EDWNet: A Novel Encoder–Decoder Architecture Network for Water Body Extraction from Optical Images

Abstract

Automated water body (WB) extraction is one of the hot research topics in the field of remote sensing image processing. To address the challenges of over-extraction and incomplete extraction in complex water scenes, we propose an encoder–decoder architecture semantic segmentation network for high-precision extraction of WBs called EDWNet. We integrate the Cross-layer Feature Fusion (CFF) module to solve difficulties in segmentation of WB edges, utilizing the Global Attention Mechanism (GAM) module to reduce information diffusion, and combining with the Deep Attention Module (DAM) module to enhance the model’s global perception ability and refine WB features. Additionally, an auxiliary head is incorporated to optimize the model’s learning process. In addition, we analyze the feature importance of bands 2 to 7 in Landsat 8 OLI images, constructing a band combination (RGB 763) suitable for algorithm’s WB extraction. When we compare EDWNet with various other semantic segmentation networks, the results on the test dataset show that EDWNet has the highest accuracy. EDWNet is applied to accurately extract WBs in the Weihe River basin from 2013 to 2021, and we quantitatively analyzed the area changes of the WBs during this period and their causes. The results show that EDWNet is suitable for WB extraction in complex scenes and demonstrates great potential in long time-series and large-scale WB extraction.

1. Introduction

Surface water bodies (WBs) refer to the general term for dynamic water and static water on the land surface, playing a crucial role in the global water cycle [1]. Changes in the area of WBs will have multiple impacts on ecosystems, water resources, and climate [2,3]. Therefore, it is crucial to monitor and understand the spatiotemporal dynamic changes of large-scale WBs [4,5].

Satellite remote sensing has the advantages of wide coverage, continuous observation, high spatial resolution, and rich band information, which makes satellite remote sensing an important tool for ground object observation and geographical information extraction [6,7,8,9]. With the development of satellite remote sensing technology, more and more satellite observation data are used to extract surface water information [10]. However, due to the strong similarity between WBs and surrounding land, buildings, vegetation, and other features, accurately segmenting WBs in complex scenes presents significant challenges [11].

In recent decades, numerous scholars have directed their attention to surface water mapping using remote sensing images [12,13]. Currently, research methods for extracting surface water from remote sensing images are primarily categorized into the single-band image threshold method, spectral index method, and image classification method.

The single-band threshold extraction method effectively and quickly extracts water information by extracting the grayscale value of a specific band in remote sensing images. For example, Klein et al. [14] employed temporal dynamic thresholds for individual datasets and used near-infrared (NIR) spectral data to detect water pixels from AVHRR and MODIS sensors. Lu et al. [15] created the Inland Surface Water Dataset in China (ISWDC) using threshold segmentation method on MODIS single-band images. However, the spectral reflection characteristics of WBs are influenced by many factors, and single-band images have a limited ability to represent the characteristics of ground objects.

The spectral index method for extracting WBs is based on the spectral reflection characteristics of WBs in different bands, utilizing values calculated from remote sensing data in different bands to characterize the features of WBs. Commonly used spectral index methods include the Normalized Difference Water Index (NDWI) [16], Modified Normalized Difference Water Index (MNDWI) [17], and Automated Water Extraction Index (AWEIsh) [18]. Some scholars have combined multiple indices for water extraction. For example, Yang and Ke [19] employed a combination of four WB indices and Bayesian decision theory to extract lake WBs, and Feng et al. [4] utilized a combination of multiple indices to extract long time-series WBs in Hunan, China. However, this method faces challenges such as determining appropriate thresholds.

In recent years, image classification has proven to be a superior method for remote sensing land feature extraction. An increasing number of scholars are employing computer vision algorithms on remote sensing images for land feature recognition and extraction. For example, Li et al. [20] utilized the random forest algorithm to draw the annual surface water distribution map of Huizhou, China, from 1986 to 2020. Duan et al. [21] extracted the WB distribution in Wuhan, China, from Gaofen-1 data using a lightweight convolutional neural network (CNN). Zhang et al. [22] extracted WBs from SAR images using the cascaded fully convolutional network (CFCN) algorithm. However, the traditional computer vision algorithm is weak in global information perception. To solve this problem, Luo et al. [23] proposed gated multiscale ConvNet (GMNet) and applied it to surface water mapping on the Qinghai–Tibet Plateau. Zhang et al. [11,24] proposed the MF-SegFormer network based on the concept of multi-scale feature fusion to extract WBs in the Weihe River basin from Landsat 8 OLI images. Pan et al. [25] proposed a new boundary enhancement network, which effectively improved the ability to extract WB boundaries. The above three algorithms all use dilated convolutions with different dilation rates to sample input features in order to construct the relationship between local and global features and ensure the integrity of water surface extraction.

Although the above methods can perceive the global information well, they cannot locate the water boundary information effectively. In order to better locate the boundaries of WBs, Miao et al. [26] proposed a method of weighted edge loss values to assign more weights to water boundary pixels, making the network more focused on edge information. Some scholars are inspired by the Vision Transformer (ViT) and attempt to use self-attention mechanism to obtain boundary features. For example, Zhong et al. [27] proposed a transformer-based deep learning neural network, NT-Net, for lake WB extraction. However, the ViT-based network is too complex for binary classification tasks such as water extraction, and the transformer has weak sensing bias ability, which prevents it from identifying the detailed texture of water boundary.

Inspired by the above work, we proposed a network model called EDWNet, designed to extract WBs in complex scenes using Landsat 8 OLI images from the Weihe River basin. To address the complexity of the existing algorithm for binary classification tasks, we employ the ResNet network as the backbone. To tackle the issue of poor accuracy in boundary positioning, we propose the CFF module. This module utilizes cross-layer feature fusion to merge the multi-layer features output by the encoder. This effectively prevents the problem of adjacent features being similar, ensuring the diversity of the fused features. To overcome challenges in capturing detailed features, we introduce the GAM module, which interacts and integrates global features to obtain more comprehensive and rich WB features, simultaneously reducing information diffusion. Additionally, in order to solve the problem of weak global WB information perception ability of the existing model, we propose the DAM module, which can fuse multiple layers of important features, learn global features, and enhance the model’s WB feature extraction ability. Finally, an auxiliary head is added to optimize the learning process of the model. This study constructs EDWNet using an end-to-end strategy, effectively avoiding the problem of traditional CNN networks relying too much on local receptive fields.

Furthermore, previous artificial intelligence algorithm-based WB extraction studies primarily focused on the use of true color composite images [28,29], and did not fully utilize the rich band information of remote sensing images. While complex scenes encompass various types of WBs and ground objects. How to use artificial intelligence algorithms to accurately extract WBs in complex scenes is a problem. Therefore, we used the Deep SHAP algorithm [30] to analyze the feature importance of bands 2 to 7 in Landsat 8 OLI images. We sorted the importance of each band and selected the most suitable band combination (RGB 763) for the algorithm to extract WBs. To verify the superiority of our selected band combination, we compared it with four commonly used band combinations. We also compared our proposed EDWNet with other mainstream deep learning algorithms. The results show that EDWNet has more advantages than other algorithms in WB extraction. In addition, we utilized the EDWNet model to extract long time-series WBs in the Weihe River Basin from 2013 to 2021, and quantitatively analyzed the changes in WB area.

2. Study Area and Data

We selected the Weihe River basin as our study area. The Weihe River is the largest first-order tributary of the Yellow River, spans 818 km and covers a total basin area of 135,000 km² (Figure 1a,b). The basin comprises various WBs, including reservoirs, lakes, small WBs, and artificial WBs, and it also covers diverse terrains such as cities, mountains, deserts, and loess plateaus. Therefore, the Weihe River Basin is a complex scene, and WB extraction is more difficult. The Weihe River Basin is located in the continental monsoon climate zone and is the transition zone between arid areas and humid areas. It is located in the fragile ecological environment zone of northwest China and is prone to natural disasters. Drought is the most frequent, severe, and harmful, and it has become the main factor restricting the national economy and agricultural development in the region [31].

Given these challenges, the Weihe River basin was selected for this study due to its unique combination of complexity and representativeness. Its varied environments make it an ideal testbed for evaluating the robustness of WB extraction models across diverse land cover types and climatic conditions. Moreover, the ecological vulnerability and importance of the Weihe River in the broader Yellow River system provide an opportunity for this study to contribute not only locally but also offer insights that could be applied to other similar basins in northwest China. In this context, we conducted a study on the spatiotemporal distribution and trends in surface WBs in the Weihe River over the past decade, aiming to provide guidance on mitigating or adapting to climate change, as well as insights for environmental protection and ecological construction in northwest China.

This study selected Landsat 8 OLI optical satellite images as the data source, which have been produced since 2013 [32]. Landsat 8 was chosen as the primary data source due to several key advantages it offers. First, Landsat 8 provides high temporal resolution, with images captured every 16 days. Second, it can provide a spatial resolution of up to 15 m, which is very suitable for large-scale WB extraction. Additionally, Landsat 8 offers rich multispectral data with multiple bands, which is particularly useful for distinguishing WBs from surrounding land features. The availability of freely accessible data also makes Landsat 8 a cost-effective choice for large-scale and long time-series studies. Compared to other satellite options, such as Sentinel-2 or MODIS, Landsat 8 strikes a balance between resolution, revisit frequency, and spectral richness, which makes it highly suitable for the complex scenes in the Weihe River Basin.

In addition, this study selected the Google Earth Engine (GEE) platform for data selection and preprocessing (https://earthengine.google.com/, (accessed on 1 October 2023)) [33]. We used surface reflectance (SR) images, which have already undergone atmospheric correction, provided by GEE. To ensure high-quality data, we selected images with less than 2% cloud cover for mosaicking. Additionally, we performed normalization to minimize brightness and contrast differences across images, enhancing consistency. Images from the wet season (1 May to 31 October 2016) were used to construct the training dataset. To address the issue of mixed pixels, we applied image pansharpening to increase the resolution from 30 m to 15 m. Previous studies have shown that the Gram–Schmidt algorithm provides the highest accuracy [34], so we employed this algorithm for pansharpening. The images before and after pansharpening areshown in Figure 1c,d, respectively. Finally, we clipped the data according to the vector boundaries of the study area.

3. Methods

3.1. Proposed Method

3.1.1. EDWNet

In order to solve the problem of incomplete extraction of WBs and incorrect division of fuzzy edges in complex scenes, we designed EDWNet to extract WBs in complex scenes. The network structure diagram is shown in Figure 2, which consists of an encoder–decoder architecture. Code will be available at https://github.com/tiany-zhang/EDWNet (accessed on 10 July 2024).

Considering that this is a binary classification task, an overly complex backbone can cause the model to have blurring and confusion problems when extracting details and texture information, resulting in missed mentions and discontinuous extraction. Therefore, the backbone in the encoder of this network uses the ResNet network. ResNet was chosen as the backbone for EDWNet due to its ability to mitigate the vanishing gradient problem, which allows for the construction of deeper networks that can extract complex features effectively. This is particularly important for the task of WB extraction from remote sensing images, where differentiating between water and similar features, such as urban structures or vegetation, requires capturing both local edge details and global contextual information. The residual connections in ResNet enable efficient learning without loss of feature richness, which is crucial for distinguishing intricate water boundaries. Compared to other backbones, like VGG or DenseNet, ResNet strikes an ideal balance between depth and computational efficiency, making it highly suitable for the large-scale and complex scenarios encountered in this study of the Weihe River Basin. In this study, we built three models of different complexity based on ResNet18, ResNet50, and ResNet101: EDWNet-S, EDWNet-M, and EDWNet-L.

EDWNet includes a Cross-layer Feature Fusion (CFF) module, Global Attention Mechanism (GAM) module [35], Deep Attention module (DAM), and uses an auxiliary head for auxiliary loss calculation [36]. Since there are differences in spectral and texture information between WBs and surrounding features, the CFF module can integrate these features and use semantic and detail information to eliminate the interference of non-water features around WBs, thereby enhancing the ability to extract edge information from WB. In addition, the CFF module can effectively prevent the problem of similar adjacent features and ensure the difference of fused features. The GAM module captures important features of WBs across both the channel and spatial dimensions. It reduces information diffusion, enhances the ability to capture contextual information, and obtains more comprehensive and detailed WB features. The DAM module is used to obtain global information from feature maps, which helps better describe local semantic information and effectively solves the problem of weak global perception ability in existing models. Finally, the auxiliary head is used for auxiliary loss calculation, which is beneficial to the optimization learning of the overall model.

Firstly, feature extraction is performed through the backbone in the encoder part. The decoder uses two CFF modules for cross-layer feature fusion on the middle layer features of the backbone and sends the fused results to two GAM modules for water feature enhancement. The highest-level semantic features output by backbone will be sent to the GAM module for feature capture. Then, the multi-layer features are upsampled and adjusted to the same size before being input into the DAM module to fuse important features obtained at different levels and suppress redundant features. Finally, the decoding head is utilized to classify each pixel in the image, while the auxiliary head calculates the auxiliary loss value.

Current image processing and feature extraction techniques face a series of challenges when performing high-level feature upsampling and multi-layer feature fusion. First, due to the weak association between high-level and low-level features, achieving effective information transfer is difficult, resulting in suboptimal preservation of image details in the fused features. Second, because of the similarity between features at different levels, high-level features may become overly similar to adjacent level features after upsampling, leading to increased redundant information and reducing the model’s sensitivity to image edges and textures. This is especially problematic for distinguishing foreground objects from background regions. Additionally, while low-level features can provide pixel-level information, they lack sufficient semantic information, making it challenging to accurately describe the overall characteristics of target objects.

To address these issues, and inspired by the feature fusion concepts from ICNet [37], we designed the CFF module. To further reduce computational load and parameter count, we utilized depthwise separable convolution. The proposed CFF module aims to automatically upsample high-level features and perform cross-layer fusion of multi-layer features, effectively preventing adjacent features from becoming overly similar. By combining pixel-level information from low-level features with semantic information from high-level features, the CFF module enhances the feature differentiation between WB and non-WB regions, improves the ability to extract edge information of WBs, and achieves pixel-level classification.

Figure 3 shows the structure diagram of the CFF module. In this module, the low-level feature maps $X_{\mathrm{low}}^{\mathrm{input}}\in (R^{{\displaystyle \frac{H}{4}}\times {\displaystyle \frac{W}{4}}\times C_2}, R^{{\displaystyle \frac{H}{8}}\times {\displaystyle \frac{W}{8}}\times C_3})$ are the outputs from Stage 1 and Stage 2 of the backbone, while the high-level feature maps $X_{\mathrm{high}}^{\mathrm{input}}\in (R^{{\displaystyle \frac{H}{8}}\times {\displaystyle \frac{W}{8}}\times C_4}, R^{{\displaystyle \frac{H}{8}}\times {\displaystyle \frac{W}{8}}\times C_5})$ are the outputs from Stage 3 and Stage 4 of the backbone. The CFF module automatically adjusts the size of the high-level feature map in the backbone to match the size of the low-level feature map through upsampling. It then employs cross-layer methods for feature fusion, ensuring the difference of the fused features. The specific equations are shown in Equations (1)–(4):

(1)${\mathrm{X}}_{\mathrm{high}}={\mathrm{Upsample}(\mathrm{X}}_{\mathrm{high}}^{\mathrm{input}},\mathrm{N})$

(2)${\mathrm{X}}_{\mathrm{high}}^{\mathrm{output}}=\left\{{\mathsf{\delta }\left\{\mathsf{\beta }\right\{\mathrm{f}}_{\mathrm{k}=3}^{\mathrm{d}=2}\right[\mathrm{MaxPool}\left({\mathrm{X}}_{\mathrm{high}}\right)+\mathrm{AvgPool}({\mathrm{X}}_{\mathrm{high}}\left)\right]\left\}\right\}\}·{\mathrm{X}}_{\mathrm{high}}$

(3)${\mathrm{X}}_{\mathrm{CFF}}={\mathsf{\delta }\left\{\mathsf{\beta }\right\{\mathrm{f}}_{\mathrm{k}=5}^{\mathrm{d}=2}{\left[\mathrm{Concat}\right(\mathrm{X}}_{\mathrm{low}}^{\mathrm{output}}{,\mathrm{X}}_{\mathrm{high}}^{\mathrm{output}}\left)\right]\left\}\right\}$

(4)${\mathrm{X}}_{\mathrm{CFF}}^{\mathrm{output}}={\mathrm{X}}_{\mathrm{CFF}}\otimes {\mathrm{X}}_{\mathrm{low}}^{\mathrm{output}}+(1-{\mathrm{X}}_{\mathrm{CFF}})\otimes {\mathrm{X}}_{\mathrm{high}}^{\mathrm{output}}$

where Upsample represents the upsampling operation, and N represents the scale factor. ${\mathrm{X}}_{\mathrm{high}}^{\mathrm{input}}$ represents the high-level input feature map, f represents the depth-wise separable convolution (DSC), k represents the convolution kernel size, d represents dilation rate, $\otimes$ denotes the Hadamard product, represents the dot product, β signifies a batch normalization (BN) operation, and δ denotes the rectified linear unit (ReLU) activation function.

Existing models often struggle with accurately capturing global spatial information, especially when handling complex scenes in remote sensing images. A primary challenge lies in the limited ability to model channel dependencies effectively, which restricts the model’s capacity to understand the global context and inter-channel relationships. While each feature map channel corresponds to specific semantic responses, these responses are interdependent and influence each other. SENet’s Squeeze-and-Excitation (SE) block [38], leverages channel-wise dependencies to refine semantic feature representation, improving the model’s focus on essential features. However, SE’s reliance on dimensionality reduction can sometimes hinder the capture of detailed channel interactions.

To address this, ECANet proposed the Efficient Channel Attention (ECA) block [39], which avoids dimensionality reduction to maintain robust channel attention. ECA demonstrated that eliminating dimensionality reduction can enhance the model’s ability to capture nuanced channel dependencies, thus improving the overall effectiveness of attention mechanisms in feature extraction.

Inspired by ECA’s approach, we designed the DAM (Figure 4). With this module, various levels of feature information are fused, followed by Global Average Pooling (GAP) to gather global information for each channel. Then, the features are extracted through a one-dimensional convolutional group with multiple sampling rates to capture dependency relationships between channels. Finally, attention weights are reapplied to the original feature map to obtain the weighted feature map. This module effectively addresses the issue of weak global information perception in existing models, ensuring that the model can more accurately capture the spatial information of WBs in remote sensing images.

Current models often struggle to balance information capture across both channel and spatial dimensions, leading to insufficient contextual information, particularly when dealing with complex backgrounds and detailed WB features. Additionally, excessive information diffusion across different feature layers limits the model’s ability to retain key features across multiple dimensions, impairing the capture and representation of three-dimensional features for WB targets. To address these issues, we introduced the GAM module.

The GAM module structure diagram is shown in Figure 5, which is a global attention mechanism that sequentially connects channel (Equation (5)) and spatial attention (Equation (6)) sub-modules. The channel attention sub-module can enhance channel and spatial dependencies and retain water information in three dimensions. Further incorporating the spatial attention submodule enhances the model’s extraction of WB features in space. The specific formula is as follows:

(5)${\mathrm{F}}_2={\mathrm{M}}_{\mathrm{C}}{(\mathrm{F}}_1)\otimes {\mathrm{F}}_1$

(6)${\mathrm{F}}_3={\mathrm{M}}_{\mathrm{S}}{(\mathrm{F}}_2)\otimes {\mathrm{F}}_2$

where M_c and M_s are the channel attention sub-module and the spatial attention sub-module, respectively.

The auxiliary head part uses the FCN Head, which serves to add an auxiliary loss. The extraction results it produces have no impact on the model and only play a role in training. Adding an auxiliary head helps optimize the learning process of the model. The main branch in the network has the highest weight, and we can balance it by increasing the weight of the auxiliary head. Experiments have shown that setting the weight of the auxiliary head to 0.4 has the best performance [36], so we have set the weight of the auxiliary head to 0.4.

The loss function of EDWNet is Loss_seg (Equation (7)), which consists of focal loss [40] and dice loss [41] with an equal weight ratio of 1:1. This ratio was chosen to balance between addressing class imbalance and achieving precise boundary accuracy, ensuring the overall performance of the model. Additionally, studies by Zhong et al. [27] and Li et al. [42] have demonstrated that a 1:1 ratio between these two losses yields optimal results for remote sensing image segmentation tasks. The formula is as follows:

(7)${\mathrm{Loss}}_{\mathrm{seg}}={\mathrm{Loss}}_{\mathrm{Focal}}+{\mathrm{Loss}}_{\mathrm{Dice}}$

where ${\mathrm{Loss}}_{\mathrm{Focal}}$ denotes the focal loss, and the ${\mathrm{Loss}}_{\mathrm{Dice}}$ denotes the dice coefficient loss.

Due to the challenge of extreme imbalance between positive and negative samples in WB segmentation, focal loss can help the model pay more attention to these fewer but important positive samples by applying a higher weight to WB areas that are difficult to classify. In this way, the model will not be dominated by negative samples, thereby improving the accuracy of detecting WBs. The specific equation for focal loss is shown in Equation (8):

(8)${\mathrm{Loss}}_{\mathrm{Focal}}{(\mathrm{p}}_{\mathrm{t}})=-{\mathsf{\alpha }}_{\mathrm{t}}{(1-{\mathrm{p}}_{\mathrm{t}})}^{\mathsf{\gamma }}{\log (\mathrm{p}}_{\mathrm{t}})$

where ${\mathrm{p}}_{\mathrm{t}}$ represents the probability of the model predicting water or non-water, ${\mathsf{\alpha }}_{\mathrm{t}}$ is used to balance the number of positive and negative samples, and γ is used to address the issue of imbalanced samples.

Dice loss optimizes the model by calculating the similarity between the predicted results and the real labels. It prioritizes global performance over the classification accuracy of individual pixels. This global consideration enhances the model’s generalization ability across different scenes, especially in identifying key areas such as WB boundaries. The specific equation for Dice Loss is shown in Equation (9):

(9)${\mathrm{Loss}}_{\mathrm{Dice}}(\mathrm{P},\mathrm{G})=1-{\displaystyle \frac{2{\displaystyle \sum _{\mathrm{h}=1}^{\mathrm{H}}{\displaystyle \sum _{\mathrm{w}=1}^{\mathrm{W}}{|\mathrm{G}}_{\mathrm{h},\mathrm{w}}\cap {\mathrm{P}}_{\mathrm{h},\mathrm{w}}|}}}{{\displaystyle \sum _{\mathrm{h}=1}^{\mathrm{H}}{\displaystyle \sum _{\mathrm{w}=1}^{\mathrm{W}}{\left(\right|\mathrm{G}}_{\mathrm{h},\mathrm{w}}|\cap {|\mathrm{P}}_{\mathrm{h},\mathrm{w}}|})}}}$

where P represents the prediction result, and G is the true value in the label image. W and H denote the width and height of the image, respectively.

3.1.2. Innovations of the EDWNet

Compared to existing segmentation methods, EDWNet introduces a unique combination of modules designed specifically for handling the challenges associated with extracting WBs in complex scenes. The CFF module integrates both low-level and high-level features, enhancing boundary precision—essential in areas where WBs are interwoven with urban and vegetation elements. GAM provides a global attention mechanism to focus on relevant features and reduce the impact of distracting background information. The DAM ensures effective global feature fusion, improving both local and global feature extraction, which is crucial for identifying complex WB structures.

Additionally, EDWNet uses ResNet as its backbone to balance model complexity and computational efficiency. Unlike DenseNet or transformer-based models, the ResNet-based approach effectively extracts features while controlling model complexity, ensuring it remains feasible for large-scale scenarios.

3.2. Contrastive Methods

3.2.1. FCN

In 2015, Long et al. [43] introduced Fully Convolutional Networks (FCN) for semantic segmentation, a framework that modifies traditional CNNs by converting fully connected layers into convolutional layers. This approach enables per-pixel classification by using deconvolution (or upsampling) layers to resize feature maps back to the original image dimensions. Unlike standard CNNs, which are limited to fixed-size inputs for image-level tasks, FCN allows segmentation of arbitrarily sized images, producing output maps that match input dimensions.

The FCN architecture progressively downscales feature maps to capture high-level features and then upsamples them, using skip connections to combine high- and low-resolution features, thereby enhancing segmentation accuracy. Applied to the PASCAL VOC dataset, FCN achieved a 20% improvement in mean intersection over union (mIoU) over previous state-of-the-art methods like SDS and R-CNN, with greater segmentation efficiency.

3.2.2. PSPNet

Pyramid Scene Parsing Network (PSPNet), proposed by Zhao et al. [36] in 2017, is a deep learning model for semantic segmentation that introduces a Pyramid Pooling Module (PPM) to improve segmentation accuracy across objects and scenes at different scales. PSPNet is well-suited for complex, panoramic segmentation tasks and performs effectively in diverse applications, such as urban and natural environments.

To address traditional FCNs’ limitations in capturing global context, PSPNet’s PPM performs feature fusion at four different levels of granularity. The coarsest level applies global average pooling to generate a single-cell output, while finer levels divide the feature map into multiple sub-regions, producing multi-cell outputs. After upsampling these feature maps to a uniform resolution, PSPNet concatenates them to form a globally aware feature representation.

3.2.3. Deeplabv3+

DeepLabV3+, introduced by Chen et al. [44], builds on the DeepLab series with an enhanced encoder–decoder architecture to improve segmentation accuracy in complex scenes. By combining an encoder–decoder structure with a pyramid pooling module, DeepLabV3+ refines object boundaries and captures rich contextual information at multiple scales. Additionally, it utilizes the Xception architecture and depthwise separable convolutions in both the Atrous Spatial Pyramid Pooling (ASPP) and decoder modules, optimizing performance while reducing computational cost. On the PASCAL VOC2012 dataset, DeepLabV3+ achieved an mIoU of 87.8%, with test dataset performance reaching up to 89% without post-processing.

