in Computer Science; Springer International Publishing: Cham, Switzerland, 2015; Volume 9351, pp. 234–241. ISBN 978-3-319-24573-7. [Google Scholar]
  29. Hübinger, C.; Fluet-Chouinard, E.; Hugelius, G.; Peña, F.J.; Jaramillo, F. Automating the Detection of Hydrological Barriers and Fragmentation in Wetlands Using Deep Learning and InSAR. Remote Sens. Environ. 2024, 311, 114314. [Google Scholar] [CrossRef]
  30. Dang, K.B.; Nguyen, M.H.; Nguyen, D.A.; Phan, T.T.H.; Giang, T.L.; Pham, H.H.; Nguyen, T.N.; Tran, T.T.V.; Bui, D.T. Coastal Wetland Classification with Deep U-Net Convolutional Networks and Sentinel-2 Imagery: A Case Study at the Tien Yen Estuary of Vietnam. Remote Sens. 2020, 12, 3270. [Google Scholar] [CrossRef]
  31. Dutt, R.; Ortals, C.; He, W.; Curran, Z.C.; Angelini, C.; Canestrelli, A.; Jiang, Z. A Deep Learning Approach to Segment Coastal Marsh Tidal Creek Networks from High-Resolution Aerial Imagery. Remote Sens. 2024, 16, 2659. [Google Scholar] [CrossRef]
  32. Yanxia, L.I.U.; Haijun, H.; Zhongfeng, Q.I.U.; Jitao, C.; Xiguang, Y. Monitoring Change and Position of Coastlines from Satellite Images Using Slope Correction in a Tidal Flat: A Case Study in the Yellow River Delta. Acta Geogr. Sin. 2012, 67, 377–387. [Google Scholar] [CrossRef]
  33. Healey, R.G.; Pye, K.; Stoddart, D.R.; Bayliss-Smith, T.P. Velocity Variations in Salt Marsh Creeks, Norfolk, England. Estuar. Coast. Shelf Sci. 1981, 13, 535–545. [Google Scholar] [CrossRef]
  34. Li, C.; Zhou, A.; Yao, A. Omni-Dimensional Dynamic Convolution. arXiv 2022, arXiv:2209.07947. [Google Scholar]
  35. Chen, L.-C.; Papandreou, G.; Kokkinos, I.; Murphy, K.; Yuille, A.L. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs 2017. IEEE Trans. Pattern Anal. Mach. Intell. 2017, 40, 834–848. [Google Scholar] [CrossRef] [PubMed]
  36. Chen, L.-C.; Zhu, Y.; Papandreou, G.; Schroff, F.; Adam, H. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018. [Google Scholar]
  37. Milletari, F.; Navab, N.; Ahmadi, S.-A. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In Proceedings of the 2016 Fourth International Conference on 3D Vision (3DV), Stanford, CA, USA, 25–28 October 2016. [Google Scholar]
  38. Lin, T.-Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal Loss for Dense Object Detection. IEEE Trans. Pattern Anal. Mach. Intell. 2020, 42, 318–327. [Google Scholar] [CrossRef]
  39. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv 2017, arXiv:1412.6980. [Google Scholar]
  40. Yu, W.; Luo, M.; Zhou, P.; Si, C.; Zhou, Y.; Wang, X.; Feng, J.; Yan, S. MetaFormer Is Actually What You Need for Vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 18–24 June 2022. [Google Scholar]
  41. Guo, M.-H.; Lu, C.-Z.; Hou, Q.; Liu, Z.; Cheng, M.-M.; Hu, S.-M. SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation. Adv. Neural Inf. Process. Syst. 2022, 35, 1140–1156. [Google Scholar]
  42. Yin, P.; Zhang, D.; Han, W.; Li, J.; Cheng, J. High-Resolution Remote Sensing Image Semantic Segmentation via Multiscale Context and Linear Self-Attention. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 9174–9185. [Google Scholar] [CrossRef]
  43. Fu, L.; Chen, J.; Wang, Z.; Zang, T.; Chen, H.; Wu, S.; Zhao, Y. MSFANet: Multi-Scale Fusion Attention Network for Mangrove Remote Sensing Lmage Segmentation Using Pattern Recognition. J. Cloud Comput. 2024, 13, 27. [Google Scholar] [CrossRef]
