Mapping the distribution and dynamics of coastal aquaculture ponds using Landsat time series data based on U²-Net deep learning model

ABSTRACT

The extraction of coastal aquaculture ponds from remote sensing images is affected by the effect of ‘same object with different spectrums’ and ‘different objects with same spectrums’. The U²-Net deep learning network is able to capture more contextual information from different scales to solve this problem and is suitable for salient object detection. This study proposed a remote sensing information extraction model of coastal ponds based on U²-Net deep learning network, completed the remote sensing information extraction of coastal aquaculture ponds in Zhoushan Archipelago from 1984 to 2022, analyzed the spatiotemporal evolution of coastal aquaculture ponds in Zhoushan Archipelago. The results showed that the developed model was more accurate with a precision rate of 96.12%, a recall rate of 95.43%, and an F1-measure of 0.96. During the study period, the area of the aquaculture ponds in the Zhoushan Archipelago demonstrated an increasing trend, expanding from 471.21 hm² in 1984–3668.55 hm² in 2022. In addition, over half of the aquaculture ponds had been active for more than 10 years. The method developed in this study is capable of rapidly and accurately mapping coastal aquaculture ponds, and thus is significant for the management of marine resources and promoting sustainable development.

KEYWORDS:

• Coastal aquaculture ponds
• U²-Net
• deep learning model
• spatiotemporal patterns
• accuracy assessment

1. Introduction

The coastal aquaculture industry has become one of the fastest-growing food industries in the twenty-first century, driven by the vigorous development of the marine economy. Aquaculture ponds are the main production mode and have provided more than half of the total seafood output (Timi and Buchmann 2023; Chen et al. 2023; Sun et al. 2024). The coastal aquaculture ponds were usually constructed by building a dike to enclose part of the sea. Cutting down mangroves and transforming wetlands and lakes became the common development strategies for expanding the production scale of aquaculture (Hou et al. 2022). Although the development of aquaculture improved the economic situation and living conditions in coastal areas (Suo and Zhang 2015; Xu et al. 2020), it also caused ecological problems such as oil spills and harmful algal blooms (Abd-Elrahman et al. 2011; Ottinger, Clauss, and Kuenzer 2016; Zhang, Xiao, and Deng 2021). Global climate change caused by the deterioration of the ecological environment and high-intensity human activities have in turn led to significant changes in water area and water quality (Gentry et al. 2017; Noto et al. 2023). Lean management has played an important role in the development of the aquaculture industry (Higgins et al. 2013; Qiu, Shen, and Xie 2023). The first and foremost goal of this strategy was to accurately obtain relevant information concerning aquaculture facilities.

In the early stages, the statistical survey method based on entity detection was the most common information extraction strategy, although this was limited by technology. Surveys were accurate and technologically friendly, but time-consuming and labor-intensive (Hardin and Jensen 2011; Safikhani et al. 2022). Presently, the entity survey method is often used for accuracy verification based on the ground truth. Remote sensing has become an efficient method of earth observation (Guo et al. 2020; Bratic, Yordanov, and Brovelli 2023). Compared with traditional methods, remote sensing technology provides a larger observation scale (Chen et al. 2022a; Zhao et al. 2023a and 2023b), more convenient access (Wang et al. 2021), richer data expression (Peterson, Sagan, and Sloan 2020), and a more complete observation perspective that can better meet the requirements of rapid and accurate extraction of aquaculture pond information (Cheng et al. 2020; Peng et al. 2022a; Mahmood, Zhang, and Li 2023).

Coastal aquaculture ponds are artificial water bodies used for aquaculture and have significant water body characteristics and man-made traces, which are usually flooded by nearshore seawater during the breeding period and partially or completely dry up after harvest. There is the regular surface shape of artificially reclaimed coastal aquaculture ponds, especially rectangular surface shapes. Remote sensing information extraction of coastal aquaculture ponds is mainly based on geometric characteristics such as spectral characteristics and rectangularity of water bodies (Sun et al. 2020; Peng et al., 2022b). At present, commonly used remote sensing information extraction methods for coastal aquaculture ponds mainly include threshold segmentation method, region growing method, object-oriented classification method, machine learning method and deep learning method. The threshold segmentation method refers to using thresholds to divide the study area into target areas and non-target areas based on the feature values or attribute values (Wang et al. 2018; Bazi et al. 2022). Affected by the ‘same object with different spectrums’ and ‘different objects with same spectrums’ in complex geographical environments, it is difficult to find a universal segmentation threshold to clearly distinguish coastal aquaculture ponds from other types of water bodies (Mcfeeters 1996; Liu and Jezek 2004; Xu 2005; Raju and Neelima 2012; Duan et al. 2020). On the other hand, efficient relational models tend to have higher complexity and are more difficult to understand and derive from the perspective of thresholding value (Alimuddin, Sumantyo, and Kuze 2012; Sabjan et al. 2022). The region growing method is an extraction method based on feature similarity judgment, The region growing method can effectively avoid the problem of many holes in the extraction results of the target area (Espindola et al. 2006). At the same time, it also has a better extraction effect for coastal aquaculture ponds with clear boundaries. However, the calculation cost is high and it is seriously affected by noise. The choice of similarity measure is the key to the quality of model extraction results (Asokan and Anitha 2019; Jia et al. 2023). Object-oriented classification is an intelligent image analysis method, which uses objects as the basic judgment unit and also considers the spatial characteristics of objects, thus effectively avoiding the impact of noise on image classification (Hossain and Chen 2019; Cui et al. 2020; Meng, Fan, and Li 2021). The basic steps of the object-oriented classification method are to first segment homogeneous pixels to form objects or patches with small spectral differences, and then use a bottom-up merging strategy based on spectral and spatial characteristics to generate the final recognition result (Ren et al. 2019). Machine learning algorithms refer to classification algorithms that continuously iterate with experience and can be divided into supervised learning and unsupervised learning (Maxwell, Warner, and Fang 2018; Gladju, Kamalam, and Kanagaraj 2022; Shirmard et al. 2022). Supervised learning includes regression and classification. Unsupervised learning uses clustering logic to reduce errors through iterative and descending operations (Xia, Guo, and Chen 2020; Xu et al. 2023). From an image perspective, deep learning algorithms focus more on mining the differences between recognition targets and context in images (Zhang, Zhang, and Du 2016; Wang, Fan, and Wang 2022a). The convolution operation is used to obtain the characteristic expression of the recognition target to identify potential targets in various situations (Cheng et al. 2020; Wang, Zhou, and Fan 2022b). Research shows that using the deep learning network such as U-Net to extract coastal aquaculture ponds in remote sensing images has better extraction results than using object-oriented methods (Lu, Shao, and Sun 2021; Chen et al. 2022b).

