DBCE-Net: A Novel Deep Learning Framework for Annual Mapping of Coastal Aquaculture Ponds in China with Sentinel-2 Data

Abstract

Despite the promising advancements of deep learning techniques in coastal aquaculture pond extraction, their capacity for large-scale mapping tasks remains relatively limited. To address this challenge, this study developed a novel deep learning framework, Dual-Branch Enhanced Network (DBCE-Net), for mapping the annual aquaculture ponds at the national scale using Sentinel-2 imagery. The DBCE-Net framework effectively mitigates the contextual information loss inherent in traditional methods and reduces classification errors by processing both down-sampled large-scale images and block images at their original resolution. The architecture comprises local feature extraction and global feature extraction, along with feature fusion and decoding. The pivotal Multi-scale Dynamic Feature Fusion (DFF) module synthesizes local and global features while incorporating complementary information, demonstrating strong robustness with smaller training areas, compared to previous methods that required a larger number of samples distributed across different regions. By applying the DBCE-Net to Sentinel-2 imagery from 2017 to 2023, we mapped the spatiotemporal distribution of coastal aquaculture ponds across all coastal counties in China, achieving an overall classification accuracy approximately 93%. The results demonstrate substantial changes in the area of coastal aquaculture ponds in China from 2017 to 2023, with the total area declining from 8970.25 km² to 8261.17 km², representing a notable decrease of 7.90%. The most pronounced reduction was observed in Shanghai, with a decrease of 38.92%, followed by Zhejiang (31.57%) and Jiangsu (19.07%). These reductions are primarily attributed to policies converting aquaculture ponds into natural wetlands. In contrast, the area of coastal aquaculture ponds in Liaoning Province slightly increased by 5.75%. This DBCE-Net demonstrates good accuracy and generalizability and is promising to further expand its application to the extraction of coastal aquaculture areas worldwide, providing important scientific value and practical significance for the global coastal aquaculture industry.

1. Introduction

Aquaculture, an essential pillar of the global food supply chain, serves a crucial function in maintaining biodiversity, ecological equilibrium, and ensuring food security. With the continual growth of the global population and the progressive depletion of marine resources, the aquaculture sector is becoming ever more vital [1]. According to the Food and Agriculture Organization of the United Nations (FAO), global fishery and aquaculture production has reached an unprecedented high, with the industry projected to play an increasingly significant role in providing food and nutrition in the future. China is the largest aquaculture nation globally, supplying over 60% of the world’s aquaculture products [2,3,4]. However, the rapid expansion of coastal pond aquaculture has precipitated a series of environmental issues, including disrupting ecological balance, damaging crucial ecosystems such as coastal wetlands, coral reefs, and seagrass beds, and adversely affecting local biodiversity and habitats [5,6]. Interactions between aquaculture ponds and surrounding marine ecosystems, such as the discharge of aquaculture wastewater, can threaten the survival and reproduction of marine organisms, resulting in marine pollution and other issues [7,8,9,10]. As the scale of aquaculture continues to change, traditional methods of manual field monitoring and statistical surveys are increasingly insufficient to meet the demands for efficient and precise management, creating an urgent need for developing efficient and reliable aquaculture monitoring methods.

In recent years, an escalating number of researchers have leveraged machine learning and deep learning techniques for the automatic extraction and mapping of aquaculture ponds. For mapping coastal aquaculture ponds, random forest algorithms and decision tree models are prevalently employed. These methods are integrated with the analysis of various features of aquaculture pond water bodies, such as spectral indices (e.g., Normalized Difference Water Index [NDWI] [11] and Modified Normalized Difference Water Index [MNDWI] [12]), morphology, and texture [13,14,15]. Existing studies have extensively applied these methods using Sentinel-2 satellite imagery in China and other large-scale coastal regions [16,17]. However, this method necessitates manually setting distinct thresholds and parameters for various regions, requiring multiple parameters nationwide, and its effectiveness is contingent upon the support of other datasets [16,17,18,19,20,21]. Additionally, methods combining edge detection and morphological processing can identify and analyze pond edges and morphological features to extract aquaculture pond information. However, these methods are limited in distinguishing coastal aquaculture ponds from similar surface water bodies [22]. Due to their high similarity in size, appearance, and spectral characteristics, these water bodies often lead to classification errors. In addition to these challenges, traditional approaches are further hindered by their limited capacity to effectively integrate spatial information, which impairs their performance in areas characterized by complex environmental conditions. Moreover, the reliance on external datasets and the necessity of adjusting thresholds for different regions at large scales introduce additional complexities, thereby diminishing both the efficiency and accuracy of the mapping process.

Deep learning methods, with their powerful feature extraction capabilities, have demonstrated exceptional performance in various remote sensing applications, including the precise identification and classification of aquaculture ponds. However, their application for large-scale extraction of aquaculture ponds on a national level remains limited. This highlights an urgent need for a highly generalizable and automated approach to accurately extract large-scale coastal aquaculture ponds. In recent years, several studies have attempted to address these challenges. Gao et al. proposed a semantic segmentation model based on D-ResUnet for high-precision extraction of marine floating raft aquaculture (FRA) areas [23]. Deng et al. introduced a coastal aquaculture network to resolve issues of boundary accuracy and complex backgrounds, successfully extracting aquaculture areas from high-resolution remote sensing images [24]. Zou et al. developed a U2-Net-based model for coastal aquaculture pond extraction and validated it in the Zhoushan Islands [25]. Ai et al. proposed the SAMALNet model, which utilizes high-resolution satellite imagery along with a self-attention mechanism and auxiliary loss network to enhance the accuracy of boundary extraction in aquaculture areas [26]. Furthermore, Liang et al. innovatively introduced the DIAS index and proposed the MAFU-Net model, achieving high-precision extraction of aquaculture ponds [27]. Despite these advancements, traditional single-branch structures in remote sensing deep learning still face inherent limitations in capturing contextual and boundary information. To address this, existing studies have incorporated pyramid structures or dilated convolutions to enhance the model’s ability to process multi-scale information [28,29]. However, these approaches often struggle with the precise differentiation of features in regions with complex geographical characteristics, such as multi-branch rivers and aquaculture ponds, due to their limited ability to handle small-scale perspectives. Additionally, edge information loss further reduces the accuracy of classification and identification.