  44. Suwanprasit, C.; Srichai, N. Impacts of Spatial Resolution on Land Cover Classification. APAN Proc. 2012, 33, 39. [Google Scholar] [CrossRef]
  45. Marques, F.; Eugenio, F.; Alfaro, M.; Marcello, J. Bathymetry Mapping Using Very High Resolution Satellite Multispectral Imagery in Shallow Coastal Waters of Protected Ecosystems. In Proceedings of the IGARSS 2019—2019 IEEE International Geoscience and Remote Sensing Symposium, Yokohama, Japan, 28 July–2 August 2019; pp. 8234–8237. [Google Scholar]
  46. Mou, K.; Gong, Z.; Qiu, H. Spatiotemporal Differentiation and Development Process of Tidal Creek Network Morphological Characteristics in the Yellow River Delta. J. Geogr. Sci. 2021, 31, 1633–1654. [Google Scholar] [CrossRef]
  47. Zhu, Y.; Zhang, X.; Jiang, B.; Su, C.; Wang, M.; Zhang, K. Comparison of Performance of Fusion Methods for Sentinel-2 and GF-2 Satellite Images and the Potential in Land Use Classification: A Case Study from Reservoirs in Hainan Island, China. In Proceedings of the 2024 12th International Conference on Agro-Geoinformatics (Agro-Geoinformatics), Novi Sad, Serbia, 15–18 July 2024; pp. 1–6. [Google Scholar]
  48. Elshamli, A.; Taylor, G.W.; Areibi, S. Multisource Domain Adaptation for Remote Sensing Using Deep Neural Networks. IEEE Trans. Geosci. Remote Sens. 2020, 58, 3328–3340. [Google Scholar] [CrossRef]
  49. Jin, H.; Mountrakis, G. Fusion of Optical, Radar and Waveform LiDAR Observations for Land Cover Classification. ISPRS J. Photogramm. Remote Sens. 2022, 187, 171–190. [Google Scholar] [CrossRef]
Remotesensing 17 00676 g001
Figure 1. Location of the study area: GF-2 images of the Yellow River Delta (RGB: 3, 2, 1 bands) (a), (b) mudflat creek area, and (c) salt marsh creek area.
Figure 1. Location of the study area: GF-2 images of the Yellow River Delta (RGB: 3, 2, 1 bands) (a), (b) mudflat creek area, and (c) salt marsh creek area.
Remotesensing 17 00676 g001
Remotesensing 17 00676 g002
Figure 2. The ODU-Net model structure.
Figure 2. The ODU-Net model structure.
Remotesensing 17 00676 g002
Remotesensing 17 00676 g003
Figure 3. CA-ASPP module structure.
Figure 3. CA-ASPP module structure.
Remotesensing 17 00676 g003
Remotesensing 17 00676 g004
Figure 4. Coordinate attention module structure.
Figure 4. Coordinate attention module structure.
Remotesensing 17 00676 g004
Remotesensing 17 00676 g005
Figure 5. Comparison of results of ablation experiments on the mudflat creek test set.
Figure 5. Comparison of results of ablation experiments on the mudflat creek test set.
Remotesensing 17 00676 g005
Remotesensing 17 00676 g006
Figure 6. Comparison of results of ablation experiments on the salt marsh creek test set.
Figure 6. Comparison of results of ablation experiments on the salt marsh creek test set.
Remotesensing 17 00676 g006
Remotesensing 17 00676 g007
Figure 7. (a) Larger spatial mudflat area; (b) prediction results for mudflat creeks; (c) larger spatial mudflat area; (d) prediction results for salt marsh creeks.
Figure 7. (a) Larger spatial mudflat area; (b) prediction results for mudflat creeks; (c) larger spatial mudflat area; (d) prediction results for salt marsh creeks.
Remotesensing 17 00676 g007
Remotesensing 17 00676 g008
Figure 8. Comparison of the edge detection results (1–3 for mudflat regions, 4–6 for salt marsh regions).
Figure 8. Comparison of the edge detection results (1–3 for mudflat regions, 4–6 for salt marsh regions).
Remotesensing 17 00676 g008
Remotesensing 17 00676 g009