In addition to being fast and convenient, the ability to support the long-term analysis of time series is also a feature of optical remote sensing, as such series may contain rich spectral information. Example platforms include Landsat and Sentinel-2 satellites (Xu and Gong 2018; Sun et al. 2023a). The steps involved long-term studies include the research model obtaining better accuracy verification results for images acquired in a certain year, then extending this to other years to obtain complete long-term series recognition results. Such long-term series observations ignore the coherence of the feature expression and the correlations between features in different phases, do not fully reflect the characteristics of the extraction target, and only extract the feature expression over a certain period (Xhardé, Long, and Forbes 2011; Yu et al. 2017; Moussa et al. 2019; Sun et al. 2023b). Another step in long-term series analysis is to construct research models based on multi-temporal samples, and a machine learning or deep learning algorithm is then used to derive the connections between the training data by mining rich data sources (Qin et al. 2020; Feizizadeh et al. 2021; Hu et al. 2022; Zou et al. 2022). Therefore, it is necessary to design a deep learning model to accurately extract coastal aquaculture ponds with different shapes, spectra, and states from different remote sensing images. U²-Net, a U-shaped deep network with two nested layers, can address the issues involved with the traditional models mentioned above (Qin et al. 2020; Batista et al. 2022; Zou et al. 2022). On the one hand, U²-Net is a two-level nested U-structure that is designed for salient object detection, which can be trained from scratch to achieve competitive performance. On the other hand, a novel ReSidual U-block capable of extracting intra-stage multi-scale features without degrading the feature map resolution is used on the bottom level, and there is a U-Net like structure that is filled by a ReSidual U-block at each stage on the top level (Qin et al. 2020; Fu et al. 2023; Sun et al. 2023). The U²-Net network has a good performance in identifying potential expressions, can obtain high accuracy without pretraining using a large amount of data and avoids the problem of insufficient model robustness caused by too large a time interval.

It is critical to explore the feasibility and effectiveness of the U²-Net model for extraction of coastal aquaculture ponds. The objective of this study was to extract information of coastal aquaculture ponds using the U²-Net model from remote sensing images and analysis the spatiotemporal characteristics. First, we developed a model based on the U²-Net deep learning model to extract aquaculture ponds in view of the uncertainty and lack of robustness of traditional extraction methods. Then, we obtained year-by-year extraction results from 1984 to 2022 to create a distribution map of the aquaculture ponds. Finally, we analyzed the evolution of the spatial and temporal patterns in the Zhoushan Archipelago during the past 40 years from quantitative and qualitative perspectives, and we examined the existence of aquaculture ponds to visualize their evolution and use in the Zhoushan Archipelago. The results of this study not only provide data support for the future development and planning of aquaculture ponds but also will aid in the transformation and upgrading of aquaculture ponds towards eco-friendly and sustainable development.

2. Study area and data sources

2.1. Study area

The Zhoushan Archipelago is situated on the outer edge of Hangzhou Bay in the East China Sea, backed by the economically developed Yangtze River Delta Economic Belt (Chen et al. 2019). This archipelago is positioned at the intersection of the Yangtze River, Qiantang River, and Yong River, and is also where the Taiwan Warm Current, Japan Cold Current, and Yellow Sea Cold Current converge. As a result, the marine fishery resources in the surrounding seawater are abundant due to significant disturbance (Wang et al. 2021). The climate surrounding Zhoushan Archipelago is typically subtropical with a mild and humid climate, abundant sunlight, and a long frost-free period throughout the year, making it well-suited for the growth of diverse biological communities (Chen et al. 2021; Zhang, Xiao, and Deng 2021). Figure 1 depicts the precise location of the Zhoushan Archipelago.

Figure 1. The location of the Zhoushan Archipelago, China.

Benefiting from a long coastline and abundant intertidal resources, which provide a unique geographical advantage, and the aquaculture industry has emerged as a significant economic pillar of the Zhoushan Archipelago. The breeding facilities of Zhoushan Archipelago include aquaculture ponds, floating rafts or cages, net cages, and purse selines, the aquacultural varieties in aquaculture ponds are Portunus trituberculatus, South America white shrimp, razor clam, bamboo shrimp and Arca subcrenata Lischke. For a long time, the development of fishery aquaculture is disorderly in Zhoushan Archipelago. In 2004, the ‘The regulations for seawall construction in Zhoushan City’ was issued, marking the aquaculture industry of coastal aquaculture ponds into the standardization stage. Thus, conducting research on the spatial distribution and evolution of these coastal aquaculture ponds holds great promise for advancing the aquaculture industry in the Zhoushan Archipelago.

2.2. Data sources

Landsat series satellite data have the advantages of multi-band observation, long time series, and open data. Landsat data have been widely used in the fields of land-use/land-cover, environmental monitoring, natural resource management, and agricultural production. Affected by the seasons, most farmers carry out farming between April and November every year. At this time, seawater is injected into the coastal aquaculture ponds, and the characteristics in the images are also obviously stable. Therefore, we select a single image with no clouds or very few clouds from April to November each year as the representative of that year. We collected remote sensing images in a total of 38 study periods of from 1984 to 2022. Because there were no images that met the requirements in 2012, the relevant information for that year was not included. The detailed information of remote sensing images used in the study is shown in Table 1.

Table 1. The detailed information of remote sensing images used in the study.

Satellite	Sensors	Imaging date	Spatial resolution	Satellite	Sensors	Imaging date	Spatial resolution
Landsat 5	TM	1984.05.09	30m	Landsat 5	TM	2002.02.13	30m
		1985.07.24				2002.10.02
		1985.11.20				2003.08.02
		1986.08.12				2003.11.15
		1986.08.19				2004.04.23
		1987.10.02				2004.08.04
		1987.10.09				2005.11.27
		1988.11.05				2006.04.20
		1988.11.12				2006.04.29
		1989.08.11				2007.07.21
		1989.12.10				2007.07.28
		1990.04.17				2008.04.18
		1990.08.14				2008.04.25
		1991.07.09				2009.10.14
		1991.07.16				2009.10.21
		1992.07.18				2010.12.20
		1992.07.27				2010.12.27
		1993.06.03				2011.05.20
		1993.06.28				2011.05.29
		1994.05.05		Landsat 8	OLI	2013.07.28
		1994.11.06				2014.12.22
		1995.08.05				2015.08.03
		1995.08.12				2015.08.12
		1996.12.13				2016.07.20
		1996.12.20				2016.07.29
		1997.10.20				2017.07.16
		1997.10.29				2017.07.23
		1998.07.03				2018.07.19
		1998.07.12				2018.07.26
		1999.09.08				2019.07.29
		1999.11.20				2019.08.07
		2000.05.14				2020.08.16
		2000.06.06				2020.09.10
		2001.07.20				2021.11.23
		2001.07.27				2022.04.28

Twenty-eight study periods from 1984 to 2011 were selected from the Landsat 5 Surface Reflection Tier 1 (Landsat 5 T1-SR) data, and 10 study periods from 2013 to 2022 were selected from the United States Geological Survey (USGS) Landsat 8 Level 2, Collection 2, Tier 1 (Landsat 8 T1-L2) data. These images have been atmospherically corrected using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS) or the Landsat Surface Reflectance Code (LaSRC) (Skakun et al. 2019; Kumar and Mehta 2023), and include cloud, shadow, water, and snow masks produced using the C Function of Mask (CFMASK) (Dashpurev et al. 2023; Zaghian et al. 2023) as well as a per-pixel saturation mask.

3. Methodology

In this study, we adopted the U²-Net deep learning model as the core component of the aquaculture pond information extraction. As shown in Figure 2, the model constructed in this study consists of the following four parts: (1) image preprocessing and dataset preparation, (2) extraction of aquaculture ponds based on the U²-Net deep learning model, (3) accuracy assessment, and (4) spatiotemporal pattern analysis.

Figure 2. Workflow of the study.

3.1. Image processing and data-set preparation

The sensors of the Landsat series of satellites can provide reflection data from the visible band (0.45 μm – 0.69 μm) to the infrared band (0.76 μm – 2.35 μm), such as the Landsat 5 TM sensor and the Landsat 8 OLI sensor. As the input of the deep learning framework, we still needed to synthesize the multi-band remote sensing images into color images. In order to find a suitable band synthesis combination, we selected 240 samples in the study area covering six types of ground features: vegetation, construction, water, bare land, tidal flats, and coastal aquaculture ponds (Figure 3a). We counted the remote sensing images used in this study and extracted the reflectance values of six bands from 240 samples. The aquaculture ponds were significantly different from other non-water body types in the Red band (0.63 μm – 0.69 μm), NIR band (0.76 μm – 0.9 μm) and SWIR1 band (1.55 μm – 1.75 μm) (Figure 3b). After trying various combinations, we chose the short-wave infrared band, the near-infrared band, and the red band for layer stacking. In the composite image, the aquaculture pond is clearly distinguished from other ground features; the boundary is clear, and there is a clear dividing line with the sea.

Figure 3. Location and spectral analysis of sample points.

To reduce the computational requirements, we cropped the acquired remote sensing images using a sliding window with a size of 512 pixels x 512 pixels, and the sliding steps in the horizontal and vertical directions of each window were 500 pixels. In this study, enlargement, reduction, rotation, and other image augmentation operations were used to expand the sample data set.

3.2. Extraction of aquaculture ponds based on the U²-Net deep learning model

The process of identifying objects that easily attract attention from images is called salient object detection (Qin et al. 2020; Peng et al. 2022; Sun et al. 2022). The earliest solution was a multilayer perceptron that was quickly replaced by fully convolutional networks (FCNs) with an up-sampling operation and skip connection structure to effectively combine local and global information (Yasir et al. 2023). The U-Net network has been proposed to adjust an FCN from two perspectives of feature extraction and feature fusion to focus on the details in the image (Falk et al. 2019). In terms of feature extraction, U-Net no longer continues the logic of the FCN, i.e. only using pooling to enlarge local information (Cheng et al. 2020). Simultaneously, in terms of feature fusion, the result of each pooled convolution of the U-Net is also used as the result of the corresponding size up-sampling operation for localization and splicing. When the input image is calculated via multi-layer convolution, the local information is no longer suitable for decomposition (Amrani, Bey, and Amamra 2022). U²-Net provides another solution to the combination of dilated convolutions that utilizes dilated convolutional groups with multiple dilation ratios only in the bottom two layers of the U-shaped structure (Qin et al. 2020). The construction of U²-Net is shown in Figure 4. With its clever model structure and superior generalization effect, the U²-Net deep learning model has been widely used in salient object detection, target recognition, image segmentation, edge detection and other problems. The unique structural design and fusion strategy enable the U²-Net model to completely extract the target area, while also ensuring that the edges are clear enough.

Figure 4. Construction of the U²-Net.

In the U²-Net model, the use of the overlap-tile strategy effectively complements the context information and alleviates the problem of insufficient training data, and the dilated convolution group in multiple relationships can effectively preserve the local information. Another improvement of the U²-Net is that it extracts more complete contextual information by building a residual structure similar to the U-Net structure. From Figure 4, the U²-Net model consists of three parts: encoding, decoding, and feature fusion. In the encoding step, the pooled convolution is replaced by a down-sampling operation to reduce feature loss. In the feature fusion step, the restoration resolution is used instead of localization to increase accuracy. The output of the U²-Net model is a saliency probability map. The local optimal solution strategy is used as the threshold for binarized segmentation. Finally, the binarized image is used in the extraction result for aquaculture ponds.

3.3. Accuracy assessment

Accuracy evaluation is an important part of testing the quality of an extraction model. We used the precision, recall, and F1-measure as metrics for the precision verification. These indexes are calculated as follows: (1) $\mathrm{Precision}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}},}$(1) (2) $\mathrm{Recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}},}$(2) (3) $F{1}_{\mathrm{measure}}={\displaystyle \frac{2^{\ast }\mathrm{Precisio}{n}^{\ast }\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}},}$(3) where TP is the number of correctly identified pond extraction result pixels; FP is the number of misidentified pond extraction result pixels, and FN is the number of unidentified pond pixels.

3.4. Spatiotemporal pattern analysis

Since the extracted result is a raster file, the raster image was converted into a vector image, and the results were then analyzed. In terms of time trend analysis, the annual aquaculture pond area was measured according to the identification and extraction results for the aquaculture ponds. By arranging the datasets of the aquaculture pond area in ascending order by year and performing linear regression, we obtained an equation for the area trend line during the study period. This equation could be used to analyze increases and decreases in the area values over the study period. In terms of spatial change analysis, we considered the 10 major islands of the Zhoushan Archipelago as a unit. In addition to discussing the changes in each island over the past 40 years, we also discussed the changes between islands. In addition, to explore the interannual operation mode of the culture ponds, we analyzed the existence of coastal aquaculture ponds.

Aslan established a set of simple rules for estimating the longevity of aquaculture ponds (Aslan et al. 2016; Aslan et al. 2021). The rules are based on the longest observation period, but the calculation results are easily affected by missing values in the observations. Specifically, the estimation results fluctuate when there are no observations in a certain year or when the observations are incorrect. As they are affected by the production cycle and the weather, the obtained remote sensing images may not necessarily contain aquaculture ponds that are in use. In particular, after the harvest period, the expression of traits in aquaculture ponds differs from that in aquaculture ponds during the production period. Thus, in the estimation, we began with the year of the first occurrence of aquaculture ponds and ended with the year of the last occurrence of aquaculture ponds as a basic rule. If there were two or more consecutive classification results of non-aquaculture ponds in the observation sequence, the sequence was divided into two observation years. The first period began with the initial occurrence of aquaculture ponds and ended with the first occurrence of non-aquaculture ponds. The second period began with the year of the second occurrence of aquaculture ponds and ended with the year of the last occurrence of consecutive aquaculture ponds.

4. Results and analysis

4.1. Aquaculture pond extraction results

We extracted the aquaculture ponds in the Zhoushan Archipelago from 1984 to 2022, and Figure 5 shows the extraction results for each year. From Figure 5, an exchange of the area of aquaculture ponds in the Zhoushan Archipelago was observed; this result is described in Section 4.3.

Figure 5. Extraction results for the aquaculture ponds in the Zhoushan Archipelago from 1984 to 2022.

4.2. Accuracy assessment

To verify the accuracy and effectiveness of the model and results, we designed a set of verification methods from the spatial and temporal dimensions. In the spatial dimension, we selected Liuheng Island, Daishan Island and Qushan Island in the Zhoushan Archipelago from the Landsat 8 image on April 8, 2022, and conducted accuracy evaluation experiments in these three areas. Based on the results (Figures 6–8), the overall accuracy of the extraction for these three islands was greater than 90% (Table 2). The results were accurate, the locations were precise, and there were no holes in the extracted blocks.

Figure 6. Extraction results for Qushan Island: (a) original image, (b) extraction result, (c) accuracy map, and (d) error matrix.

Figure 7. Extraction results for Liuheng Island: (a) original image, (b) extraction result, (c) accuracy map, and (d) error matrix.

Figure 8. Extraction results for Daishan Island: (a) original image, (b) extraction result, (c) accuracy map, and (d) error matrix.

Table 2. Accuracy assessment and visual interpretation of the results of coastal aquaculture pond extraction on Liuheng Island, Daishan Island, and Qushan Island.

Study area	Region number		Area (hm^2)		Precision	Recall rate	F-measure
	Ground Truth	Our	Ground Truth	Our
Liuheng Island	19	19	10.33	10.3	94.47	93.86	0.94
Daishan Island	7	8	5.80	5.94	91.1	93.18	0.92
Qushan Island	2	2	3.55	3.69	92.79	96.45	0.95
Average					92.79	94.50	0.94

In the temporal dimension, we selected four images of Landsat 8 from January to April 2022 for accuracy verification experiments in Liuheng Island (Figures 9 and 10 and Table 3). The results from January, February, March, and April 2022 were accurate, the locations were precise, and there were no holes in the extracted blocks. The extraction from February 2022 was relatively poor (Table 3). Although the positioning was accurate, the accuracy was relatively poor, and there were many missed areas.