In this study, we developed DBCE-Net (Dual-Branch Enhanced Network), an innovative deep learning framework specifically tailored for the mapping of remote sensing imagery of coastal aquaculture ponds. This framework employs a dual-branch structure: the main branch concentrates on accurately capturing local features of high-resolution images, while the Context Edge Supplement Branch (CESB) processes down-sampled images, supplementing global contextual and boundary information, thereby enhancing overall scene understanding. Through the Multi-scale Dynamic Feature Fusion (DFF) module, we achieved efficient integration of local details with global information, markedly enhancing the model’s accuracy in identification and classification within complex backgrounds.

The primary contributions of this study are

(1)

Design a dual-branch DBCE-Net deep learning framework that combines high-resolution small image features with low-resolution large image features, effectively addressing the information loss problem inherent in traditional single-branch networks.

(2)

Propose an innovative Multi-scale Dynamic Feature Fusion (DFF) module to effectively integrate complementary information from the main branch and CESB, capturing and utilizing multi-scale features.

(3)

Selecting typical aquaculture pond regions in China and applying the small sample to large-scale coastal areas, achieving extensive extraction and mapping of aquaculture ponds across China.

Through these innovations, DBCE-Net effectively addresses key challenges in analyzing remote sensing imagery of coastal aquaculture environments, such as improving classification accuracy, enhancing boundary identification accuracy, and reducing spectral similarity interference.

2. Study Areas and Data

2.1. Study Areas

China’s coastline begins at the Yalu River estuary and extends to the Beilun River estuary, spanning a total length of 18,000 km and characterized by its winding paths and numerous islands [30]. These unique natural conditions provide an exceptionally favorable environment for the thriving development of the marine aquaculture industry. Coastal marine aquaculture primarily involves the cultivation of fish, shellfish, seaweed, shrimp, and crabs. These aquaculture activities play a pivotal role in promoting local economic development and are also essential for maintaining regional ecological balance [31].

In this study, we selected Chinese coastal counties and a 10 km buffer zone extending seaward from the coastline as the study area (Figure 1). This choice was motivated by the significant annual changes in coastline position and coastal land use caused by natural and human factors, which render static coastline data insufficient to capture the dynamic characteristics of aquaculture ponds. The 10 km buffer zone provides better coverage of these spatial variations while using counties as the basic statistical unit, offering more accurate and targeted decision making and resource management support to local governments.

2.2. Data and Preprocessing

The Sentinel-2 satellite series constitutes a pivotal element of the European Space Agency’s (ESA) remote sensing observation program, encompassing two key satellites: Sentinel-2A and Sentinel-2B. This dual-satellite configuration achieves a revisit period of at least once every 5 days and offers a wide field of view of up to 290 km. Equipped with Multispectral Instruments (MSI) featuring 13 spectral bands spanning visible to near-infrared wavelengths and spatial resolutions ranging from 10 to 60 m, Sentinel-2 provides crucial data for diverse applications [5,32,33]. With its 10 m spatial resolution, rich spectral information across 13 bands, and high-frequency revisit period, Sentinel-2 significantly reduces cloud interference, supporting large-scale and long-term aquaculture monitoring [6,34,35]. The L1C level data comprise raw, georeferenced multispectral images widely used for agricultural monitoring, urban planning, and environmental protection [36]. Sentinel-2’s wide coverage, high temporal resolution, and spectral diversity make it invaluable for land cover and dynamic change detection, particularly in remote sensing research.

It is important to note that many aquaculture ponds undergo cleaning and drying operations from December to February, which may result in imagery from this period not accurately reflecting typical conditions in these areas. To enhance data quality, this study focuses on remote sensing imagery from March to November each year, synthesizing annual median images for subsequent monitoring. Cloud masking is performed using the QA60 band, with a filtering criterion set to CLOUDY_PIXEL_PERCENTAGE less than 20% to effectively identify and exclude cloudy pixels.

3. Method

3.1. Creation of Sample Dataset

The primary objective of this study is to utilize semantic segmentation techniques for the binary classification of surface water bodies in coastal areas, distinctly categorizing them into aquaculture ponds and background. By accurately identifying aquaculture ponds within coastal water bodies, we aim to conduct a detailed analysis and research on these key areas, while classifying all other non-aquaculture surface water bodies as background to concentrate on the identification and study of aquaculture ponds.

The Multispectral Imaging Instrument (MSI) on Sentinel-2 satellite is capable of capturing image data in 13 spectral bands (

Table 1. Sentinel-2 bands and their central wavelengths.