Figure 9. Semantic segmentation results of different models on the mudflat creek test set.
Figure 9. Semantic segmentation results of different models on the mudflat creek test set.
Remotesensing 17 00676 g009
Remotesensing 17 00676 g010
Figure 10. Semantic segmentation results of different models on the salt marsh creek test set.
Figure 10. Semantic segmentation results of different models on the salt marsh creek test set.
Remotesensing 17 00676 g010
Table 1. Technical parameters of Gaofen-2 satellite sensor.
Table 1. Technical parameters of Gaofen-2 satellite sensor.
ParametersPanchromatic/Multispectral Cameras
Spectral ScopePanchromatic0.45~0.90 μm
Multispectral0.45~0.52 μm
0.52~0.59 μm
0.63~0.69 μm
0.77~0.89 μm
Spatial ResolutionPanchromatic0.8 m
Multispectral3.2 m
Table 2. Evaluation of various metrics for semantic segmentation in ablation experiments on mudflat creeks.
Table 2. Evaluation of various metrics for semantic segmentation in ablation experiments on mudflat creeks.
ModelF1 ScoreKappa CoefficientRecallAccuracy
Vanilla U-Net0.79290.77530.71840.9687
U-Net + Resnet Block0.80750.78980.75620.9688
U-Net + Resnet Block + CA-ASPP0.81100.79390.77970.9701
ODU-Net0.83200.81690.83940.9746
Table 3. Evaluation of various metrics for semantic segmentation in ablation experiments on marsh creeks.
Table 3. Evaluation of various metrics for semantic segmentation in ablation experiments on marsh creeks.
ModelF1 ScoreKappa CoefficientRecallAccuracy
Vanilla U-Net0.75620.74530.73690.9665
U-Net + Resnet Block0.79040.78020.77020.9799
U-Net + Resnet Block + CA-ASPP0.81000.80090.78890.9811
ODU-Net0.83190.82320.92100.9837
Table 4. Comparison of tidal creek segmentation results from different models on mudflats.
Table 4. Comparison of tidal creek segmentation results from different models on mudflats.
ModelF1 ScoreKappa
Coefficient		RecallAccuracy
DeeplabV3+0.65500.63540.54050.9620
PSPNet0.63560.61540.52140.9604
PoolFormer0.77290.75790.69250.9703
SegNeXt0.77960.76450.71400.9725
ODU-Net0.83200.81690.83940.9746
Table 5. Comparison of tidal creek segmentation results from different models on salt marshes.
Table 5. Comparison of tidal creek segmentation results from different models on salt marshes.
ModelF1 ScoreKappa
Coefficient		RecallAccuracy
DeeplabV3+0.61930.60530.52830.9720
PSPNet0.62760.61330.56820.9716
PoolFormer0.71320.70440.64140.9816
SegNeXt0.76090.75200.75850.9825
ODU-Net0.83190.82320.92100.9837
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Chen, B.; Zhang, Q.; Yang, N.; Wang, X.; Zhang, X.; Chen, Y.; Wang, S. Deep Learning Extraction of Tidal Creeks in the Yellow River Delta Using GF-2 Imagery. Remote Sens. 2025, 17, 676. https://doi.org/10.3390/rs17040676

AMA Style

Chen B, Zhang Q, Yang N, Wang X, Zhang X, Chen Y, Wang S. Deep Learning Extraction of Tidal Creeks in the Yellow River Delta Using GF-2 Imagery. Remote Sensing. 2025; 17(4):676. https://doi.org/10.3390/rs17040676

Chicago/Turabian Style

Chen, Bojie, Qianran Zhang, Na Yang, Xiukun Wang, Xiaobo Zhang, Yilan Chen, and Shengli Wang. 2025. "Deep Learning Extraction of Tidal Creeks in the Yellow River Delta Using GF-2 Imagery" Remote Sensing 17, no. 4: 676. https://doi.org/10.3390/rs17040676

APA Style

Chen, B., Zhang, Q., Yang, N., Wang, X., Zhang, X., Chen, Y., & Wang, S. (2025). Deep Learning Extraction of Tidal Creeks in the Yellow River Delta Using GF-2 Imagery. Remote Sensing, 17(4), 676. https://doi.org/10.3390/rs17040676

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop