Nation-Scale Mapping of Coastal Aquaculture Ponds with Sentinel-1 SAR Data Using Google Earth Engine

Global rapid expansion of the coastal aquaculture industry has made great contributions to enhance food security, but has also caused a series of ecological and environmental issues. Sustainable management of coastal areas requires the explicit and efficient mapping of the spatial distribution of aquaculture ponds. In this study, a Google Earth Engine (GEE) application was developed for mapping coastal aquaculture ponds at a national scale with a novel classification scheme using Sentinel-1 time series data. Relevant indices used in the classification mainly include the water index, texture, and geometric metrics derived from radar backscatter, which were then used to segment and classify aquaculture ponds. Using this approach, we classified aquaculture ponds for the full extent of the coastal area in Vietnam with an overall accuracy of 90.16% (based on independent sample evaluation). The approach, enabling wall-to-wall mapping and area estimation, is essential to the efficient monitoring and management of aquaculture ponds. The classification results showed that aquaculture ponds are widely distributed in Vietnam’s coastal area and are concentrated in the Mekong River Delta and Red River delta (85.14% of the total area), which are facing the increasing collective risk of climate change (e.g., sea level rise and salinity intrusion). Further investigation of the classification results also provides significant insights into the stability and deliverability of the approach. The water index derived from annual median radar backscatter intensity was determined to be efficient at mapping water bodies, likely due to its strong response to water bodies regardless of weather. The geometric metrics considering the spatial variation of radar backscatter patterns were effective at distinguishing aquaculture ponds from other water bodies. The primary use of GEE in this approach makes it replicable and transferable by other users. Our approach lays a solid foundation for intelligent monitoring and management of coastal ecosystems.


Introduction
Aquaculture is an important source of food, nutrition, income, and livelihood for hundreds of millions of people around the world [1,2]. Driven by the increasing human population, the need to improve social-economic benefits and the escalating protein demands, aquaculture has been one of the fastest-growing food production sectors in the world over the past decades [3][4][5].

Sentinel-1 SAR data
Using the Google Earth Engine code editor, we retrieved all Sentinel-1 data (C-band Synthetic Aperture Radar with a frequency of 5.405 GHz), within 2020 (1 January-30 April), covering the entire study area. Sentinel-1 C-band SAR instruments support operations in single-polarization (VV and HH) and dual-polarization (VV + VH and HH + HV). In our study, we collected a total of 410 dualpolarized (VV + VH) images with the Interferometric Wide Swath and Ground Range Detected (GRD) format. Sentinel-1 data with the GRD format were detected, multi-looked, and projected to ground range using an Earth ellipsoid model, which has a spatial resolution of 5 m by 20 m with a ground sampling distance of 10 m [49]. All the Sentinel-1 images were collected by the GEE platform, which had been preprocessed using the European Space Agency's (ESA) Sentinel-1 Toolbox including orbit restitution, thermal noise removal, terrain correction, and radiometric calibration [50].
Instead of using the original images, the median images of VV and VH were calculated as a median value composite using the images from January to April, respectively. Compared with the original images, using median images could reduce the impact of speckle noise and enhance the permanent or stable low scatters. In addition, median images were able to differentiate aquaculture ponds from seasonal regulated water areas because the water level in aquaculture ponds was consistent throughout the year.

Training and validation datasets
We constructed two independent datasets including 3500 training polygons (i.e., 1200 aquaculture polygons and 1000 non-aquaculture polygons) and 1200 validation points (i.e., 600 aquaculture points and 600 non-aquaculture points) (Figure 3a,b). Training polygon datasets were used to obtain the class features and corresponding thresholds for the decision tree and validation point datasets were utilized to assess accuracy.

Sentinel-1 SAR data
Using the Google Earth Engine code editor, we retrieved all Sentinel-1 data (C-band Synthetic Aperture Radar with a frequency of 5.405 GHz), within 2020 (1 January-30 April), covering the entire study area. Sentinel-1 C-band SAR instruments support operations in single-polarization (VV and HH) and dual-polarization (VV + VH and HH + HV). In our study, we collected a total of 410 dual-polarized (VV + VH) images with the Interferometric Wide Swath and Ground Range Detected (GRD) format. Sentinel-1 data with the GRD format were detected, multi-looked, and projected to ground range using an Earth ellipsoid model, which has a spatial resolution of 5 m by 20 m with a ground sampling distance of 10 m [49]. All the Sentinel-1 images were collected by the GEE platform, which had been preprocessed using the European Space Agency's (ESA) Sentinel-1 Toolbox including orbit restitution, thermal noise removal, terrain correction, and radiometric calibration [50].
Instead of using the original images, the median images of VV and VH were calculated as a median value composite using the images from January to April, respectively. Compared with the original images, using median images could reduce the impact of speckle noise and enhance the permanent or stable low scatters. In addition, median images were able to differentiate aquaculture ponds from seasonal regulated water areas because the water level in aquaculture ponds was consistent throughout the year.

Training and validation datasets
We constructed two independent datasets including 3500 training polygons (i.e., 1200 aquaculture polygons and 1000 non-aquaculture polygons) and 1200 validation points (i.e., 600 aquaculture points and 600 non-aquaculture points) (Figure 3a,b). Training polygon datasets were used to obtain the class features and corresponding thresholds for the decision tree and validation point datasets were utilized to assess accuracy.

Ancillary datasets
We used some ancillary datasets to define our study sites and assist the classification. For example, we implemented a high spatial resolution (30 m) map (Credit: Laboratory of Digital Earth Sciences, Aerospace Information Research Institute, Chine Academy of Sciences) with a detailed coastline to mask sea and regions not included in the study area. We also used the digital elevation model (DEM) (acquired by the Shuttle Radar Topography Mission with 30 m spatial resolution) to  Training polygon datasets were acquired by visual interpretation from Sentinel-1 images. We first obtained water objects derived from Sentinel-1 images and then attributed their properties (aquaculture polygons or non-aquaculture polygons) by referencing the high-resolution optical images in Google Earth. Aquaculture ponds are mainly water bodies closed by dams, so they have obvious water characteristics. Due to their constituent and distribution characteristics, aquaculture ponds have obvious distinctions from other objects in the image including darker colors, regular shapes (mainly rectangular), and distribution in rich water areas. Based on the above features, the principles and rules used to pick up the training polygons were as follows: (1) the training polygons should be distributed evenly in the study area; (2) their sizes and shapes should be representative in the whole study area; (3) their locations should be representative such as located in estuaries, shorelines, and near rivers. Furthermore, expert experience and knowledge are also very important when selecting the training polygons. Based on the above criteria and methods, the training polygons sets were collected.
Moreover, validation datasets were obtained by stratified random sampling on high-resolution optical images from Google Earth for the accuracy assessment.

Ancillary datasets
We used some ancillary datasets to define our study sites and assist the classification. For example, we implemented a high spatial resolution (30 m) map (Credit: Laboratory of Digital Earth Sciences, Aerospace Information Research Institute, Chine Academy of Sciences) with a detailed coastline to mask sea and regions not included in the study area. We also used the digital elevation model (DEM) (acquired by the Shuttle Radar Topography Mission with 30 m spatial resolution) to remove the effects of hillshade, which may very likely lead to misclassification of aquaculture ponds because hillshade can induce similar returns as ponds in Sentinel-1 images.

Extraction of Potential Aquaculture Ponds
Since aquaculture ponds are mainly shallow water bodies, they can be easily confused with other water bodies including lakes, reservoirs, and rivers. Therefore, we conducted the mapping of aquaculture ponds by two steps: retrieving potential aquaculture ponds and then secondary classification to catch the "true" aquaculture ponds.
To find the potential aquaculture ponds from satellite images, we adopted the integration of object-based segmentation and decision tree classification. A water index specifically developed for Sentinel-1 dual-polarized bands (Sentinel-1 Dual-Polarized Water Index, SDWI [51]) was utilized to capture shallow water. Fully considering the signal difference between water and other objects in dual-polarized bands, SDWI can enhance the water body information and simultaneously eliminate the effect of soil and vegetation. SDWI (Equation (1)) can efficiently extract low-backscatter water bodies under complex conditions [51], and thus can be used to map aquaculture ponds. We computed a threshold on SDWI to define shallow water bodies. Since water and the other ground objects can create two distinct modes on the histogram of SDWI, we used the valley (SDWI = 0.3) between the two peaks in the histogram of training polygons as the threshold to separate water and the other objects. The pixels were classified as water body when SDWI was greater than 0.3 where VV and VH represent the pixel value in VV median images and VH median images. The logarithmic function is used to make the value of VV and VH more manageable and comparable for different ground objects. The connected component segmentation (CCS) algorithm was chosen for obtaining the objects of water bodies. The CCS algorithm approach detects objects by considering the connectivity of focal pixels with neighboring pixels (e.g., eight neighbors), where the connected pixels are merged into the same objects [52]. We then generated a total of six metrics for segmented objects representing the geometric and texture features (Table 1). Among the six metrics, five of them have been widely used to measure the geometry of objects including area, perimeter, shape index (SI), compactness, and extent ratio (ER) [53][54][55]. Since seasonal water bodies including paddy fields and salt ponds are generally characterized by rough surface with distinctive high backscatters in the dry season, the pixels of seasonal water bodies are normally coarser than that of aquaculture ponds. To distinguish aquaculture ponds from seasonal water bodies, we constructed and adopted the seasonal water sensitivity index (SWSI) computed with gray level co-occurrence matrix (GLCM) to enhance the seasonal features [56][57][58]. We used the (ee.glcmTexture) function in the GEE to obtain texture features. In this step, the kernel is a 3 × 3 square, resulting in four GLCMs with offsets (−1, −1), (0, −1), (1, −1), and (−1, 0). Next, the four directional bands for each feature were averaged to one band. Based on the function, the GLCMs of the median composition of the VV image and VH image were calculated. Finally, the sum average (Savg) of each GLCM was selected to construct the SWSI. Table 1. Geometric and texture features derived from water body objects.

Feature/Metric Description References
Area Number of pixels per object [53] Perimeter Approximates the contour as a line through the centers of border pixels using a 4-connectivity [53] Extent ratio Ratio of pixels in the object to pixels in the total bounding box [53] Shape index Border length feature of image object divided by four times the square root of its area [54] Compactness Degree to which an object is compact [55] SWSI The sum of the sum average of VV and VH median images [58] Finally, the features in Table 1 were used as input variables to classify potential aquaculture ponds with a decision tree ( Figure 4). The thresholds in the decision tree were obtained by training polygons (Table 2).

Secondary Classification
Based on the potential aquaculture ponds, we employed a secondary classification to exclude non-aquaculture objects including hillshade, discontinuous rivers, and isolated ponds, which have similar geometric and texture characteristics as aquaculture ponds.
First, we identified and excluded hillshade objects from potential aquaculture ponds ( Figure 5). Hillshade information was acquired using the hillshade function (ee.Terrain.hillshade) in the GEE. The input of function was DEM provided by the global SRTM Version 3.0 dataset with a spatial

Secondary Classification
Based on the potential aquaculture ponds, we employed a secondary classification to exclude non-aquaculture objects including hillshade, discontinuous rivers, and isolated ponds, which have similar geometric and texture characteristics as aquaculture ponds.
First, we identified and excluded hillshade objects from potential aquaculture ponds ( Figure 5). Hillshade information was acquired using the hillshade function (ee.Terrain.hillshade) in the GEE. The input of function was DEM provided by the global SRTM Version 3.0 dataset with a spatial resolution of 30 m. The default of 270 degrees and 45 degrees were set as the input of azimuth and elevation, respectively. The final hillshade map was obtained by threshold segmentation. In order to obtain the optimal threshold, a histogram was obtained by selecting some regions of interest ( Figure 6). The histogram showed an obvious unimodal feature, with a mean value (M) of 176.70 and a standard deviation (SD) of 33.76. Therefore, the threshold for extracting the hillshade map was determined by M ± 0.5SD (i.e., pixels with gray values between 159.82 and 193.58 were not hillshade, but the others with gray > 193.58 and gray < 159.82 were hillshade). We then used the hillshade map to remove hillshade from potential aquaculture ponds.