Spectral Range	Band Name	Central Wavelength (nm)
Visible	Blue (B2)	490
	Green (B3)	560
	Red (B4)	665
Near-Infrared	NIR (B8)	842
Shortwave Infrared	SWIR1 (B11)	1610
	SWIR2 (B12)	2190
Other Bands	B1	443
	B5	705
	B6	740
	B7	783
	B8A	865
	B9	935
	B10	1375

), covering spectra from visible light to near-infrared and shortwave infrared light. Among them, B2 (blue), B3 (green), B4 (red), B8 (near-infrared), and B11 (shortwave infrared) exhibit significant spectral differences, making them effective for distinguishing aquaculture ponds, salt fields, and dikes. In this study, we selected bands B2, B3, B4, B8, and B11. The spectral differences of these bands are crucial for extracting key land features in coastal areas, particularly for distinguishing aquaculture ponds from other water bodies or background features.

The Sentinel-2 satellite series constitutes a pivotal element of the European Space Agency’s (ESA) remote sensing observation program, encompassing two key satellites: Sentinel-2A and Sentinel-2B. This dual-satellite configuration achieves a revisit period of at least once every 5 days and offers a wide field of view of up to 290 km. These satellites are outfitted with Multispectral Instruments (MSI), featuring 13 spectral bands spanning wavelengths from visible to near-infrared, with spatial resolutions ranging from 10 to 60 m [5,32,33].

Sentinel-2 satellite imagery, with its 10 m median spatial resolution and rich spectral information across 13 bands, coupled with a high-frequency revisit period of 5 days, significantly reduces cloud interference, thereby supporting large-scale and long-term aquaculture dynamic monitoring needs [6,34,35]. The L1C level data provided by the Sentinel-2 series consist of raw, georeferenced multispectral images that are widely recognized for their broad applicability [36]. These data play a pivotal role in various remote sensing analyses, encompassing agricultural monitoring, urban planning, and environmental protection. The primary features of satellite data include wide coverage, high temporal resolution, and a rich selection of spectral bands, rendering Sentinel-2 extremely valuable for land cover and dynamic change monitoring. Particularly, its comprehensive optical imaging capabilities make it an indispensable tool in the field of remote sensing research.

Accurately distinguishing between salt pans, aquaculture ponds, embankments, and paddy fields is a critical step in the extraction and classification of coastal aquaculture ponds. To address this challenge, this study selected three regions along China’s coastline, spanning from north to south: Bohai Bay, Yancheng, and Shantou, as the study areas for spectral band analysis. A total of 161 embankment samples, 251 aquaculture pond samples, 92 salt pan samples, and 159 paddy field samples were collected.

Through an in-depth analysis of the spectral reflectance characteristics of these sample points, it was found that Sentinel-2 satellite imagery bands B2 (blue), B3 (green), B4 (red), B8 (near-infrared), and B11 (shortwave infrared) exhibited significant spectral reflectance differences among the four land cover types (Figure 2). Notably, paddy fields showed high spectral similarity to aquaculture ponds in the B2, B3, and B4 bands, making them difficult to differentiate. However, significant differences were observed in the B8 and B11 bands, especially during the vegetative growth stage of paddy fields. By incorporating these two bands into the analysis, combined with the robust feature extraction capabilities of deep learning models, the classification model was able to reduce confusion between paddy fields and aquaculture ponds. Consequently, these five bands were selected for subsequent analysis to enhance classification accuracy and improve the reliability and applicability of the classification model.

It is important to note that many aquaculture ponds undergo cleaning and drying operations from December to February [37], which may result in imagery from this period not accurately reflecting typical conditions in these areas. To enhance data quality, this study focuses on remote sensing imagery from March to November each year, synthesizing annual median images for subsequent monitoring. Cloud masking is performed using the QA60 band, with a filtering criterion CLOUDY_PIXEL_PERCENTAGE set to less than 20% to effectively identify and exclude cloudy pixels.

The characteristics of aquaculture ponds in China exhibit significant regional diversity, particularly the stark contrast between the northern and southern regions. The north is characterized by large-scale aquaculture ponds and salt fields, while the south features smaller, more complex ponds with diverse attributes. To capture this diversity, we selected Sentinel-2 satellite imagery from 2021 and chose Bohai Bay (Figure 1a) in the north, along with Hangzhou Bay and Sanmen Bay in Zhejiang Province (Figure 1b) in the south, as regions for training sample selection. These areas not only represent typical features of aquaculture ponds in northern and southern China, but also include land features such as salt pan in the north and rice fields in the south that are easily confused with aquaculture ponds. The image information for each area is presented in

Table 2. Image information for different areas.

Area	Image Size	Image Data (Date Range)
Bohai Bay	16,516 × 16,636	2021-03-01–2021-12-01
Hangzhou Bay	14,978 × 4464	2021-03-01–2021-12-01
Sanmen Bay	7436 × 3136	2021-03-01–2021-12-01

. These regions encapsulate the representative features of aquaculture ponds across China, enabling the model to generalize effectively for nationwide and multi-year applications. The sample data consist of high-resolution images provided by Google Earth [38], which were annotated through visual interpretation methods. These annotations classified the data into aquaculture ponds and background categories to ensure the precision and practical applicability of the training dataset.

During this study, we generated a total of 802 images, each with a pixel size of 512 × 512. These images encompass samples from various geographical environments in coastal areas, including lakes, rivers, aquaculture ponds, and salt pans. For the semantic segmentation task, we explicitly categorized these images into aquaculture ponds and background to meet the requirements of our study.

In this study, an 80/20 split was employed to partition the training and validation datasets. Specifically, 628 images were used for training the model, while 174 images were allocated to the validation set. This ratio is widely acknowledged as a standard practice in deep learning tasks, particularly for medium-sized datasets [39], due to its balance between efficiency and representativeness. The chosen partitioning strategy ensures that a sufficient number of samples are available for training, allowing the model to effectively learn the underlying data features, while preserving a representative subset for validation. This approach facilitates a robust evaluation of the model’s generalization ability and mitigates the risk of performance degradation caused by overfitting.