3.2.4. Segmenter

Segmenter, proposed by Strudel et al. [45], is a ViT-based semantic segmentation model that uses the transformers’ self-attention mechanism for precise pixel-level segmentation. Segmenter divides the input image into patches, mapping them into a linear embedding sequence processed by a transformer encoder with L layers, each containing Multi-Head Self-Attention and Feed-Forward Network sub-layers. This structure enables each image patch to capture global context, generating rich semantic features. The encoder output–a globally informed embedding sequence–then enters the Mask Transformer, which decodes it along with class embeddings. Finally, the features are upsampled, and an Argmax operation assigns class labels to each pixel, producing the final segmentation map.

3.2.5. SETR

Segmentation Transformer (SETR) [46] is a semantic segmentation model based on the ViT architecture, which replaces traditional CNN encoders with a pure transformer structure. Unlike CNN-based encoders that rely on repeated downsampling to increase the receptive field—often at the expense of capturing long-range dependencies—SETR leverages the self-attention mechanism of the transformers to effectively capture global context within images. For the decoding stage, SETR introduces three strategies: naive upsampling, progressive upsampling, and multi-level feature fusion. This transformer-based approach offers an alternative to CNN-based models like FCN and U-Net, with experimental results indicating that SETR learns semantic features that outperform those of traditional CNN architectures. In this study, we used the PUP structure as the decoder, as it showed the best performance in the original paper.

3.2.6. SegNext

SegNext [47] is a semantic segmentation model based on the ViT that combines the strengths of CNNs and the transformers to enhance segmentation in complex scenes. The model’s encoder integrates an efficient attention mechanism using multi-scale convolutional features, which are aggregated through element-wise multiplication to capture contextual information and long-range dependencies. By employing lightweight strip convolutions and 1 × 1 convolutions, SegNext reduces computational costs while maintaining high performance. Demonstrating strong results on multiple benchmarks, SegNext achieved a 2.0% improvement in mIoU on the ADE20K dataset, with performance on par with or better than state-of-the-art methods at a reduced computational load.

4. Experimental Setup

In this study, Landsat 8 OLI images from 2013 with a resolution of 15 m were selected for the Weihe River Basin to create the dataset. The corresponding ground truth values were obtained through visual interpretation. Subsequently, the images and their corresponding labels were cropped into 256 × 256 small images, obtaining a total of 16,610 pairs of sample data. We divided all the data into training, validation, and test datasets in a ratio of 6:2:2. This division ratio was chosen based on common empirical practices in deep learning. Dividing the dataset in this way allows for a balanced allocation, where the majority of the data are dedicated to training, ensuring robust model learning. A 20% allocation each for validation and testing provides sufficient data for model tuning and unbiased evaluation, enhancing the model’s reliability and generalization capability on unseen data. Specifically, 9966 pairs of samples were randomly selected for the training dataset, 3322 pairs for the validation dataset, and the remaining 3322 pairs for the test dataset. In order to reduce model overfitting and improve accuracy, we employed a series of methods to increase the training datasets, including flipping, rotating, cropping, deforming, and scaling. Ultimately, the training sample set was expanded to 24,000 pairs of samples.

In order to determine the most suitable band combination for water extraction in Landsat 8 OLI images, we analyzed bands 2 to 7 based on their feature importance. Subsequently, we selected the three most crucial bands for combination to construct the datasets. In addition, to verify whether this combination is more conducive to the algorithm’s extraction of WBs, we also selected several commonly used band combinations for comparative experiments. To evaluate the performance of our proposed algorithm, we selected multiple mainstream semantic segmentation algorithms for comparative experiments.

The experiments were conducted under the MMSegmentation framework [48]. The computer configuration is shown in Table 1. The batch size was set to 28 due to GPU memory limitations. The number of epochs was chosen as 100 to ensure sufficient training iterations for the model to converge effectively without overfitting. During preliminary experiments, it was observed that the model’s performance plateaued around 100 epochs, indicating that this number provides a good balance between computational efficiency and model accuracy. Meanwhile, we set the optimizer to AdamW [49] with an initial learning rate of 6 × 10⁻⁵ and a weight decay of 0.01, and used a combined loss function, summing Focal Loss and Dice Loss with a 1:1 weight. All comparative algorithms were trained and evaluated under the same software and hardware conditions.

Finally, we employed four performance indicators—Dice, Precision, Recall, and mIoU—to evaluate the accuracy of different extraction results. The expressions of these indicators are presented in Equations (10), (11), (12) and (13), respectively:

(10)$\mathrm{Dice}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{n=1}^N\frac{2\times TP}{FP+FN+2\times TP}\times 100\%}$

(11)$\mathrm{Precision}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{\mathrm{n}=1}^N\frac{TP}{TP+FP}\times 100\%}$

(12)$\mathrm{Recall}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{\mathrm{n}=1}^N\frac{TP}{TP+FN}}\times 100\%$

(13)$mIoU={\displaystyle \frac{1}{C}}{\displaystyle \sum _{i=1}^C\frac{TP}{TP+FP+FN}}\times 100\%$

where TP represents True Positive, FP stands for False Positive, FN denotes False Negative, C represents the number of classes, and N indicates the total number of samples.

5. Results

5.1. Data Selection

Previous research has proven that the false color composite image composed of bands 5, 6, and 4 (RGB 564) is more suitable for WB extraction models [11]. However, not all commonly used bands were analyzed, so we consider using the Deep SHAP algorithm to evaluate the importance of commonly used bands (bands 2–7) in Landsat8 OLI images. As shown in Figure 6, it can be seen that for both background and WB extraction, band7 has the highest importance, followed by band3, band6, band2, band5, and band4. Their SHAP values for WB extraction are 1.11 × 10⁻², 6.85 × 10⁻³, 3.91 × 10⁻³, 1.71 × 10⁻³, 1.52 × 10⁻³, and 9.5 × 10⁻⁴, respectively. It can be seen that the three most important bands are band 3, band 6, and band 7. This may be because the WB absorbs most of the band 4 (red) and band 3 (green), but the reflectivity in the green band is slightly higher than red light, so it can better distinguish WBs and vegetation. And the reflectance of water in the Short Wavelength Infra-Red (SWIR) 1 and 2 bands is very low. Therefore, we chose band3, band6, and band7 data to construct a false color composite image (RGB 763).

To determine whether the false color composite image (RGB 763) we created offers advantages in WB extraction over other band combinations, we conducted comparative experiments with five commonly used band combinations (https://geodata.pku.edu.cn/index.php?c=content&a=show&id=536, accessed on 1 February 2024). The experimental configurations are detailed in Table 2. All experiments used the EDWNet-M algorithm with 100 epochs. The results show that the band combination with the highest loss value on the validation dataset was RGB 432, with a minimum loss value of 0.129, followed by RGB 543, RGB 652, RGB 654, and RGB 632 (Figure 7). The band combination with the lowest loss value was RGB 763, with a loss of 0.099, representing a reduction of 1% to 30% compared to the other combinations. This indicates that the dataset constructed using RGB 763 is more effective in aiding the model’s recognition of WBs. Therefore, in all subsequent experiments, we used the images constructed by RGB 763 as the data source.

5.2. Comparison with Various Methods

In order to analyze the performance of different semantic segmentation models in WB extraction, we selected FCN, PSPNet, DeeplabV3+, Segmenter, SETR, and SegNext as comparative algorithms. We calculated the Dice, precision, recall, and mIoU (Table 3) of each algorithm on the test datasets by using the confusion matrices. We found that EDWNet-L had the highest Dice, Recall, and mIoU values (87.3%, 85.9%, 88.7%), followed by EDWNet-M (87.0%, 85.2%, 88.4%). EDWNet-S is higher than the four comparative algorithms in terms of Dice and mIoU, but slightly lower in Recall compared to Deeplab V3+. The Precision of EDWNet-L and EDWNet-M is the same (88.8%), followed by EDWNet-S (87.6%). It is evident that the performance of our proposed EDWNet significantly outperforms the other comparative algorithms. Compared with the four comparative algorithms, EDWNet-L exhibited improvements ranging from 4.8% to 41.3% in Dice, 6.5% to 39.6% in Precision, 3.4% to 42.7% in Recall, and 3.6% to 22.7% in mIoU.

In addition, we also calculated the accuracy of all models on the overall graph. The results are shown in Table 4. The descending order of Dice and mIoU is as follows: EDWNet-L, EDWNet-M, EDWNet-S, DeeplabV3+, SETR, FCN, SegNext, PSPNet, Segmenter. For Precision, the descending order is as follows: EDWNet-M, EDWNet-L, EDWNet-S, DeeplabV3+, SETR, FCN, PSPNet, SegNext, Segmenter. The descending order of Recall is as follows: EDWNet-M, EDWNet-L, DeeplabV3+, EDWNet-S, SETR, FCN, SegNext, PSPNet, Segmenter.