Secondary Classification
Based on the potential aquaculture ponds, we employed a secondary classification to exclude non-aquaculture objects including hillshade, discontinuous rivers, and isolated ponds, which have similar geometric and texture characteristics as aquaculture ponds.
First, we identified and excluded hillshade objects from potential aquaculture ponds ( Figure 5). Hillshade information was acquired using the hillshade function (ee.Terrain.hillshade) in the GEE. The input of function was DEM provided by the global SRTM Version 3.0 dataset with a spatial resolution of 30 m. The default of 270 degrees and 45 degrees were set as the input of azimuth and elevation, respectively. The final hillshade map was obtained by threshold segmentation. In order to obtain the optimal threshold, a histogram was obtained by selecting some regions of interest ( Figure  6). The histogram showed an obvious unimodal feature, with a mean value (M) of 176.70 and a standard deviation (SD) of 33.76. Therefore, the threshold for extracting the hillshade map was determined by M ± 0.5SD (i.e., pixels with gray values between 159.82 and 193.58 were not hillshade, but the others with gray > 193.58 and gray < 159.82 were hillshade). We then used the hillshade map to remove hillshade from potential aquaculture ponds.  In the process of the decision tree classification above, some discontinuous streams could be misclassified as aquaculture ponds due to their similar shape and spectral characteristics in SAR images (like A and B in Figure 7a). A morphological technique was then chosen to remove the discontinuous streams ( Figure 7). First, all potential stream objects were dilated to fill the gaps between subsections of streams by creating an inside and outside buffer for each object (Figure 7b). Then, we eroded the dilation objects by creating an inside buffer for each object (Figure 7c) and obtained the continuous streams (Figure 7d). The continuous streams could be easily excluded because of their specific geometric characteristics using geometric features such as the shape index.
The last step was to remove regular lakes, farm reservoirs, and impounding reservoirs from the potential aquaculture pond layer. Aquaculture ponds bisected by narrow and glittering dams or roads generally present an agglomeration distribution (e.g., A in Figure 8a), while most of the lakes and reservoirs are isolated (e.g., B and C in Figure 8a). Therefore, a step searching neighborhood features was used to remove the non-aquaculture objects from potential aquaculture pond objects because aquaculture ponds are generally located close to each other (Figure 8). Remote Sens. 2020, 12, x FOR PEER REVIEW 9 of 19 In the process of the decision tree classification above, some discontinuous streams could be misclassified as aquaculture ponds due to their similar shape and spectral characteristics in SAR images (like A and B in Figure 7a). A morphological technique was then chosen to remove the discontinuous streams (Figure 7). First, all potential stream objects were dilated to fill the gaps between subsections of streams by creating an inside and outside buffer for each object (Figure 7b). Then, we eroded the dilation objects by creating an inside buffer for each object (Figure 7c) and obtained the continuous streams (Figure 7d). The continuous streams could be easily excluded because of their specific geometric characteristics using geometric features such as the shape index. The last step was to remove regular lakes, farm reservoirs, and impounding reservoirs from the potential aquaculture pond layer. Aquaculture ponds bisected by narrow and glittering dams or roads generally present an agglomeration distribution (e.g., A in Figure 8a), while most of the lakes and reservoirs are isolated (e.g., B and C in Figure 8a). Therefore, a step searching neighborhood features was used to remove the non-aquaculture objects from potential aquaculture pond objects because aquaculture ponds are generally located close to each other (Figure 8).  In the process of the decision tree classification above, some discontinuous streams could be misclassified as aquaculture ponds due to their similar shape and spectral characteristics in SAR images (like A and B in Figure 7a). A morphological technique was then chosen to remove the discontinuous streams ( Figure 7). First, all potential stream objects were dilated to fill the gaps between subsections of streams by creating an inside and outside buffer for each object (Figure 7b). Then, we eroded the dilation objects by creating an inside buffer for each object (Figure 7c) and obtained the continuous streams (Figure 7d). The continuous streams could be easily excluded because of their specific geometric characteristics using geometric features such as the shape index. The last step was to remove regular lakes, farm reservoirs, and impounding reservoirs from the potential aquaculture pond layer. Aquaculture ponds bisected by narrow and glittering dams or roads generally present an agglomeration distribution (e.g., A in Figure 8a), while most of the lakes and reservoirs are isolated (e.g., B and C in Figure 8a). Therefore, a step searching neighborhood features was used to remove the non-aquaculture objects from potential aquaculture pond objects because aquaculture ponds are generally located close to each other ( Figure 8).

Accuracy Assessment
Accuracy assessment for aquaculture ponds was performed based on the validation sets. A confusion matrix was calculated and standard performance measures (i.e., overall accuracy, producer's and user's accuracy) were derived from the matrix [59].

Mapping and Validation of Aquaculture Ponds
Based on the validation sample sets, the user's, producer's accuracy (UA and PA) and the overall accuracy (OA) were obtained ( Table 3). The OA and Kappa coefficient were 90.16% and 0.8, respectively.

Accuracy Assessment
Accuracy assessment for aquaculture ponds was performed based on the validation sets. A confusion matrix was calculated and standard performance measures (i.e., overall accuracy, producer's and user's accuracy) were derived from the matrix [59].

Mapping and Validation of Aquaculture Ponds
Based on the validation sample sets, the user's, producer's accuracy (UA and PA) and the overall accuracy (OA) were obtained ( Table 3). The OA and Kappa coefficient were 90.16% and 0.8, respectively.  Figure 9 shows an example of the result of each step including processed images, segmentation, and classification results. Compared with the original SAR image (Figure 9a), the median image ( Figure 9b) can largely improve the representation of water because noise on single images has been removed. Based on median images, the water index SDWI was calculated ( Figure 9c) and has the capability to enhance the difference between water (high values) and other objects (low values), which corresponded to the two modes in the histogram (Figure 9d). Then, through a simple dip test, the threshold was obtained from the histogram of SDWI (SDWI > 0.3) to get the water body pixels. We then conducted the CCS algorithm to obtain the objects of water bodies (Figure 9e). Finally, through rule construction derived from training polygons (Figure 3a) and decision tree execution, potential aquaculture ponds were extracted (Figure 9f). Aquaculture ponds were then classified with a secondary classification based on the above results of potential aquaculture ponds. As shown in Figure 10, aquaculture ponds were largely clustered and occupied the water-rich areas along the rivers or the coast. Seasonal water bodies such as rice farms were Aquaculture ponds were then classified with a secondary classification based on the above results of potential aquaculture ponds. As shown in Figure 10, aquaculture ponds were largely clustered and occupied the water-rich areas along the rivers or the coast. Seasonal water bodies such as rice farms were much smaller in size and located away from rivers and coastal areas. Isolated water bodies such as lakes, reservoirs, and farmland reservoirs were also identified and distributed randomly.
Remote Sens. 2020, 12, x FOR PEER REVIEW 12 of 19 much smaller in size and located away from rivers and coastal areas. Isolated water bodies such as lakes, reservoirs, and farmland reservoirs were also identified and distributed randomly.

Spatial Distribution of Coastal Aquaculture Ponds
The distribution of coastal aquaculture ponds in Vietnam in 2020 showed very high spatial variability concentrated in the north and south (e.g., MRD and RRD) and very few in central regions (e.g., SCR) ( Figure 11). The total area of aquaculture ponds was 4262.94 km 2 , with the largest area in the MRD (3405.05 km 2 ), followed by SCR (257.89 km 2 ), RRD (224.33 km 2 ), NCR (177.94 km 2 ), SR (132.67 km 2 ), and NR (65.08 km 2 ). The proportional area of aquaculture ponds in each region is shown in Figure 12b. Aquaculture ponds concentrated most of the landscape in MRD and RRD, accounting for about 85.14% of total area, while they covered a small proportion (less than 10%) in other regions.
Among the six sub-regions, MRD had the largest area and density of aquaculture ponds (7.95 ha/km 2 ), and NCR had a minimum density of 0.42 ha/km 2 (Figure 12a). The density of NR, RRD, SCR, and SR was also small, about 0.63 ha/km 2 , 2.02 ha/km 2 , 0.54 ha/km 2 and 1.28 ha/km 2 , respectively.
We also examined the spatial distribution of aquaculture ponds along their distance to the coastline by computing the area of ponds in every 1 km buffer zone of the coastline (Figure 12c). Aquaculture ponds were mostly distributed within 15 km from the coastline, accounting for 70.88% of the total area, with highly varied pond area ranging from 701.22 km 2 to 63.79 km 2 . Between the distance of 15 km and 35 km from the coastline, the area of aquaculture ponds rapidly decreased and only covered 26.13% of the total area with an average pond area of 47.74 km 2 . In regions away from

Spatial Distribution of Coastal Aquaculture Ponds
The distribution of coastal aquaculture ponds in Vietnam in 2020 showed very high spatial variability concentrated in the north and south (e.g., MRD and RRD) and very few in central regions (e.g., SCR) ( Figure 11). The total area of aquaculture ponds was 4262.94 km 2 , with the largest area in the MRD (3405.05 km 2 ), followed by SCR (257.89 km 2 ), RRD (224.33 km 2 ), NCR (177.94 km 2 ), SR (132.67 km 2 ), and NR (65.08 km 2 ). The proportional area of aquaculture ponds in each region is shown in Figure 12b. Aquaculture ponds concentrated most of the landscape in MRD and RRD, accounting for about 85.14% of total area, while they covered a small proportion (less than 10%) in other regions. the coastline (distance greater than 35 km), aquaculture ponds were very few and the mean area of aquaculture ponds was only 7.26 km 2 , accounting for only 2.99% of the total area.  (a) Aquaculture pond density (ha/km 2 ) in the study area); proportional area of aquaculture ponds in different regions is shown in (b), and area of aquaculture ponds in every 1 km buffered zone from the coastline is shown in (c). Note: aquaculture density = aquaculture pond area/regional area.