3.2. Data Augmentation

Compared to traditional sliding window methods, the sample augmentation strategy proposed in this study effectively avoids over-training on repetitive regions, thereby mitigating the risk of model overfitting. Given the limitation of training sample size, we designed and implemented an online data augmentation strategy to enhance sample diversity in real-time during the training process. Specifically, during the construction of the Dataloader, we introduced random parameters that allow, with a certain probability, operations such as rotation, cropping, and scaling to be applied to samples before training each batch, thus generating augmented samples dynamically. This online augmentation approach eliminates the need for additional storage of augmented data, enabling the optimal utilization of limited samples while significantly improving the model’s generalization capability across diverse scenarios.

3.3. Model Structure

In the design of DBCE-Net (Dual-Input Context Enhanced Network), the network architecture primarily comprises the main branch, the Context Edge Supplement Branch (CESB), and the feature fusion and decoding modules (Figure 3), where DBCE-Net outputs the final segmentation map exclusively from the main branch, and with CESB enhancing the network’s performance by supplementing contextual information.

In this study, we selected ResNeSt [40] as the encoder for our model due to its substantial performance improvement and efficient capture of complex data features. ResNeSt, a significant evolution of ResNet [41], represents the latest generation of deep convolutional neural networks. By introducing the Split-Attention mechanism, it maintains controllability over network depth and parameter scale while markedly enhancing the network’s feature representation capabilities. This mechanism allows the network to perform adaptive recalibration of features within each block, thereby strengthening the model’s ability to capture key features and improving overall performance.

During the feature extraction stage, DBCE-Net’s main branch employs ResNeSt200e as the encoder, concentrating on extracting detailed information from high-resolution remote sensing images. Meanwhile, CESB processes images with fixed center points and expanded sizes, primarily to supplement the contextual and boundary information missing due to cropping in the main branch. This structural design enables the Network to comprehensively address contextual loss due to cropping in remote sensing images and also enhances the network’s multi-scale capabilities.

In the feature fusion stage, the Dynamic Feature Fusion (DFF) module intelligently integrates features from the main branch and CESB. In the decoding phase, DBCE-Net gradually restores compressed feature maps to their original size, ultimately outputting precise aquaculture pond segmentation maps exclusively from the main branch.

Through this multi-branch structure and innovative loss function design, DBCE-Net fully exploits the rich information provided by remote sensing imagery, thereby enhancing the accuracy and efficiency of aquaculture pond extraction. The workflow for aquaculture pond classification based on DBCE-Net (Figure 4) is shown below.

3.3.1. Context-Enhanced Supplementary Branch

When extracting aquaculture ponds, traditional networks often struggle due to the lack of global contextual information and boundary features. These limitations result in frequent misclassifications, such as mistakenly identifying complex, multi-branch rivers or lakes as aquaculture ponds, or failing to extract the ponds completely. This issue primarily stems from the inherent constraints of single-branch networks, whose feature extraction capabilities are limited by their local receptive fields, making it difficult to capture the broader spatial relationships and global background information within the image.

The fundamental principle of the Context-Enhanced Supplementary Branch (CESB) lies in its ability to process larger-scale input images and extract expanded contextual information to compensate for the main branch’s inability to independently capture spatial relationships and boundary details. In the task of aquaculture pond extraction, CESB effectively recognizes the global relationships between ponds and surrounding geographical features, such as rivers, lakes, and vegetation. This capability aids in distinguishing visually similar but fundamentally different objects, while also enhancing the completeness and accuracy of pond boundary delineation. Unlike traditional single-branch networks, which primarily focus on local feature extraction, CESB emphasizes macro-level feature representation, significantly improving the network’s understanding of complex scenes.

Specifically, CESB dynamically integrates the localized features extracted by the main branch with its own global contextual information. This integration not only prevents the misclassification of rivers or lakes as aquaculture ponds but also reduces the risk of incomplete pond extraction caused by insufficient local information. By supplementing the main branch with expanded contextual features, CESB addresses challenges such as boundary ambiguity and classification errors, leading to substantial improvements in segmentation accuracy and robustness.

To achieve this, CESB incorporates ResNeSt50 as its lightweight encoder, striking an optimal balance between computational efficiency and feature extraction capacity. The lightweight design of ResNeSt50 minimizes the computational overhead introduced by CESB, while ensuring the effective capture of additional contextual features essential for enhancing segmentation performance. This design consideration makes CESB particularly suitable for computationally demanding tasks requiring high efficiency.

The outputs generated by CESB do not replace the feature representations of the main branch. Instead, they undergo fine-grained integration through a Dynamic Feature Fusion module, which ensures seamless incorporation of the contextual information extracted by CESB into the localized feature representations of the main branch. By combining global contextual information with fine-grained local details, the fusion module enables the network to produce unified, hierarchical feature representations. This approach enhances overall segmentation accuracy and robustness, particularly for aquaculture pond extraction. It effectively differentiates aquaculture ponds from other water bodies and ensures accurate delineation of pond boundaries.

3.3.2. Dynamic Feature Fusion Module

To effectively integrate the backbone features and contextual boundary features extracted by the dual-branch network, we designed the Dynamic Feature Fusion (DFF) module. The DFF module is designed to learn the complementary relationship between high-resolution features and contextual boundary features, resulting in enhanced feature representations that encompass both local details and global semantics. This is particularly beneficial in the context of aquaculture pond extraction, where it demonstrates effective effectiveness.