As shown in Table 3, although EDWNet-L and EDWNet-M achieved the highest performance among the models, they also have relatively high parameter counts (206.49 M and 187.5 M), second only to SETR (309.86 M). On the other hand, EDWNet-S has only 49.47 M parameters, and its accuracy is only slightly lower than EDWNet-L and EDWNet-M, achieving an effective balance between performance and parameter size, providing high accuracy with lower computational requirements.

In addition, we conducted additional experiments using a dataset constructed with RGB 564. The evaluation results are shown in Table 5, where EDWNet once again outperformed other methods on the test dataset. EDWNet-L achieved the highest performance across all metrics, with an mIoU of 88.3%, followed by EDWNet-M (87.8%), EDWNet-S (86.7%), DeeplabV3+ (85.1%), SETR (83.8%), FCN (81.7%), SegNext (80.6%), PSPNet (77.6%), and Segmenter (70.8%). Furthermore, compared to Table 3, it can be seen that, except for SETR, which experienced a 0.2% decrease in accuracy, all other algorithms achieved higher accuracy when using models constructed with the RGB 763 dataset compared to those constructed with the RGB 564 dataset.

Based on the above experimental results, we can see that the overall accuracy of EDWNet-M and EDWNet-L has an advantage, and the accuracy between the two is not much different. Following closely is EDWNet-S. Although its accuracy is reduced, the training time can be greatly shortened. The accuracy of the three is greatly improved compared to the six comparison algorithms, which shows that EDWNet is more suitable for WB extraction. Furthermore, based on the results of different algorithms on the RGB 564 dataset, it can be seen that EDWNet consistently outperforms other algorithms in extraction performance, and that RGB 763 further enhances the model’s performance compared to other band combinations.

5.3. Ablation Experiment

To further understand the impact of each module in EDWNet, as shown in Table 6, four different comparative experiments were established in this study, and accuracy validation was performed on the test dataset. The baseline model is EDWNet-M, and P2-P4 refer to the baseline model with DAM, GAM, and CFF removed, respectively.

From the results, it can be seen that. The P1 model has the highest accuracy, indicating that when three modules are added simultaneously, the model performance is optimal, and the number of parameters in the model is also the highest. When removing the three modules separately, the accuracy decreased, with P3 showing the most significant decrease, followed by P2, and finally P4. This indicates that GAM contributes the most to model performance, followed by DAM and finally CFF. However, it can also be observed that adding GAM leads to a significant increase in the number of model parameters.

6. Discussions

6.1. Typical Area Extraction Comparison

In order to validate the model’s effectiveness in extracting WBs, we conducted a qualitative analysis of all algorithms’ results. As shown in Figure 8, each column represents a different scene, with a total selection of seven diverse scenes. The first row displays a pseudo-color composite image of a real scene, the second row depicts the labeled image of each scene, and the third to eighth rows exhibit the prediction results of the comparison algorithms. The last three rows are the extraction results of our proposed EDWNet under different backbones.

Yellow dotted lines indicate areas where WBs are incorrectly classified as the background, while red dotted lines indicate areas where the background is incorrectly classified as WBs. The results show that each model has varying degrees of misclassification, and EDWNet based on three different backbones has less misclassification. In these scenes, as shown in Figure 8a, PSPNet, Segmenter, and SETR are unable to distinguish lakes from ponds with close intervals, leading to the problem of over-extraction. As shown in Figure 8b,c, FCN and SegNext may perform poorly on small WBs in urban scenes due to the influence of urban background, but they perform better on small WBs in mountains. PSPNet and Segmenter perform poorly on small WBs in both scenes. In the wide river scene, except for EDWNet-L and EDWNet-M, other algorithms have defects in detail. In addition, each algorithm performs well on the lake main body and mountain shadow scenes.

In summary, it is evident that EDWNet, based on three different backbones, demonstrates superior extraction performance compared to the comparative algorithms. EDWNet-M and EDWNet-L show similar overall performance, both better than EDWNet-S. This indicates that our proposed EDWNet has obvious advantages for extracting WBs in complex scenes.

6.2. Comparison of the Weihe River’s Mainstream

To further study the extraction performance of different methods, we intercepted the mainstream in the “Xi’an-Xianyang” section. As shown in Figure 9, it shows the spatial distribution of the main stream section of the Weihe River Basin in the “Xi’an-Xianyang” section. The green rectangular dashed outline indicates the focal area, which is located downstream of the Weihe River Basin. We constructed a 100 m buffer zone based on the main stream of the Weihe River, ensuring that the analysis area focused only on the river’s main channel. Using this buffer zone, we clipped the results generated by all algorithms, as well as the label data, to standardize the analysis area and avoid interference from unrelated surrounding regions.

We analyzed the river width results extracted by various methods and evaluated the accuracy using R² (Equation (17)) and RMSE (Equation (18)). Equations (14)–(16) provide the formulas for calculating river width:

(14)${\mathrm{N}}_{\mathrm{i}}=\mathrm{GETNUM}\left(\mathrm{Lat}\mathrm{itude}\right[1]:\mathrm{Latitude}[\mathrm{n}\left]\right), \mathrm{i}\in \left(\mathrm{Longitude}\right[1],\mathrm{Longitude}[\mathrm{m}\left]\right)$

(15)$W_{actual,i}={\mathrm{N}}_{\mathrm{i}}\times \mathrm{R}$

(16)${\overline{W}}_{\mathrm{actual}}={\displaystyle \frac{1}{m}}{\displaystyle \sum _{i=\mathrm{Longitude}\left[1\right]}^{\mathrm{Longitude}\left[\mathrm{m}\right]}{W}_{\mathrm{actual},i}}$

(17)$R^2=1-{\displaystyle \frac{{\displaystyle {\sum }_{i=\mathrm{Longitude}\left[1\right]}^{\mathrm{Longitude}\left[\mathrm{m}\right]}{(W_{\mathrm{actual},i}-W_{\mathrm{pred},i})}^2}}{{\displaystyle {\sum }_{i=\mathrm{Longitude}\left[1\right]}^{\mathrm{Longitude}\left[\mathrm{m}\right]}{(W_{\mathrm{actual},i}-{\overline{W}}_{\mathrm{actual}})}^2}}}$

(18)$RMSE=\sqrt{{\displaystyle \frac{1}{m}}{\displaystyle \sum _{i=\mathrm{Longitude}[1]}^{\mathrm{Longitude}[\mathrm{m}]}{(W_{\mathrm{pred},i}-PreWidt{h}_i)}^2}}$

where n and m represent the number of grid points in the meridional and latitudinal directions, respectively, of the study area. GETNUM represents the number of meridional river grid points when the longitude is i. R is the resolution of the remote sensing image. ${\overline{W}}_{\mathrm{actual}}$ represents the average actual river width, W_pred,i and W_actual,i represent the predicted river width and the actual river width when the longitude is i, respectively.

Figure 10 shows the river width extracted by various methods at different longitudes. From Figure 10a, it can be seen that the Segmenter extraction results have a significant deviation from the labels, while EDWNet-M and EDWNet-L have the smallest deviation from the labels. However, due to the large number of line segments, in order to more intuitively see the difference between each method and the label, we calculated the difference between the predicted river width using each method and the label (Figure 10b). From the results, it can be seen that most of the values of EDWNet-M and EDWNet-L are close to 0, followed by EDWNet-S and Deeplab V3+, while the value of Segmenter differs the most from 0.

Figure 11 shows the scatter plots of river widths extracted by different methods and actual river widths. We can see that the accuracy is ranked in descending order as follows: EDWNet-M, EDWNet-L, EDWNet-S, Deeplab V3+, FCN, SETR, SegNext, Segmenter. The R² and RMSE are 0.961, 0.96, 0.948, 0.944, 0.936, 0.916, 0.9, 0.899, 0.735 and 22.765 m, 23.123 m, 26.019 m, 27.939 m, 29.376 m, 34.09 m, 40.378 m, 40.378 m, 82.317 m, respectively.

In summary, EDWNet performs better than other algorithms in river WB extraction, among which EDWNet-M performs best.

6.3. Long Time-Series WB Extraction in the Weihe River Basin Basin

We applied the EDWNet-M to WB extraction in the Weihe River Basin from 2013 to 2021. Firstly, we obtained Landsat 8 OLI images of the Weihe River Basin during the flood season from 2013 to 2021 (1 May to 30 October) through the GEE platform, and concatenated the images with the lowest cloud cover. Then we used our proposed model to perform long time-series WB extraction in the Weihe River Basin. It can be seen that the overall water system of the Weihe River Basin presents a pattern of more in the south and less in the north, with uneven distribution of WBs. In addition, the overall scale of the WBs is small, but it contains many types of WBs, and there are many small WBs, which also greatly tests the algorithm’s performance. Figure 12 shows the false color composite images (RGB: 763) and EDWNet extraction results of the Weihe River Basin from 2014, 2016, 2018, and 2020. It can be seen that EDWNet has better water extraction performance in complex scenes and has strong generalization ability.