Figure 9. Error matrix and random sample point classification results for Liuheng Island from January to April 2022.

Figure 10. Accuracy map of aquaculture pond extraction.

Table 3. Accuracy assessment and visual interpretation results of coastal aquaculture pond extraction on Liuheng Island from January to April 2022.

Study case	Region number		Area (hm^2)		Precision	Recall rate	F-measure
	Ground Truth	Our	Ground Truth	Our
Jan. 2022	19	19	10.27	9.75	94	95.32	94.66
Feb. 2022	16	13	4.88	4.46	81.16	82.67	81.91
Mar. 2022	20	19	9.16	8.5	92.33	94.18	93.24
Apr. 2022	19	19	10.46	9.79	94.85	93.62	94.23
Average					90.59	91.45	91.01

We conducted an accuracy assessment on all of the even-numbered year Landsat 8 images of the Zhoushan Archipelago from 1984 to 2022 (Table 4). The accuracy verification results were generally good; the average value of precision, the recall rate, and F-measures were 93.23, 97.76, and 93.97, respectively.

Table 4. Accuracy assessment of coastal aquaculture pond extraction in the Zhoushan Archipelago in all even-numbered years from 1984 to 2022.

Study case	Precision	Recall rate	F-measure	Study case	Precision	Recall rate	F-measure
1984	81.10	88.68	84.72	2004	95.26	95.72	95.49
1986	91.63	92.83	92.23	2006	93.35	95.26	94.30
1988	92.66	94.72	93.68	2008	92.20	94.56	93.37
1990	95.44	94.68	95.06	2010	94.75	95.82	95.28
1992	91.12	95.69	93.35	2014	96.67	95.14	95.90
1994	89.52	95.20	92.28	2016	95.00	94.57	94.78
1996	93.61	95.46	94.53	2018	95.57	96.06	95.82
1998	95.06	94.62	94.84	2020	95.90	95.59	95.75
2000	93.62	94.69	94.15	2022	96.14	95.43	95.78
2002	92.86	95.63	94.23	average	93.23	94.76	93.97

In general, the aquaculture pond extraction model based on the U²-Net was verified as accurate, and the assessment results were accurate and thus could provide a foundation for the follow-up analysis.

4.3. Spatiotemporal analysis of aquaculture ponds in the Zhoushan Archipelago from 1984 to 2022

4.3.1. Temporal analysis

We calculated the area of the aquaculture ponds in the Zhoushan Archipelago over the past 40 years in ascending order of time and constructed a trend line (Figure 11). As shown in the figure, the change in area of the aquaculture ponds in the Zhoushan Archipelago from 1984 to 2022 was 3197.35 hm², exhibiting an overall increasing trend. The area of the aquaculture ponds increased from 471.21 hm² in 1984–3668.55 hm² in 2022, with an average annual change of 84.14 hm²/a and an average annual growth rate of 5.55%. The area reached a peak in 2004. The development process of the aquaculture ponds in the Zhoushan Archipelago can be divided into two stages. From 1984 to 2004, the aquaculture industry was vigorously developed within the archipelago, and the area of the aquaculture ponds increased significantly, with an average annual increase of 188.67 hm²/a and an average annual growth rate of 11.62%. From 2005 to 2022, the development entered a stable period, with an average annual change of −32.01 hm²/a and an average annual growth rate of −0.81%.

Figure 11. Area of aquaculture ponds and change trend in the Zhoushan Archipelago from 1984 to 2022.

4.3.2. Spatial analysis

We analyzed the spatial distribution of the aquaculture ponds on the ten largest islands in the Zhoushan Archipelago, and the results are shown in Figure 12. From the figure, there were differences in the changes in the spatial pattern of the aquaculture ponds on the different islands. However, most of the islands had significantly increased aquaculture pond areas compared to the original areas. The specific changes are described below.

1. Zhoushan Island

Figure 12. Area of aquaculture ponds on ten large islands.

A spatial distribution map of the coastal aquaculture ponds and corresponding statistics was created. Figure 13 shows the extraction results and the change trend for the Zhoushan Island aquaculture ponds from 1984 to 2022. As shown in Figure 13a, the area of the aquaculture ponds on Zhoushan Island exhibited an initially increasing and then decreasing development pattern. The area increased from the initial 158.37 hm² in 1984, reached a peak of 833.69 hm² in 2013, and then decreased to 299.20 hm² in 2022. The average annual change in the increasing stage was 23.29 hm²/a, and the average annual percentage rate of change was 5.89%. The rate in the decreasing stage was −59.39 hm²/a, and the average annual rate of change was −10.76%.

Figure 13. Extraction results for aquaculture ponds on Zhoushan Island from 1984 to 2022.

In addition to the initial increase and subsequent decrease in area, the spatial changes of the aquaculture ponds tended to be more centralized (Figure 13b). In 1984, the ponds were only lightly distributed along the island's northern coast. In 2004, the second peak in aquaculture ponds accounted for about one third of the northern coast. Then, the aquaculture area began to shift to both ends of the island. By 2022, only the northwest end of the island and the east end were sparsely distributed.

(2)	Daishan Island

The area of aquaculture ponds on Daishan Island only fluctuated slightly after the rapid formation during the development period. Figure 14 shows the extraction results and change trend for the Daishan Island aquaculture ponds from 1984 to 2022. As shown in Figure 14a, from 1984 to 1988, the area of the Daishan Island aquaculture ponds increased rapidly, from 64.71 hm² in 1984–673.31 hm² in 1988, with an average annual rate of change of 152.15 hm²/a and an average annual percentage rate of change of 79.60%. From 1988 to 2022, the area decreased slightly, with an average annual rate of change of −6.03hm²/a and an average annual percentage rate of change of −1.06%.