In the extraction of aquaculture ponds, the DFF module improves segmentation performance by enhancing the precise capture of boundaries and the overall perception of pond regions. Aquaculture ponds are often located in complex geographical environments, such as areas with branching rivers or adjacent water bodies. Their boundaries are susceptible to interference, and the internal texture information is relatively sparse, making it challenging for traditional segmentation methods to distinguish ponds from other water bodies. By dynamically adjusting the weight distribution between backbone features and contextual boundary features, the DFF module effectively mitigates issues such as ambiguous boundaries, incomplete feature representation, and misclassification, enabling accurate and comprehensive extraction of pond regions.

In the design of the DFF module (Figure 5), three dilated convolutions with different dilation rates (2, 3 and 4) are first applied to the contextual boundary features. These operations aggregate rich global contextual information through multi-scale dilated convolutions [28], enlarging the receptive field and capturing long-range relationships within the image, thereby enhancing the semantic representation of features. For aquaculture pond segmentation, this global perception capability enables the network to better distinguish ponds from surrounding environments, reducing the likelihood of misclassification.

The DFF module further incorporates two feature processing and fusion branches. The first branch applies a 1 × 1 convolution followed by a Sigmoid activation function and performs element-wise multiplication with the original feature map. This process enhances the significant regions of the feature map while suppressing less important areas, achieving feature selection and weight allocation. The second branch begins with a pooling operation to reduce spatial dimensions and aggregate feature information. It then applies two sequential 1 × 1 convolutions followed by Sigmoid activation functions, culminating in element-wise multiplication with the original feature map. This design refines the information within the feature map, ensuring the accurate extraction of features relevant to aquaculture ponds.

By combining the outputs of the two branches, the DFF module achieves efficient feature fusion and dynamic weighting, ensuring precise extraction of critical features. Through the introduction of a dynamic weight learning mechanism, the DFF module not only adjusts weight distributions across feature channels but also optimizes weights in the spatial dimension. This flexibility allows the module to adapt to various image scenes. In aquaculture pond segmentation, the DFF module significantly enhances global perception, contextual boundary information processing, and local detail capture, resulting in improved accuracy and robustness.

3.3.3. Training Parameter Settings and Post-Processing

In this study, to meet the high computational demands of deep learning, we selected Python as the development language and PyTorch as the primary framework. CUDA (version 11.2) was employed for GPU acceleration in the experiments, running on a system equipped with an NVIDIA GeForce RTX 3090 GPU and an Intel(R) Xeon(R) Gold 6226R CPU @ 2.90 GHz. The Python version used in development was 3.8.5. Additionally, the GDAL library was utilized for handling geospatial data. After a series of tests, we determined the optimal training configuration: 200 training epochs, a batch size of 24, the Adam optimizer, an initial learning rate set to 1 × 10⁻⁴, and the ReLU activation function. We chose Focal Loss [42] and Dice Loss [43] as the loss functions. The training loss curve is illustrated (Figure 6).

In addition, it is essential to differentiate aquaculture ponds from independent reservoirs, which, although often possessing similar shapes and sizes, are distinct entities. These independent reservoirs cannot be reliably distinguished from aquaculture ponds based solely on spectral and geometric characteristics. Consequently, to account for the spatial distribution patterns of aquaculture ponds, the distance between the nearest neighboring objects of potential aquaculture ponds was calculated. A threshold was then established to effectively exclude independently distributed reservoirs. The results from the training process indicate that a minimum distance threshold of 50 m is optimal; if no other water bodies are located within 50 m of a given water body, it is classified as an independent reservoir.

$$\mathrm{FocalLoss}=-{\alpha }_t{(1-p_t)}^{\gamma }log\left(p_t\right)$$

(1)

$$\mathrm{DiceLoss}=1-{\displaystyle \frac{2TP}{2TP+FP+FN}}$$

(2)

$$\mathrm{TotalLoss}=\mathrm{FocalLoss}+\mathrm{DiceLoss}$$

(3)

$p_t$ is the model’s estimated probability for the true class, ${\alpha }_t$ is a balancing factor for class imbalance, and $\gamma$ is a focusing parameter used to reduce the impact of easy examples. $TP$ is the number of true positives, $FP$ is the number of false positives, $FN$ is the number of false negatives.

3.4. Evaluation Metrics

We used the F1 score [44], precision, recall, and OA as metrics to evaluate the performance of the model. We also calculated confusion matrices for accuracy assessment. The formulas for the evaluation metrics are as follows:

$$\mathrm{Precision}={\displaystyle \frac{TP}{TP+FP}}$$

(4)

$$\mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}}$$

(5)

$$\mathrm{F}1\mathrm{Score}=2·{\displaystyle \frac{\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(6)

$$\mathrm{OA}={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(7)

where $TP$ is the number of true positive samples (i.e., samples correctly predicted as aquaculture ponds), $TN$ is the number of true negative samples (i.e., samples correctly predicted as non-aquaculture ponds), $FN$ is the number of false negative samples (i.e., aquaculture pond samples incorrectly predicted as non-aquaculture ponds), and $FP$ is the number of false positive samples (i.e., non-aquaculture pond samples incorrectly predicted as aquaculture ponds).

Our sampling points were initially generated based on the 2021 ESA dataset [45], covering various water bodies such as rivers, reservoirs, lakes, farmland, seasonal wetlands, and rice fields (Figure 7). To address the issue of annual variations in the distribution of aquaculture ponds, we also rely on high-resolution Google Earth images manually annotated each year to migrate samples to different years.