We used random stratified sampling to evaluate the accuracy of the extraction results for each year. Specifically, we randomly selected 2000 points in areas with more WBs on each image and used high-resolution images from Google Earth Engine for auxiliary validation. The actual category of the selected points was determined through visual interpretation techniques. Subsequently, Dice and mIoU metrics were used to evaluate the accuracy of the extraction results for each year. As shown in Figure 13a, both Dice and mIoU for each year exceeded 84%. From 2013 to 2021, Dice and mIoU were relatively stable, with the highest Dice in 2016 (87.9%) and the highest mIoU in both 2014 and 2016 (88.1%).

In addition, we conducted long time-series statistics on the WB area of the Weihe River Basin during the wet season from 2013 to 2021 (Figure 13b). It can be seen that the overall trend shows a decrease followed by an increase. The WB area showed a declining trend from 2013 to 2017, and an ascending trend from 2017 to 2021. The minimum WB area in 2017 was 1184.99 km², and the maximum WB area in 2021 was 1376.69 km². Meanwhile, the yellow rectangle in Figure 12 highlights a local magnified view of a certain section of the Weihe River mainstream. It can be observed that in 2014 and 2020, the river channel was wider, and surface water flow was greater, whereas in 2016 and 2018, the river channel was narrower with less surface water flow. This may be one of the reasons for the overall trend of first decreasing and then increasing in WB from 2013 to 2021.

This may be related to complex climate change trends and extreme climate events. Deng et al. [50] found that the year with the most average annual high-temperature days in the Weihe River Basin from 1980 to 2020 was 2017. And the average high-temperature days gradually increased from 2013 to 2017, while they gradually decreased from 2017 to 2020 (Figure 14). High-temperature weather will affect the evaporation of WBs, which may lead to changes in WB area. This may be a factor affecting the WB area in the Weihe River Basin.

Furthermore, Zou et al. [3] observed that ENSO (El Niño–Southern Oscillation) events have a significant impact on rainfall and runoff in the Weihe River Basin. El Niño events tend to reduce rainfall and runoff, while La Niña events increase them. The NINO 3.4 index, which represents the average sea surface temperature anomalies in the central equatorial Pacific (5°N–5°S, 170°W–120°W), is a key indicator used to monitor the strength and phase of ENSO events, including El Niño and La Niña. A higher NINO 3.4 index value generally corresponds to an El Niño event, indicating warmer sea surface temperatures, while a lower value is associated with a La Niña event, indicating cooler conditions.

Figure 15 shows the most severe El Niño event on record, which occurred between 2014 and 2016. As illustrated in Figure 15, the NINO 3.4 index significantly increased during this period, reflecting a strong El Niño phenomenon. The sharp rise in the NINO 3.4 index during this time likely led to a reduction in rainfall and runoff in the Weihe River Basin, which in turn caused a gradual decline in the surface WB area of the Weihe River Basin starting from 2014. From 2017 to 2021, it was generally a La Niña-dominated period, especially in the latter half of 2017, as well as in 2020 and 2021. This likely led to increased rainfall and runoff in the Weihe River Basin, resulting in an overall upward trend in surface WB balance in the Weihe River Basin after 2017.

7. Conclusions and Future Perspectives

In this study, we proposed a new algorithm called EDWNet for extracting WBs from Landsat 8 OLI images. EDWNet is an encoder–decoder structure designed specifically for high-precision WB extraction in complex environments. EDWNet incorporates the CFF module to enhance the fusion of multi-layer features, the GAM to reduce information diffusion, and the DAM to capture global features more effectively. Additionally, we used the Deep SHAP algorithm to select RGB 763 and construct the dataset, and conducted quantitative evaluation using Dice, Precision, Recall, and mIoU. Finally, we utilized the established WB extraction model to extract and analyze WBs in the Weihe River Basin from 2013 to 2021.

The results show that using EDWNet-M to extract WB, the dataset constructed with RGB 763 has the lowest validation loss of 0.099 compared to other commonly used band combinations. This indicates that the false color composite image (RGB 763) is more suitable for WB extraction. Additionally, the accuracy of each algorithm on the test dataset is ranked from high to low as follows: EDWNet-L, EDWNet-M, EDWNet-S, Deeplab V3+, SETR, FCN, SegNext, PSPNet, and Segmenter. Furthermore, EDWNet also achieved the highest R² and the lowest RMSE in river width extraction for the Weihe River’s “Xi’an-Xianyang” section. Overall, the EDWNet model demonstrates the highest accuracy and best performance, outperforming other methods.

Overall, we present a new method for accurately extracting WBs from optical remote sensing images in complex environments. However, this model has only been applied to the Weihe River Basin. In the future, we will use more datasets to verify the performance of EDWNet. Moreover, due to the high cost of producing high-quality labels, we will also attempt to explore the construction of a semi-supervised model based on EDWNet. This approach could potentially reduce the reliance on extensive labeled data while maintaining or even enhancing the model’s performance in various scenes.

Author Contributions

Conceptualization, T.Z. and W.J.; methodology, T.Z. and C.Q.; validation, T.W. and Y.R.; formal analysis, T.Z. and Y.R.; investigation, Y.F. and Z.H.; resources, W.L. and L.J.; data curation, T.W. and Y.R.; writing—original draft preparation, T.Z. and W.J.; writing—review and editing, T.Z. and W.L.; visualization, T.Z. and C.Q.; supervision, W.L. and L.J.; funding acquisition, W.L. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.

Acknowledgments

The authors would like to thank three anonymous reviewers for their helpful comments and suggestions, which significantly improved the quality of our paper.

Conflicts of Interest

Author Yuan Fang was employed by the company Shaanxi Water Development Ecological Technology R&D Co., Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Footnotes

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.

Figures and Tables

Figure 1. Spatial location and scope of the Weihe River Basin study area in the People’s Republic of China: (a) is the administrative divisions of China, (b) is the true color Landsat 8 OLI image of the study area, (c) is the image before pansharpening in a randomly selected area, and (d) is the image after pansharpening in a randomly selected area.

Figure 2. EDWNet model structure.

Figure 3. CFF module structure.

Figure 4. DAM module structure.

Figure 5. GAM module structure.

Figure 6. SHAP values of bands 2 to 7 in Landsat 8 OLI images.

Figure 7. Validation loss of EDWNet in different band combination images.

Figure 8. Classification results of WBs using different methods: (a–c) the scenario with small WBs, (d,e) the scenario with a reservoir, (f) the scenario with a wide river channel, (g) the scenario with shadows of hills. The yellow dotted line indicates WBs misclassified as background, while the red dotted line indicates pixels misclassified as WBs.

Figure 9. Spatial distribution of the main stream of the “Xi’an-Xianyang” section in the Weihe River Basin.

Figure 10. Results of river width extraction using different methods in the Weihe River “Xi’an-Xianyang” section: (a) line graph of river width extracted using different methods at different longitudes, and (b) difference between river width extracted using different methods and true width.

Figure 10. Results of river width extraction using different methods in the Weihe River “Xi’an-Xianyang” section: (a) line graph of river width extracted using different methods at different longitudes, and (b) difference between river width extracted using different methods and true width.

Figure 11. Scatter plot of label river width and extracted river width in different methods.

Figure 12. Extraction maps of the Weihe River Basin in 2014, 2016, 2018, and 2020. The left side are the original images, and the right side are the WB extraction results. The yellow color represents the background, the blue color represents the extracted WB. The area inside the yellow rectangle is a local magnified view of a certain section of the Weihe River mainstream.

Figure 13. Long time-series WB extraction results in the Weihe River Basin from 2013 to 2021: (a) WB extraction accuracy and (b) WB area changes.

Figure 14. Average high temperature days in the Weihe River Basin from 2013 to 2020.

Figure 15. NINO 3.4 index from 2013 to 2021.

Table 1

Platform configuration.

CPU	GPU	Internal Memory	System
Inter i9-13900k	RTX 4090	24 GB	Ubuntu

Table 2

Design of experiments.

Method	Backbone	Band Combinations
EDWNet-M	ResNet-50	RGB 763
		RGB 432
		RGB 543
		RGB 652
		RGB 632
		RGB 564

Table 3

Table of extraction results in accuracy evaluation for different methods on the test dataset.