Figure 14. Extraction results for aquaculture ponds on Daishan Island from 1984 to 2022.

The spatial distribution of the aquaculture ponds on Daishan Island exhibited a trend of centralization (Figure 14b). In 1984, the aquaculture ponds were scattered on the island, and then, aquaculture centers quickly formed along Niluo Mountain and the Daishan waterway on the south side of the island. In 2005, only the large-scale aquaculture area centered on Niluoshan and the northern coast of Qiulongshanzui remained.

(3)	Liuheng Island

The aquaculture pond area exhibited an increasing trend on Liuheng Island. Figure 15 shows the extraction results and change trend for the Liuheng Island aquaculture ponds from 1984 to 2022. The overall change trend in the area of the aquaculture ponds on Liuheng Island was increasing, from 37.53 hm² in 1984–868.89 hm² in 2022, with an average annual rate of change of 21.88 hm²/a and an average annual percentage rate of change of 8.62% (Figure 15a). Among the years, 1998–2004 were the years with the most pronounced changes, with an average annual rate of change of 100.94 hm²/a and an average annual percentage rate of change of 18.66%.

Figure 15. Extraction results for aquaculture ponds on Liuheng Island from 1984 to 2022.

The spatial distribution of the aquaculture ponds on Liuheng Island exhibited a trend of expansion to the southern coast (Figure 15b). The aquaculture ponds were initially distributed along the coast of the island. With the intensification of aquaculture, the small-scale aquaculture ponds disappeared on the northern coast, and the aquaculture ponds on the southern coast expanded significantly.

(4)	Jintang Island

The area of Jintang Island's aquaculture ponds increased; however, it eventually returned to the original level. Figure 16 shows the extraction results and change trend for the Jintang Island aquaculture ponds from 1984 to 2022 (Figure 16a). The area of the aquaculture ponds on Jintang Island did not change significantly. Among the extraction results for the past 40 years, the peak value was 138.47 hm² in 1994, and the area of the aquaculture ponds rarely exceeded 50.00 hm² in other years.

Figure 16. Extraction results for aquaculture ponds on Jintang Island from 1984 to 2022.

The main change areas on Jintang Island were Tai O on the west coast of the island and Nantouzui on the southeast corner (Figure 16b). In 1984, the aquaculture ponds on Liuheng Island were only sparsely distributed in Bandit's Bay and Lion's Mouth. Since 2002, two stable independent ponds appeared in Nantouzui on the southeast corner. Tai O's aquaculture ponds gradually receded.

(5)	Zhujiajian Island

The area of the aquaculture ponds on Zhujiajian Island also increased significantly. Figure 17 shows the extraction results and change trend for the Zhujiajian Island aquaculture ponds from 1984 to 2022. The area of the aquaculture ponds on Zhujiajian Island exhibited an increasing trend (Figure 17a). The area increased from 0 hm² in 1984–312.48 hm² in 2022, with an average annual rate of change of 8.22 hm² and an average annual percentage rate of change of 16.32%.

Figure 17. Extraction results for aquaculture ponds on Zhujiajian Island from 1984 to 2022.

The aquaculture ponds on Zhujiajian Island tended to converge towards the west coast in terms of the spatial distribution (Figure 17b). Originally located in the Waishanzui area on the southern part of the island, the area remained stable since the establishment of the aquaculture pond in 1994, reaching Dengfeng in 2005 and then gradually dying out. However, after the reclamation, large-scale development of aquaculture on the west coast of the island began in 2002. After 2005, the ponds began to gradually aggregate and were finally concentrated in the reclamation area in the west.

(6)	Qushan Island

The area of the aquaculture ponds on Qushan Island increased steadily. Figure 18 shows the extraction results and change trend for the Qushan Island aquaculture ponds from 1984 to 2022. The area of the aquaculture ponds on Qushan Island increased from 16.93 hm² in 1984–310.96 hm² in 2022, with an average annual rate of change of 7.74 hm²/a and an average annual percentage rate of change of 7.96% (Figure 18a). The first stage of the growth of the aquaculture ponds on Qushan Island occurred in 1984–2008, and the area of the aquaculture ponds fluctuated around 100 hm². The second stage of the growth of the aquaculture ponds on Qushan Island occurred during 2009–2022, and the area of the aquaculture ponds fluctuated around 250 hm².

Figure 18. Extraction results for aquaculture ponds on Qushan Island from 1984 to 2022.

The aquaculture ponds on Qushan Island exhibited a centralized spatial distribution. From 1984 to 1988, a small number of aquaculture ponds were distributed around the island (Figure 18b). After 1989, a center of coastal aquaculture ponds formed and gradually expanded in the bay in the middle of the south coast of the island and the bay in the middle of the north bank.

(7)	Taohua Island

The area of the aquaculture ponds on Taohua Island remained basically unchanged after rapid initial growth. Figure 19 shows the extraction results and change trend for the Taohua Island aquaculture ponds from 1984 to 2022. The area of aquaculture ponds on Peach Blossom Island increased (Figure 19a). Overall, the aquaculture ponds on Taohua Island only changed slightly after the initial rapid formation. The area increased from 0 hm² in 1984–168.64 hm² in 2022, with an average annual rate of change of 4.44 hm²/a and an average annual percentage rate of change of 14.45%.

Figure 19. Extraction results for aquaculture ponds on Taohua Island from 1984 to 2022.

The spatial distribution of the aquaculture ponds on Taohua Island remained unchanged (Figure 19b). The first aquaculture area started in 1988, with only slight expansion after the establishment of aquaculture on the northern coast of the island. Another aquaculture area remained unchanged since 1993 when the aquaculture was initiated in the bay on the west side.