Discussion
In this study, we proposed an approach for mapping aquaculture ponds at a national scale using Sentinel-1 SAR images with the Google Earth Engine (GEE) platform. The application in mapping aquaculture ponds in Vietnam showed high accuracy. Our result of aquaculture ponds showed the the coastline (distance greater than 35 km), aquaculture ponds were very few and the mean area of aquaculture ponds was only 7.26 km 2 , accounting for only 2.99% of the total area.  , and area of aquaculture ponds in every 1 km buffered zone from the coastline is shown in (c). Note: aquaculture density = aquaculture pond area/regional area.

Discussion
In this study, we proposed an approach for mapping aquaculture ponds at a national scale using Sentinel-1 SAR images with the Google Earth Engine (GEE) platform. The application in mapping aquaculture ponds in Vietnam showed high accuracy. Our result of aquaculture ponds showed the Figure 12. (a) Aquaculture pond density (ha/km 2 ) in the study area); proportional area of aquaculture ponds in different regions is shown in (b), and area of aquaculture ponds in every 1 km buffered zone from the coastline is shown in (c). Note: aquaculture density = aquaculture pond area/regional area.
Among the six sub-regions, MRD had the largest area and density of aquaculture ponds (7.95 ha/km 2 ), and NCR had a minimum density of 0.42 ha/km 2 (Figure 12a). The density of NR, RRD, SCR, and SR was also small, about 0.63 ha/km 2 , 2.02 ha/km 2 , 0.54 ha/km 2 and 1.28 ha/km 2 , respectively.
We also examined the spatial distribution of aquaculture ponds along their distance to the coastline by computing the area of ponds in every 1 km buffer zone of the coastline (Figure 12c). Aquaculture ponds were mostly distributed within 15 km from the coastline, accounting for 70.88% of the total area, with highly varied pond area ranging from 701.22 km 2 to 63.79 km 2 . Between the distance of 15 km and 35 km from the coastline, the area of aquaculture ponds rapidly decreased and only covered 26.13% of the total area with an average pond area of 47.74 km 2 . In regions away from the coastline (distance greater than 35 km), aquaculture ponds were very few and the mean area of aquaculture ponds was only 7.26 km 2 , accounting for only 2.99% of the total area.

Discussion
In this study, we proposed an approach for mapping aquaculture ponds at a national scale using Sentinel-1 SAR images with the Google Earth Engine (GEE) platform. The application in mapping aquaculture ponds in Vietnam showed high accuracy. Our result of aquaculture ponds showed the spatial heterogeneity of aquaculture ponds and provides significant information for monitoring and managing aquaculture and coastal ecosystems.

A Transferable Approach
Our classification framework could be generalized to map aquaculture ponds at a large scale in other areas. Our approach identified aquaculture ponds as objects with low radar backscatter values and a clustering pattern in space. Through capturing the surface roughness, radar backscatter images are efficient at recognizing water bodies. However, other water bodies can be easily misclassified as aquaculture ponds with only the backscatter information [41]. Through a refined process incorporating the geometric and texture metrics, our approach can exclude seasonal water bodies such as rice farms, river segmentations, lakes, and reservoirs. With the high spatial resolution of Sentinel−1 SAR images, we were able to identify very small ponds that had been neglected using coarser resolution images such as Landsat [5,14]. Our results can enhance the feasibility and application of SAR images for mapping aquaculture farms with a better understanding of the role of temporal images and the geometric features of radar backscatters. The approach presented is available via GEE code application and can be easily transferred to map aquaculture ponds in other regions, making it possible to map aquaculture ponds at regional and even global scales.
Several uncertainties should be considered when applying our approach to other regions. First, a longer time series data frame should be considered such as data covering the entire year from January to December. In this study, we only had access to the data from January to April in 2020 because most of the aquaculture ponds in coastal zones covered by water all year round in Southeast Asia such as Vietnam. Therefore, our datasets did not affect the mapping accuracy. Although most rice farms only conserve water part of the year, some irrigation reservoirs have water year-round and are easily misclassified as aquaculture ponds. Additionally, the misclassification could also lie in the temporal change that rice farms may have been converted to aquaculture ponds when the SAR data were acquired. In addition, the method for removing discontinuous rivers might fail where aquaculture ponds with loss of dividing dam clusters. Therefore, specific thresholds need to be changed in some special places. There might exist some other circumstances leading to the misidentification of aquaculture ponds that were not presented in this study.