4. Results

4.1. Accuracy Assessment

The results (

Table 3. Confusion matrix and F1 scores for the aquaculture pond map generated by DBCE-Net in China from 2017 to 2023.

Year	Class	AP	Non-AP	Precision	Recall	F1 Score	OA
2017	AP	728	68	91.5%	89.7%	90.5%	92.7%
	Non-AP	84	1209	93.5%	94.8%	94.1%
2018	AP	745	67	91.7%	89.8%	90.7%	92.8%
	Non-AP	85	1210	93.4%	94.9%	94.1%
2019	AP	745	58	92.8%	89.8%	91.2%	93.2%
	Non-AP	85	1220	93.5%	95.5%	94.5%
2020	AP	726	49	93.7%	89.9%	91.7%	93.7%
	Non-AP	82	1228	93.7%	96.2%	94.9%
2021	AP	773	61	92.7%	88.7%	90.7%	92.5%
	Non-AP	98	1179	92.3%	95.1%	93.7%
2022	AP	713	60	92.2%	89.5%	90.9%	93.1%
	Non-AP	84	1217	93.5%	95.3%	94.4%
2023	AP	687	55	92.6%	88.6%	90.6%	93.1%
	Non-AP	86	1230	93.3%	95.7%	94.5%

) indicate that the overall accuracy evaluation metric for the aquaculture pond map produced in this study averages around 93%. Specifically, the precision and F1 score for aquaculture ponds both exceed 90%. The confusion matrix results demonstrate a high level of agreement between the aquaculture pond map generated in this study and the validation sample points. The training model established based on 2019 samples maintains comparable accuracy in other years, indicating that the model exhibits good temporal consistency and stability. The model’s ability to sustain high accuracy across different years underscores its reliability in temporal detection tasks.

To further verify the accuracy of the contours and shapes of aquaculture ponds predicted by this method, we compared the ponds generated using this method with polygons manually delineated from high-resolution Google Earth images.

The comparison results indicate that the extracted contours and shapes are highly consistent with the manually delineated high-precision results, as illustrated in the figure. This comparison validates the reliability of the method proposed in this study and demonstrates its effectiveness in extracting aquaculture pond information (Figure 8).

We observed that the aquaculture pond extraction results obtained using this method are highly consistent with visual interpretation results in most cases, exhibiting significant overlap and extraction performance.

However, we also noticed some errors in extracting pond edges, resulting in an overall extracted aquaculture pond area that is slightly smaller than the actual area. This phenomenon is primarily due to the limitations of spatial resolution, which affect the precise extraction of aquaculture pond boundaries. To further improve the accuracy of boundary extraction, future research could consider using higher resolution imagery or developing more advanced image processing algorithms to mitigate the impact of spatial resolution on the results.

4.2. Verification of Typicality of Selected Data

To verify the representativeness of the training data and the generalization ability of the model, we conducted statistical analysis and comparative analysis on the provincial sample points from all years.

(

Table 4. Accuracy in different provinces.

Type	Province	Class	AP	Non-AP	Precision	Recall	F1 Score	OA
With Training	Hebei	AP	83	11	93.30%	88.30%	90.70%	92.7%
		Non-AP	6	134	92.40%	95.70%	94.00%
	Shandong	AP	876	103	91.10%	89.50%	90.20%	92.3%
		Non-AP	86	1412	93.20%	94.30%	93.70%
	Zhejiang	AP	201	27	91.40%	88.20%	89.70%	95.8%
		Non-AP	19	842	96.90%	97.80%	97.30%
Without Training	Liaoning	AP	1226	114	97.20%	91.50%	94.30%	92.8%
		Non-AP	35	704	86.10%	95.30%	90.40%
	Jiangsu	AP	845	93	90.00%	90.10%	90.00%	89.8%
		Non-AP	94	807	89.70%	89.60%	89.60%
	Shanghai	AP	47	14	88.70%	77.00%	82.50%	90.0%
		Non-AP	6	134	90.50%	95.70%	93.10%
	Fujian	AP	311	82	92.60%	79.10%	85.30%	91.0%
		Non-AP	25	773	90.40%	96.90%	93.50%
	Taiwan	AP	315	36	90.80%	89.70%	90.30%	92.8%
		Non-AP	32	566	94.00%	94.60%	94.30%
	Guangdong	AP	921	96	92.10%	90.60%	91.30%	94.1%
		Non-AP	79	1876	95.10%	96.00%	95.50%
	Guangxi	AP	158	18	92.40%	89.80%	91.10%	93.3%
		Non-AP	13	274	93.80%	95.50%	94.60%
	Hainan	AP	125	11	87.40%	91.90%	89.60%	96.1%
		Non-AP	18	598	98.20%	97.10%	97.60%

) shows that the accuracy between provinces with training samples and provinces without training samples is basically the same. This indicates that our training samples have high representativeness and exhibit strong generalization ability.

4.3. Ablation Experiments

To validate the impact of the proposed DFF (Dynamic Feature Fusion) module on model performance, we conducted a series of ablation experiments, comparing the classification performance of the model with and without the DFF module. We selected samples from the year 2023 to ensure the precision and reliability of the experimental results.

(

Table 5. Performance of DBCE-Net with and without DFF module.