(8)	Changtu Island

The change in the aquaculture ponds on Changtu Island was similar to that on Taohua Island. The area remained almost unchanged since its formation but has increased slightly in recent years. Figure 20 shows the extraction results and change trend for the Changtu Island aquaculture ponds from 1984 to 2022. The area of the aquaculture ponds on Changtu Island increased (Figure 20a). The area on Changtu Island has also changed slightly after the rapid initial formation. The area increased from 33.45 hm² in 1984–239.25 hm² in 2022, with an average annual rate of change of 5.42 hm² and an average annual percentage rate of change of 5.31%. During the study period, 1984–1985 was the first stage of rapid growth, and the area of the aquaculture ponds increased by five times compared to the original area. Then, the area entered a plateau period from 1985 to 2011, the area of aquaculture ponds remained at roughly 160 hm². The second period of rapid growth was 2011–2013, with an increase of 38.89%. The aquaculture pond area then remained at 240 hm².

Figure 20. Extraction results for aquaculture ponds in Changtu Island from 1984 to 2022.

The spatial distribution of the aquaculture ponds on Changtu Island remained unchanged, centered on two breeding sites in the middle of the island, and continued to expand around these two centers (Figure 20b).

(9)	Xiushan Island

The area of the aquaculture ponds on Xiushan Island decreased to almost half of the original size. Figure 21 shows the extraction results and change trend for the Xiushan Island aquaculture ponds from 1984 to 2022. The area of the aquaculture ponds on Xiushan Island exhibited a downward trend (Figure 21a). The area decreased from 51.15 hm² in 1984–21.31 hm² in 2022, with an average annual rate of change of −0.79 hm²/a and an average annual percentage rate of change of −2.28%.

Figure 21. Extraction results for aquaculture ponds on Xiushan Island from 1984 to 2022.

The aquaculture ponds on Xiushan Island moved northward (Figure 21b). The aquaculture ponds on Xiushan Island were mainly distributed on the north and south coasts. The area of the coastal aquaculture ponds on the north side remained unchanged, and the main change was the disappearance of the coastal aquaculture ponds on the south side.

(10)	Sijiao Island

The area of the aquaculture ponds on Sijiao Island has only increased slightly in recent years. Figure 22 shows the extraction results and change trend for the Sijiao Island aquaculture ponds from 1984 to 2022. The area increased from 0 hm² in 1984–114.555 hm² in 2022, with an average annual rate of change of 3.01 hm² and an average annual percentage rate of change of 13.29% (Figure 22a).

Figure 22. Extraction results for aquaculture ponds on Sijiao Island from 1984 to 2022.

The spatial distribution of the aquaculture ponds on Sijiao Island remained unchanged, and the ponds were mostly distributed in the southern bay of the island. After the reclamation project was completed, the aquaculture area expanded (Figure 22b).

There are many islands in the Zhoushan Archipelago, and the changes in the culture ponds on the different islands varied. The geographical distribution of the aquaculture ponds was mostly along the coast and lesser inland. The changes in the area of the aquaculture ponds were affected by many factors, including the geographical conditions, reclamation projects, and island positioning. Under the constraints of the geographical conditions, the expansion of the aquaculture ponds often did not include reclaimed land inland; instead, the ponds expanded into the ocean. Thus, most of the changes in aquaculture ponds on the islands, especially the changes involved for the islands with increasing trends (taking Liuheng Island, Qushan Island, Daishan Island, Dachangtu Island, and Sijiao Island as examples), were often due to reclamation projects. In addition, there were islands that entered a stable stage after the initial formation of the aquaculture ponds, and the aquaculture area did not change. The decreasing trends were influenced by the positions and policies of the islands themselves. Zhoushan City advocates and implements a highly-standardized intensive model for coastal aquaculture ponds.

4.3.3. Analysis of the existence of aquaculture ponds in the Zhoushan Archipelago

The analysis of the existence of aquaculture ponds is important for evaluating the ecological quality of aquaculture and for adjusting sustainable development strategy. We calculated the history and area of the coastal aquaculture ponds in the Zhoushan Archipelago. The results of the statistical analysis are shown in Figure 23. Chen et al. (2021) analyzed the changes in the coastline of the Zhoushan Archipelago and found that areas with large fluctuations have undergone different degrees of reclamation. After the completion of the projects, many areas are used for the construction of aquaculture ponds. This also demonstrates that in coastal areas where land resources are not abundant, reclamation is the main aquaculture pond construction method (Figure 23a). As shown in Figure 23b, the proportion of the coastal aquaculture ponds within the Zhoushan Archipelago existing for two years was the highest, with an area of 2018.16 hm². Nearly 50% of the coastal aquaculture ponds had a history of over 10 years. Nearly 30% of the coastal aquaculture ponds had a history of over 20 years, and nearly 15% of the coastal aquaculture ponds had a history of over 30 years. The longest persisting coastal aquaculture pond had an aquaculture history of over 38 years and an area of 53.55 hm². The aquaculture ponds over 33 years old were mostly distributed on Daishan Island, Liuheng Island, Changtushan Island, and Taohua Island. The larger islands had relatively more coastal aquaculture ponds and longer persistence. In addition, we also observed that most of the aquaculture ponds were located on the edges of the islands.

Figure 23. Analysis of existence for aquaculture ponds in the Zhoushan Archipelago.

5. Discussion

In this study, we focused on obtaining the spatiotemporal changes in the aquaculture ponds in the Zhoushan Archipelago, China, using the U²-net model. Although we achieved information extraction for the coastal aquaculture ponds, there are still aspects that can be improved.