Implications for Sustainable Management of Aquaculture and Ecosystem Conservation
We have witnessed the spatial heterogeneity of aquaculture ponds along Vietnam's coastal regions through the application of our approach. Further exploration of our mapping results can provide significant insights into the sustainable development of aquaculture, coastal area management, and protection of the coastal ecosystems. Current programs focus on global fisheries, for example, the FAO provides statistical data on the yield and export rate. However, limited attention is paid to the sustainable planning and management of the aquaculture ponds, likely due to the lack of statistical standards and methods. The explicit spatial distribution of aquaculture ponds at a large scale is a first crucial step toward supporting better decision-making for aquaculture management and ecosystem conservation. For example, our maps could provide the status of aquaculture along the coastal area. Aquaculture ponds have taken the space of the near coastal region in Vietnam, especially in the northern and southern coast (Figure 12), by replacing rice fields, mangroves, and natural water bodies [7,28,60]. Shifting from rice fields to aquaculture ponds (e.g., shrimp farms) could be a way to mitigate the risk of sea level rise and salinity intrusion. With the increasing demand of human nutrients and the government's hope to improve social and economic benefits, aquaculture is predicted to increase continuously. In this context, the untreated waste from aquaculture [9,60] and the damage on natural ecosystems such as replacing mangroves may push the coastal environment to its sustainable limitation and lead the ecosystem to another state that may be difficult to reverse, for example, harmful algal blooms and dead zones. Figure 13 shows that along the coastal line of Vietnam, the water quality is poor where there is a large area of aquaculture ponds. Therefore, the spatial distribution of aquaculture ponds is fundamental for the investigation of environmental problems and ecosystem conservation [5,12,27,61,62].

Conclusions
Since sustainable and efficient aquaculture management is one of the most important challenges for coastal ecosystem conservation, the accurate mapping of aquaculture ponds is a crucial first step toward understanding the spatial heterogeneity and supporting better decision-making for ecosystem conservation and aquaculture management. In this paper, we presented an approach to efficiently and accurately map aquaculture ponds over the coastal region of Vietnam. The approach relied on the use of radar backscatters from Sentinel-1 SAR images with a novel classification scheme. The workflow was implemented entirely on the Google Earth Engine platform to leverage the computation power and the ease of data-access. This application, based on Google Earth Engine, can be easily transferred and applied by potential users to other regions for the same purpose.
With this approach, we were able to map the aquaculture ponds with very high accuracy along the coastal region of Vietnam. The successful implementation of our approach indicates that median radar backscatter with the dual-polarized (VV + VH) is efficient for extracting water bodies. However, the accurate mapping of aquaculture ponds needs more information such as geometry and texture to remove other objects with a similar return on the radar backscatter. Our approach is designed to fill the need for accurate mapping of aquaculture ponds and can be applied from a local to regional scale.

Conclusions
Since sustainable and efficient aquaculture management is one of the most important challenges for coastal ecosystem conservation, the accurate mapping of aquaculture ponds is a crucial first step toward understanding the spatial heterogeneity and supporting better decision-making for ecosystem conservation and aquaculture management. In this paper, we presented an approach to efficiently and accurately map aquaculture ponds over the coastal region of Vietnam. The approach relied on the use of radar backscatters from Sentinel-1 SAR images with a novel classification scheme. The workflow was implemented entirely on the Google Earth Engine platform to leverage the computation power and the ease of data-access. This application, based on Google Earth Engine, can be easily transferred and applied by potential users to other regions for the same purpose.
With this approach, we were able to map the aquaculture ponds with very high accuracy along the coastal region of Vietnam. The successful implementation of our approach indicates that median radar backscatter with the dual-polarized (VV + VH) is efficient for extracting water bodies. However, the accurate mapping of aquaculture ponds needs more information such as geometry and texture to