Type	Class	AP	Non-AP	Precision	Recall	F1 Score	OA
DBCE-Net (with DFF)	AP	687	55	92.6%	88.6%	90.6%	93.1%
	Non-AP	88	1230	93.3%	95.7%	94.5%
DBCE-Net (without DFF)	AP	708	100	87.6%	91.4%	89.5%	91.9%
	Non-AP	67	1185	94.6%	92.2%	93.4%

) indicates that the introduction of the DFF module increased the model’s overall accuracy from 91.9% to 93.1%, an improvement of 1.2%, demonstrating the effectiveness of DFF in enhancing the overall performance of the model. Specifically, the DFF module improved the precision of the AP category from 87.6% to 92.6%, although the recall slightly decreased (from 91.4% to 88.6%). Nevertheless, the F1 score increased by 1.1%, indicating that the DFF module provides significant benefits in extracting aquaculture pond features.

4.4. Comparative Experiments with Single-Branch Structure

To verify the effectiveness of improvements made by DBCE-Net compared to traditional single-branch structure networks, we selected validation sample points from 2023 for comparative experiments. This comparative experiment aims to demonstrate the performance enhancements achieved by DBCE-Net. By comparing the performance of DBCE-Net and single-branch structure networks at the same validation sample points, we can directly evaluate the advantages of the improved network structure in terms of processing efficiency and accuracy.

(

Table 6. Confusion matrix of DBCE-Net and other single-branch networks.

Type	Class	AP	Non-AP	Precision	Recall	F1 Score	OA
DBCE-Net	AP	687	55	92.6%	88.6%	90.6%	93.1%
	Non-AP	88	1230	93.3%	95.7%	94.5%
UNet	AP	683	101	87.1%	88.1%	87.6%	90.6%
	Non-AP	92	1184	92.8%	92.1%	92.5%
U2Net	AP	674	66	91.1%	87.0%	89.0%	91.9%
	Non-AP	101	1219	92.3%	94.9%	93.6%
PSPNet	AP	624	71	89.8%	80.5%	84.9%	89.2%
	Non-AP	151	1214	88.9%	94.5%	91.6%
SegNet	AP	590	115	83.7%	76.1%	79.7%	85.4%
	Non-AP	185	1170	86.3%	91.1%	88.6%

) below presents the comparison results of DBCE-Net and other single-branch networks, revealing a significant increase in accuracy.

It is evident that the single-branch structure of networks results in misclassification of rivers and lakes due to insufficient multi-scale processing capability and loss of edge information, leading to decreased accuracy (Figure 9). These results indicate that DBCE-Net possesses significant advantages over traditional single-branch structures in handling boundary information and complex scenes, thereby enhancing overall segmentation performance.

5. Discussion

5.1. Comparison with Other Datasets

We further compared the results of this study with other 10 m resolution land cover datasets, including the FROM-GLC10 dataset [46], the CWaC dataset [47] the ESA land cover dataset [45], and the China Aquaculture Ponds V1 [48]. Although these datasets do not specifically classify aquaculture ponds, they are generally categorized as a type of water body in land cover classification systems. Therefore, by comparing the extraction results of water body categories in these datasets, we can indirectly evaluate their performance in identifying aquaculture ponds.

The comparison (Figure 10) reveals that while these datasets can generally extract the locations of aquaculture ponds, their accuracy is relatively low, with prevalent issues of misclassification and omission. Specifically, the ESA dataset exhibits blurred boundaries of aquaculture ponds, with closely connected ponds making it difficult to discern the shape and boundaries of individual ponds, thereby failing to effectively distinguish between aquaculture ponds and surrounding land, which is inconsistent with the expected performance of 10 m high-resolution data. The FROM-GLC10 dataset performs better in extracting aquaculture ponds, effectively distinguishing between aquaculture ponds and land, but fails to differentiate between dikes and aquaculture ponds. The CWaC dataset is relatively superior in pond boundary recognition but exhibits lower accuracy in extracting small aquaculture ponds. The China Aquaculture Ponds V1 dataset demonstrates an enhanced ability to effectively distinguish between salt fields and aquaculture ponds, achieving improved overall accuracy. Nevertheless, for certain small-scale aquaculture ponds, the dataset only managed to extract a minimal portion of these features, resulting in incomplete and less comprehensive extraction outcomes. In this study, the total area of aquaculture ponds in 2022 was calculated to be 8592 square kilometers, exceeding the area covered by the CAV1 dataset (7919 square kilometers). This larger area includes a more comprehensive extraction of small aquaculture ponds, demonstrating the reliability of our study. However, unlike the CAV1 dataset, our study did not exclude abandoned aquaculture ponds. Moreover, most of the aforementioned datasets effectively distinguish between salt pans and aquaculture ponds. Overall, compared to existing land cover datasets, the method proposed in this paper more accurately identifies the boundaries of aquaculture ponds, more thoroughly extracts small aquaculture ponds, and successfully distinguishes between salt pans and aquaculture ponds. These results indicate that the method proposed in this paper exhibits higher accuracy and discriminative capability in identifying aquaculture ponds.

5.2. Area and Spatial Distribution of Coastal Aquaculture Ponds in China

The spatial distribution of coastal aquaculture ponds, along with their national and provincial regional delineations for the year 2023 (Figure 11A). These aquaculture ponds are extensively dispersed along the Chinese coastline, covering a total area of 8290.52 km². Coastal aquaculture ponds are densely distributed in the following four regions: Bohai Bay, coastal plains of Jiangsu Province, Beilin Bay of Guangdong Province, and the Pearl River Delta (Figure 11A). These areas account for over 70% of the total coastal aquaculture ponds in China. Shandong has the largest area of coastal aquaculture ponds (2271.30 km², 27.4%), followed by Guangdong (1205.18 km², 14.5%), Jiangsu (1094.28 km², 13.2%), and Liaoning (1092.78 km², 13.2%) (Figure 11B,C). Collectively, these four provinces account for 68.3% of the total area of coastal aquaculture ponds in China.