Method	Backbone	Dice	Precision	Recall	mIoU	Params/M (106)
FCN	ResNet-50	79.8%	81.8%	78.0%	83.2%	49.49
PSPNet	ResNet-50	73.6%	77.1%	70.3%	79.0%	48.97
DeeplabV3+	ResNet-50	83.3%	83.4%	83.1%	85.6%	43.58
Segmenter	ViT-Base	61.8%	63.6%	60.2%	72.3%	117.74
SETR	ViT-Base	80.5%	82.3%	78.8%	83.6%	309.86
SegNext	MSCAN	76.9%	78.5%	75.3%	81.1%	48.77
EDWNet-S	ResNet-18	85.2%	87.6%	82.9%	87.1%	49.47
EDWNet-M	ResNet-50	87.0%	88.8%	85.2%	88.4%	187.5
EDWNet-L	ResNet-101	87.3%	88.8%	85.9%	88.7%	206.49

Table 4

Table of extraction results in accuracy evaluation for different methods on the whole study area.

Method	Backbone	Dice	Precision	Recall	mIoU
FCN	ResNet-50	78.8%	80.4%	77.2%	82.4%
PSPNet	ResNet-50	73.2%	76.7%	70.0%	78.8%
DeeplabV3+	ResNet-50	83.0%	83.1%	83.0%	85.5%
Segmenter	ViT-Base	57.1%	59.4%	54.9%	69.9%
SETR	ViT-Base	80.0%	82.0%	78.1%	83.3%
SegNext	MSCAN	74.1%	75.5%	72.8%	79.4%
EDWNet-S	ResNet-18	83.8%	85.5%	82.1%	86.0%
EDWNet-M	ResNet-50	85.8%	87.2%	87.2%	87.5%
EDWNet-L	ResNet-101	86.0%	86.7%	86.7%	87.7%

Table 5

Table of accuracy evaluation on the test dataset using different methods on the RGB 564 dataset.

Method	Backbone	Dice	Precision	Recall	mIoU
FCN	ResNet-50	78.1%	79.4%	76.5%	81.7%
PSPNet	ResNet-50	71.9%	75.7%	68.4%	77.6%
DeeplabV3+	ResNet-50	82.6%	82.7%	82.5%	85.1%
Segmenter	ViT-Base	58.8%	60.5%	57.3%	70.8%
SETR	ViT-Base	80.7%	81.2%	80.1%	83.8%
SegNext	MSCAN	75.4%	76.9%	74.1%	80.6%
EDWNet-S	ResNet-18	84.7%	87.2%	82.4%	86.7%
EDWNet-M	ResNet-50	86.7%	88.2%	84.6%	87.8%
EDWNet-L	ResNet-101	86.9%	88.5%	85.4%	88.3%

Table 6

Results of ablation experiment.

Project	Model	mIoU	Params/M (106)
P1	Baseline	87.8%	187.5
P2	Baseline-(DAM)	87.0%	179.53
P3	Baseline-(GAM)	86.9%	79.35
P4	Baseline-(CFF)	87.2%	186.09

References

1. Hou, X.; Feng, L.; Tang, J.; Song, X.-P.; Liu, J.; Zhang, Y.; Wang, J.; Xu, Y.; Dai, Y.; Zheng, Y. et al. Anthropogenic transformation of Yangtze Plain freshwater lakes: Patterns, drivers and impacts. Remote Sens. Environ.; 2020; 248, 111998. [DOI: https://dx.doi.org/10.1016/j.rse.2020.111998]

2. Keller, P.S.; Catalan, N.; von Schiller, D.; Grossart, H.-P.; Koschorreck, M.; Obrador, B.; Frassl, M.A.; Karakaya, N.; Barros, N.; Howitt, J.A. et al. Global CO₂ emissions from dry inland waters share common drivers across ecosystems. Nat. Commun.; 2020; 11, 2126. [DOI: https://dx.doi.org/10.1038/s41467-020-15929-y]

3. Zou, Z.; Xiao, X.; Dong, J.; Qin, Y.; Doughty, R.B.; Menarguez, M.A.; Zhang, G.; Wang, J. Divergent trends of open-surface water body area in the contiguous United States from 1984 to 2016. Proc. Natl. Acad. Sci. USA; 2018; 115, pp. 3810-3815. [DOI: https://dx.doi.org/10.1073/pnas.1719275115]

4. Feng, S.; Liu, S.; Zhou, G.; Gao, C.; Sheng, D.; Yan, W.; Wu, Y.; Gao, H.; Jia, J.; Wang, Z. et al. Long-term dense Landsat observations reveal detailed waterbody dynamics and temporal changes of the size-abundance relationship. J. Hydrol. Reg. Stud.; 2022; 41, 101111. [DOI: https://dx.doi.org/10.1016/j.ejrh.2022.101111]

5. Bui, T.A.; Lee, P.J. SWRNet: A deep learning approach for Small surface Water area Recognition onboard satellite. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2023; 16, pp. 10369-10380. [DOI: https://dx.doi.org/10.1109/JSTARS.2023.3328118]

6. Voigt, S.; Kemper, T.; Riedlinger, T.; Kiefl, R.; Scholte, K.; Mehl, H. Satellite image analysis for disaster and crisis-management support. IEEE Trans. Geosci. Remote Sens.; 2007; 45, pp. 1520-1528. [DOI: https://dx.doi.org/10.1109/TGRS.2007.895830]

7. Zhang, J.; Xing, M.; Sun, G.C.; Chen, J.; Li, M.; Hu, Y.; Bao, Z. Water body detection in high-resolution SAR images with cascaded fully-convolutional network and variable focal loss. IEEE Trans. Geosci. Remote Sens.; 2020; 59, pp. 316-332. [DOI: https://dx.doi.org/10.1109/TGRS.2020.2999405]

8. Su, H.; Zhang, T.; Lin, M.; Lu, W.; Yan, X.-H. Predicting subsurface thermohaline structure from remote sensing data based on long short-term memory neural networks. Remote Sens. Environ.; 2021; 260, 112465. [DOI: https://dx.doi.org/10.1016/j.rse.2021.112465]

9. Su, H.; Lu, W.; Wang, A.; Zhang, T. AI-Based Subsurface Thermohaline Structure Retrieval from Remote Sensing Observations. Artificial Intelligence Oceanography; Springer Nature: Singapore, 2023; pp. 105-123.

10. Vörösmarty, C.J.; McIntyre, P.B.; Gessner, M.O.; Dudgeon, D.; Prusevich, A.; Green, P.; Glidden, S.; Bunn, S.E.; Sullivan, C.A.; Liermann, C.R. et al. Global threats to human water security and river biodiversity. Nature; 2010; 467, pp. 555-561. [DOI: https://dx.doi.org/10.1038/nature09440]

11. Zhang, T.; Qin, C.; Li, W.; Mao, X.; Zhao, L.; Hou, B.; Jiao, L. Water Body Extraction of the Weihe River Basin Based on MF-SegFormer Applied to Landsat8 OLI Data. Remote Sens.; 2023; 15, 4697. [DOI: https://dx.doi.org/10.3390/rs15194697]

12. Wei, Z.; Jia, K.; Liu, P.; Jia, X.; Xie, Y.; Jiang, Z. Large-scale river mapping using contrastive learning and multi-source satellite imagery. Remote Sens.; 2021; 13, 2893. [DOI: https://dx.doi.org/10.3390/rs13152893]

13. Tang, H.; Lu, S.; Ali Baig, M.H.; Li, M.; Fang, C.; Wang, Y. Large-scale surface water mapping based on Landsat and Sentinel-1 images. Water; 2022; 14, 1454. [DOI: https://dx.doi.org/10.3390/w14091454]

14. Klein, I.; Dietz, A.J.; Gessner, U.; Galayeva, A.; Myrzakhmetov, A.; Kuenzer, C. Evaluation of seasonal water body extents in Central Asia over the past 27 years derived from medium-resolution remote sensing data. Int. J. Appl. Earth Obs. Geoinf.; 2014; 26, pp. 335-349. [DOI: https://dx.doi.org/10.1016/j.jag.2013.08.004]

15. Lu, S.; Ma, J.; Ma, X.; Tang, H.; Zhao, H.; Baig, M.H.A. Time series of the Inland Surface Water Dataset in China (ISWDC) for 2000–2016 derived from MODIS archives. Earth Syst. Sci. Data; 2019; 11, pp. 1099-1108. [DOI: https://dx.doi.org/10.5194/essd-11-1099-2019]

16. Gao, B.C. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ.; 1996; 58, pp. 257-266. [DOI: https://dx.doi.org/10.1016/S0034-4257(96)00067-3]

17. Xu, H. Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. Int. J. Remote Sens.; 2006; 27, pp. 3025-3033. [DOI: https://dx.doi.org/10.1080/01431160600589179]

18. Feyisa, G.L.; Meilby, H.; Fensholt, R.; Proud, S.R. Automated Water Extraction Index: A new technique for surface water mapping using Landsat imagery. Remote Sens. Environ.; 2014; 140, pp. 23-35. [DOI: https://dx.doi.org/10.1016/j.rse.2013.08.029]