5.1. Reliability and significance of the methodology

The information extraction of the aquaculture ponds based on remote sensing images is often affected by the phenomenon of homogenous foreign bodies. Different from the pixel-based classification method, the deep learning method uses prior knowledge to consider the pixel block as the unit. In addition, the methods differ from the general image-based classification methods’ use of the same idea for the methodology, instead emphasizing the context information and highlighting the target recognition. We utilized the support vector machine (SVM), U-Net, and U²-Net to compare the pixel-based classification and deep learning methods. Figure 24 shows the extraction results of the different classification methods, and Table 5 shows that the U²-Net had a higher recognition accuracy. The results obtained using the proposed method have a clear boundary and high precision.

Figure 24. The extraction results obtained using the SVM-based method, U-Net deep learning model-based method, and the U²-Net deep learning model-based method. The red areas represent the coastal aquaculture ponds.

Table 5. Accuracy evaluation of the SVM-based method, U-Net deep learning model-based method, and U²-Net deep learning model-based method.

Model	Precision (%)	Recall (%)	F1-measure
SVM	84.64	63.41	0.72
U-Net	88.32	82.81	0.85
U^2-Net	96.12	95.43	0.96

5.2. Error analysis

In this study, the key to evaluating the accuracy of the findings was the accuracy of the aquaculture extraction. In the present study, we treated the aquaculture pond identification problem as a salient object detection problem. Therefore, the quality of the image data easily affected the model recognition results. In the data training process, since the preparation of the training samples was primarily based on the results of the visual interpretation, it was difficult to avoid errors caused by the subjectivity of the observers. In the target detection process, the lack of saliency status of the target recognition object in the image to be recognized also affected the recognition accuracy. In addition, because the Landsat data used in this study have spatial-resolution of 30 m, some of the aquaculture ponds with a small breeding area had too few pixels in the remote sensing images, and the feature expression was unclear. In fact, if the area of real coastal aquaculture ponds is less than 900 m², it is almost impossible to observe it in Landsat remote sensing images, and the characteristic loss of many small coastal aquaculture ponds is serious. If higher spatial-resolution satellite remote sensing images such as Sentinel and Gaofen-2 are used, the minimum area of aquaculture ponds that can be observed can be reduced by at least nine times. It is of great help in terms of the accuracy of model identification, faithful reflection of real data, and even sample labeling. In addition to the data quality, the limitations of the model itself were also one of the factors that affected the recognition results. The U²-Net network relies on contextual information and object identifier saliency expression. This leads to poor generalization of the recognition results of some aquaculture ponds that are too small. Second, when there are multiple aquaculture ponds in the image (there are multiple salient targets in the image), the identification pressure of the small aquaculture ponds increases, thereby reducing the identification accuracy of the small aquaculture ponds.

5.3. Driving factors of change in aquaculture ponds

Aquaculture pond changes are influenced by natural factors and human processes. The Zhoushan Archipelago is rich in marine fishery resources. In the past, people chose fishing-based fishery production. With the continuous development of science and technology and the strengthening of environmental protection, people have begun to use the fishery production method of nearshore aquaculture, as it is more friendly to the marine ecological environment. With the strengthening of supervision, the high-yield and refined management model led many farmers to convert original low-yield salt fields into aquaculture ponds. The multi-temporal satellite remote sensing images show that reclamation projects contributed to the increase in the area of ⁣the aquaculture ponds.

6. Conclusions

In this study, the U²-net deep learning framework was used to extract the areas of aquaculture ponds in the Zhoushan Archipelago, and the spatiotemporal changes in the aquaculture ponds in the Zhoushan Archipelago were revealed. The main contributions of this study are as follows.

1. The method used in this study could accurately extract the information relevant to aquaculture ponds, and the average precision, recall, and F1-measure of the method were 96.12%, 95.43%, and 0.96, respectively.

2. Through the analysis, we found that the area of ⁣the aquaculture ponds increased from 471.21 hm² in 1984–3668.55 hm² in 2022, with an average annual rate of change of 84.14 hm²/a and an average annual growth rate of 5.55%.

3. Based on the analysis results for ten islands, Daishan Island, Liuheng Island, Sijiao Island, Zhujiajian Island, Taohua Island, Changtu Island, and Qushan Island exhibited clear increasing trends. Jintang Island and Xiushan Island exhibited clear decreasing trends. Zhoushan Island exhibited a trend of increasing initially and then decreasing. In addition, based on the existence map, the aquaculture ponds with a long existence were mostly distributed on the islands with clear increasing trends such as Daishan Island, Changtu Island, Taohua Island, and Liuheng Island.

The impact and mechanism of human farming activities on the surrounding environment are still unknown. From accurately determining the spatial location and existence analysis of coastal aquaculture ponds, it is possible to carry out soil and water quality analyses in the vicinity of the relevant areas and to reveal the mechanism of the impact of farming activities on the ecological environment for ecological prevention and restoration.

In the future, we will focus on (1) reducing the effect of mixed pixels to improve the area statistics of the remote sensing information extraction results of coastal aquaculture ponds; (2) combining the analysis with feature transformation methods such as the tasseled cap transformation to improve the model’s information extraction accuracy.

Acknowledgments

The authors would like to thank the editors and the anonymous reviewers for their outstanding comments and suggestions that helped to improve the technical quality and presentation of this manuscript. We also thank the United States Geological Survey (www.usgs.gov), the National Aeronautics and Space Administration (www.nasa.gov), and the Chinese Academy of Science (www.ids.ceode.ac.cn) for the free availability of Landsat remote sensing images. We thank LetPub (www.letpub.com) for linguistic assistance and pre-submission expert review.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available on request from the corresponding author.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China [grant nos 41971296, 42122009, 42171311, and 42271340], in part by the Zhejiang Province Pioneering Soldier and Leading Goose R&D Project [grant no 2023C01027], in part by the Public Projects of Ningbo City [grant no 2021S089], and in part by the Ningbo Science and Technology Innovation 2025 Major Special Project [grant nos 2021Z107 and 2022Z032].