5.3. Area and Spatial Distribution of Coastal Aquaculture Ponds in China

This study detailed the spatial changes of coastal aquaculture ponds in China from 2017 to 2023, offering an important perspective on the spatial dynamics of coastal aquaculture. During this period, the total area of coastal aquaculture ponds exhibited a continuous downward trend, decreasing from 8970.25 km² to 8261.17 km² (Figure 12C), resulting in a net loss of 708.83 km², a decline of 7.90%. Additionally, existing studies reveal that before 2016, the area of coastal aquaculture ponds in China generally exhibited an expansion trend. Nevertheless, this study reveals a significant reversal in the trend, starting around 2017, wherein the area dedicated to aquaculture ponds has begun to decline. This finding represents a critical inflection point in the spatial dynamics of coastal aquaculture development in China.

We calculated the annual change rate (ACR) for each province in the entire coastal region (Figure 11A). The changes in the aquaculture area varied significantly across different provinces, highlighting the differences in aquaculture development strategies, natural conditions, and policy orientations among regions. The reduction in the aquaculture area was most significant in Shanghai Province, decreasing by 124.41 square kilometers, with a reduction rate of 38.92%. This was followed by Zhejiang and Jiangsu, which decreased by 31.57 square kilometers and 219.60 square kilometers, with reduction rates of 31.57% and 19.07% (Figure 11B), respectively. The changes observed in Jiangsu’s coastal areas are primarily attributed to natural coastal erosion and the implementation of the “Three Zones” policy along with land reclamation initiatives, with the most severely affected regions experiencing an average annual shoreline retreat of approximately 80 m, in Zhejiang Province, ecological protection efforts have led to the establishment of several wetland protection areas and large-scale projects to convert high-pollution, illegally constructed aquaculture ponds back to farmland. According to the Shandong Fisheries Statistical Yearbook released by the Shandong Provincial Department of Agriculture in recent years (http://nync.shandong.gov.cn/, accessed on 1 November 2024), the trends in the aquaculture pond area align closely with those calculated in our study. The area of aquaculture ponds are basically consistent with the trend we have calculated. In contrast, the aquaculture area in Liaoning exhibited an increasing trend, which is also reflected in the Liaoning Province National Economic and Social Development Statistical Bulletin (https://tjj.ln.gov.cn/, accessed on 1 November 2024), indicating a yearly increase in the area of aquaculture ponds.

As reported in the Fisheries Statistics Bulletin issued by the Ministry of Agriculture and Rural Affairs of the People’s Republic of China since 2017 (http://www.moa.gov.cn/, accessed on 20 October 2024), the area of aquaculture ponds in China has been declining in recent years. This observed trend confirms our research findings, thus confirming the reliability of our study. In addition, data show that although the aquaculture area decreased, the decline in aquatic product production in coastal areas was relatively small. This indicates that the structure of aquaculture ponds in coastal areas of China has been optimized, resulting in a significant increase in yield per unit area.

5.4. Limitations of the Method and Future Directions for Improvement

This study employed a pixel-based semantic segmentation approach [49], which was extensively utilized in aquaculture pond extraction due to its capability to directly classify each pixel into specific categories. However, compared to object-based methods, pixel-based segmentation is often prone to challenges such as over-segmentation and under-segmentation. In this study, under-segmentation is particularly prevalent due to the complex shapes and spatial distributions of aquaculture ponds, resulting in multiple adjacent ponds being erroneously merged into a single entity.

In contrast, object-based methods leverage spectral [27,50,51], spatial, and contextual information to group pixels into meaningful objects, effectively mitigating issues of over-segmentation and under-segmentation. Nevertheless, these methods, including object-based image analysis (OBIA), often require additional preprocessing steps and manual parameter tuning, which increases computational complexity and limits their scalability for large-scale automated applications.

Despite these limitations, this study adopted a pixel-based semantic segmentation method due to its simplicity and its suitability for large-scale automated analyses, particularly when applied to moderate spatial resolution datasets, such as Sentinel-2 imagery. Future research directions could explore the integration of object-based methods or hybrid approaches to address the segmentation challenges encountered. Future research could address these challenges by incorporating higher-resolution remote sensing data, such as PlanetScope [52], WorldView imagery [53], and the Gaofen series [54], to enhance the precision of pond boundary delineation, thereby overcoming the limitations imposed by spatial resolution on segmentation performance.

6. Conclusions

This study proposes a dual-branch deep learning model to improve the accuracy of identifying coastal aquaculture ponds in China. By integrating a Context-Enhanced Supplementary Branch (CESB), the model enhances the recognition of ponds with varying shapes and sizes while reducing misclassification of rivers and lakes. Since 2017, aquaculture pond areas have declined by 10.05%, reflecting policy adjustments and ecological protection efforts. The model demonstrates strong generalization capabilities, effectively adapting to diverse temporal and spatial data with limited training samples. Compared to traditional single-branch models, CESB more accurately captures contextual and boundary features in broader scenes. Beyond its applicability to China’s coastal ecological monitoring, this method has global potential, enabling detailed aquaculture pond mapping and spatiotemporal change analysis worldwide. This research offers new insights into sustainable coastal ecosystem development, emphasizes the environmental impact of aquaculture pond changes, and provides valuable references for policy making, ecological conservation, and interdisciplinary studies, advancing resource management and environmental protection.