19. Yang, B.; Ke, X. Analysis on urban lake change during rapid urbanization using a synergistic approach: A case study of Wuhan, China. Phys. Chem. Earth A/B/C; 2015; 89, pp. 127-135. [DOI: https://dx.doi.org/10.1016/j.pce.2015.10.005]

20. Li, K.; Xu, E. High-accuracy continuous mapping of surface water dynamics using automatic update of training samples and temporal consistency modification based on Google Earth Engine: A case study from Huizhou, China. ISPRS J. Photogramm. Remote Sens.; 2021; 179, pp. 66-80. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2021.07.009]

21. Duan, Y.; Zhang, W.; Huang, P.; He, G.; Guo, H. A New Lightweight Convolutional Neural Network for Multi-Scale Land Surface Water Extraction from GaoFen-1D Satellite Images. Remote Sens.; 2021; 13, 4576. [DOI: https://dx.doi.org/10.3390/rs13224576]

22. Zhang, T.; Su, H.; Yang, X.; Yan, X. Remote sensing prediction of global subsurface thermohaline and the impact of longitude and latitude based on LightGBM. Natl. Remote Sens. Bull.; 2020; 24, pp. 1255-1269. [DOI: https://dx.doi.org/10.11834/jrs.20200007]

23. Luo, X.; Hu, Z.; Liu, L. Investigating the seasonal dynamics of surface water over the Qinghai–Tibet Plateau using Sentinel-1 imagery and a novel gated multiscale ConvNet. Int. J. Digit. Earth; 2023; 16, pp. 1372-1394. [DOI: https://dx.doi.org/10.1080/17538947.2023.2198266]

24. Zhang, T.; Li, W.; Feng, X.; Ren, Y.; Qin, C.; Ji, W.; Yang, X. Super-Resolution Water Body Extraction Based on MF-SegFormer. IGARSS 2024—2024 IEEE International Geoscience and Remote Sensing Symposium; IEEE: New York, NY, USA, 2024; pp. 9848-9852.

25. Pan, H.; Chen, H.; Hong, Z.; Liu, X.; Wang, R.; Zhou, R.; Zhang, Y.; Han, Y.; Wang, J.; Yang, S. et al. A Novel Boundary Enhancement Network for Surface Water Mapping Based on Sentinel-2 MSI Data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2023; 16, pp. 9207-9222. [DOI: https://dx.doi.org/10.1109/JSTARS.2023.3308046]

26. Miao, Z.; Fu, K.; Sun, H.; Sun, X.; Yan, M. Automatic water-body segmentation from high-resolution satellite images via deep networks. IEEE Geosci. Remote Sens. Lett.; 2018; 15, pp. 602-606. [DOI: https://dx.doi.org/10.1109/LGRS.2018.2794545]

27. Zhong, H.F.; Sun, Q.; Sun, H.M.; Jia, R.-S. NT-Net: A semantic segmentation network for extracting lake water bodies from optical remote sensing images based on transformer. IEEE Trans. Geosci. Remote Sens.; 2022; 60, 5627513. [DOI: https://dx.doi.org/10.1109/TGRS.2022.3197402]

28. Tambe, R.G.; Talbar, S.N.; Chavan, S.S. Deep multi-feature learning architecture for water body segmentation from satellite images. J. Vis. Commun. Image Represent.; 2021; 77, 103141. [DOI: https://dx.doi.org/10.1016/j.jvcir.2021.103141]

29. Yu, Y.; Huang, L.; Lu, W.; Guan, H.; Ma, L.; Jin, S.; Yu, C.; Zhang, Y.; Tang, P.; Liu, Z. et al. WaterHRNet: A multibranch hierarchical attentive network for water body extraction with remote sensing images. Int. J. Appl. Earth Obs. Geoinf.; 2022; 115, 103103. [DOI: https://dx.doi.org/10.1016/j.jag.2022.103103]

30. Scott, M.; Lee, S.-I. A unified approach to interpreting model predictions. Adv. Neural Inf. Process. Syst.; 2017; 30, pp. 4765-4774.

31. Wu, J.; Zhou, L.; Mo, X.; Zhou, H.; Zhang, J.; Jia, R. Drought monitoring and analysis in China based on the Integrated Surface Drought Index (ISDI). Int. J. Appl. Earth Obs. Geoinf.; 2015; 41, pp. 23-33. [DOI: https://dx.doi.org/10.1016/j.jag.2015.04.006]

32. Roy, D.P.; Wulder, M.A.; Loveland, T.R.; Woodcock, C.E.; Allen, R.G.; Anderson, M.C.; Helder, D.; Irons, J.R.; Johnson, D.M.; Kennedy, R. et al. Landsat-8: Science and product vision for terrestrial global change research. Remote Sens. Environ.; 2014; 145, pp. 154-172. [DOI: https://dx.doi.org/10.1016/j.rse.2014.02.001]

33. Tamiminia, H.; Salehi, B.; Mahdianpari, M.; Quackenbush, L.; Adeli, S.; Brisco, B. Google Earth Engine for geo-big data applications: A meta-analysis and systematic review. ISPRS J. Photogramm. Remote Sens.; 2020; 164, pp. 152-170. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2020.04.001]

34. Su, H.; Wei, S.; Qiu, J.; Wu, W. RaftNet: A new deep neural network for coastal raft aquaculture extraction from Landsat 8 OLI data. Remote Sens.; 2022; 14, 4587. [DOI: https://dx.doi.org/10.3390/rs14184587]

35. Liu, Y.; Shao, Z.; Hoffmann, N. Global attention mechanism: Retain information to enhance channel-spatial interactions. arXiv; 2021; arXiv: 2112.05561

36. Zhao, H.; Shi, J.; Qi, X.; Wang, X.; Jia, J. Pyramid scene parsing network. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Honolulu, HI, USA, 21–26 July 2017; pp. 2881-2890.

37. Zhao, H.; Qi, X.; Shen, X.; Shi, J.; Jia, J. ICNet for real-time semantic segmentation on high-resolution images. Proceedings of the European Conference on Computer Vision (ECCV); Munich, Germany, 8–14 September 2018; pp. 405-420.

38. Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Salt Lake City, UT, USA, 18–23 June 2018; pp. 7132-7141.

39. Wang, Q.; Wu, B.; Zhu, P.; Li, P.; Zuo, W.; Hu, Q. ECA-Net: Efficient channel attention for deep convolutional neural networks. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; Seattle, WA, USA, 13–19 June 2020; pp. 11534-11542.

40. Lin, T.Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal loss for dense object detection. Proceedings of the IEEE International Conference on Computer Vision; Venice, Italy, 22–29 October 2017; pp. 2980-2988.

41. Milletari, F.; Navab, N.; Ahmadi, S.A.; Net, V. Fully convolutional neural networks for volumetric medical image segmentation. Proceedings of the 2016 Fourth International Conference on 3D Vision (3DV); Stanford, CA, USA, 25–28 October 2016; IEEE: New York, NY, USA, 2016; pp. 565-571.

42. Li, Z.; He, W.; Li, J.; Lu, F.; Zhang, H. Learning without Exact Guidance: Updating Large-scale High-resolution Land Cover Maps from Low-resolution Historical Labels. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; Seattle, WA, USA, 16–22 June 2024; pp. 27717-27727.

43. Long, J.; Shelhamer, E.; Darrell, T. Fully convolutional networks for semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Boston, MA, USA, 7–12 June 2015; pp. 3431-3440.

44. Chen, L.C.; Zhu, Y.; Papandreou, G.; Schroff, F.; Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. Proceedings of the European Conference on Computer Vision (ECCV); Munich, Germany, 8–14 September 2018; pp. 801-818.

45. Strudel, R.; Garcia, R.; Laptev, I.; Schmid, C. Segmenter: Transformer for semantic segmentation. Proceedings of the IEEE/CVF International Conference on Computer Vision; Montreal, QC, Canada, 11–17 October 2021; pp. 7262-7272.

46. Zheng, S.; Lu, J.; Zhao, H.; Zhu, X.; Luo, Z.; Wang, Y.; Fu, Y.; Feng, J.; Xiang, T.; Torr, P.H.S. et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; Nashville, TN, USA, 19–25 June 2021; pp. 6881-6890.

47. 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, pp. 1140-1156.

48. MMSegmentation Contributors. MMSegmentation: OpenMMLab Semantic Segmentation Toolbox and Benchmark. GitHub Repository. 2020; Available online: https://github.com/open-mmlab/mmsegmentation (accessed on 10 July 2023).

49. Loshchilov, I.; Hutter, F. Decoupled Weight Decay Regularization. arXiv; 2017; arXiv: 1711.05101

50. Deng, T.; Geng, G.; Yang, R.; Zhang, B. Temporal and spatial variation characteristics of high temperature and heat wave in the Weihe River Basin from 1980 to 2020. Arid Land Geogr.; 2023; 46, pp. 211-221.