Coastline extraction using remote sensing: a review

ABSTRACT

Coastlines are important basic geographic elements and mapping their spatial and attribute changes can help monitor, model and manage coastal zones. Traditional studies focused on the accuracy of extraction methods and the evolution characteristics of coastlines. Thanks to the advances in remote sensing for earth observations, recent coastline extraction studies can reveal detailed ocean-land interaction changes. In this review, we aim to identify key milestones in coastline extraction using remote sensing by associating the emergence of major research topics with the occurrence of multiple application fields, multiple data sources, and multiple algorithms. Specifically, we define coastlines that can be applied to different application fields, summarize the characteristics of coastline products, and analyze the principles, advantages and disadvantages of methods. On this basis, we discussed the development direction and the challenges involved. This study provides practical insights that can be incorporated into the future development of coastline extraction approaches using remote sensing technologies.

KEYWORDS:

• Coastline
• remote sensing
• image analysis
• review
• implementation

1. Introduction

The coastline is the dividing line between the ocean and the mainland, and is one of the basic elements of topographic maps and charts; it is also recognized by the International Commission for Geographic Data as one of the 27 surface elements (Chen et al. 2019, 2019; Gens 2010; Yang et al. 2022). The areas along coastlines represent the transition regions between land and sea, the distribution center of many materials and energy, and strong coupling regions of various processes (e.g. physical, chemical, biological, and geological) (Liu et al. 2013a; Hou et al. 2020; Mujabar and Chandrasekar 2013). Influenced by land, ocean, atmosphere, and human activities, coastlines have unique geographic and dynamic characteristics; and their position, trend, and form change constantly (Bera and Maiti 2019; Gens 2010; Ghosh, Kumar, and Roy 2015; Wang et al. 2017, 2019b; Wu et al. 2020).

Coastlines have important ecological functions and resource value (Angulo, Lessa, and de Souza 2006). With improved transportation, large population growth, rapid economic development, and increasingly prominent geographic importance, the areas along coastlines have become one of the regions with the most frequent and intense human activities (Bera and Maiti 2019; Ghosh, Kumar, and Roy 2015; Konko et al. 2020). More than half of the world’s population lives within the range of 60 km from a coastline, and two-thirds of cities with a population of more than 2.5 million are located near tidal estuaries (Green et al. 1996; Liu et al. 2019; Mimura 2013; Small and Nicholls 2003). However, rapid economic development and limited land resources have led to significant changes in the natural ecological environment of coastal areas (Vassilakis and Papadopoulou-Vrynioti 2014; Zhang and Hou 2020). Since the 20th century, the economic centers of the world’s coastal countries have shifted to coastal areas (Mimura 2013; Small and Nicholls 2003). With the construction of various coastal projects and the rapid development of the regional economy, the natural attributes of coastlines have been reduced rapidly, and their original production capacity and ecological function have changed significantly (Bell, Bird, and Plater 2016; Green et al. 1996). From 1980 to 2015, China’s coastline increased by 3,000 km while its natural coastline decreased by 50% (Wang, Hou, and Wang 2017; Wu, Hou, and Xu 2014). Therefore, many scholars have studied how to objectively understand the evolution characteristics of the temporal and spatial patterns of coastlines, and quantitatively evaluate how they are impacted by human activities and how they dynamically respond to human disturbance.

Remote sensing technologies can obtain a large amount of information regarding the spatial structures and detailed textures of ground objects from airborne and satellite platforms and are used in many applications, including urban monitoring, disaster management, agriculture, and forestry (Im 2020; Im, Park, and Takeuchi 2019; Klemas 2012; Yang et al. 2022, 2023). Several remote sensing techniques are commonly used to detect and monitor coastlines (Boak and Turner 2005; Chen et al. 2019, 2019; Nandi et al. 2016). Traditionally, coastlines were extracted using photogrammetry (Chen et al. 2014, 2019, 2021). The launch of satellites with sensors in the visible and infrared electromagnetic spectra, such as MODIS, Landsat TM/ETM+/OLI, Sentinel 1/2FormoSat-2, SPOT, IKONOS, Quickbird, CBERS, and Gaofen-1/2, has offered an alternative to field measurements (Bell, Bird, and Plater 2016; Chen et al. 2019; Chu et al. 2013; Gens 2010; Jia et al. 2021; Li and Gong 2016; Wang et al. 2019b; Wu et al. 2020; Zhang et al. 2014). The development of airborne SAR interferometry and airborne laser ranging has allowed overcoming the limitations of optical electromagnetic spectra to delineate coastlines under cloud cover (Bell, Bird, and Plater 2016). Placing high-definition cameras along the coasts also improves the real-time performance of coastline monitoring (Angnuureng et al. 2016; Bracs et al. 2016).

Previous reviews of coastline extraction using remote sensing technology have mainly focused on existing indicators, methods and spatio-temporal characteristics, providing few details about development and future research direction. With the diversification of observation platforms and advancement of sensors, remote sensing data sources are increasingly used for coastline detection. In the current study, our primary goal is to review the current status of the technologies for coastline extraction by using remote sensing imagery. First, we provide a definition of coastline. Then, products and data sources are summarized. Finally, the methodologies for coastline extraction using remote sensing technologies are described in detail, clarifying their advantages and disadvantages and their development and future research direction are discussed.

2. Definition of coastline

A coastline is an important geomorphological type in coastal zones, providing valuable resources for human survival and development (Vos et al. 2019). As the dividing line between land and sea, the coastline changes position with the ebb and flow of the tide. The definitions and locations of a coastline in different research fields and applications is shown in Figure 1. In physical geography, the coastline is the upper boundary affected by the ocean’s storm surge. In marine management and geomatics, the coastline refers to the boundary between sea and land at the mean high water surface of spring tides, also known in oceanography as the mean higher high water (MHHW). In remote sensing, a coastline is the water boundary at the moment of remote sensing imaging and is also known as instantaneous water. From the perspective of shipping safety, a coastline is defined as the boundary between sea and land at the mean low water surface of spring tides, also known in oceanography as mean higher low water. With the regular movement of the tide, a boundary exists between sea and land at the mean low water surface of low tides, also known in oceanography as mean lower low water (MLLW).

Figure 1. Definitions and locations of a coastline in different research fields and applications.

In Oxford Learner’s Dictionaries, “shoreline,” “coastline,” and “waterline” mean the edge of the sea, the ocean, or a lake; the land along a coast; and the level that the water reaches along the side of a ship, respectively. None of these terms reflects the dynamic and professional characteristics of the land-water boundary. The term “shoreline” is used by the coastal research community whereas the term “waterline” is used in water transportation, and “coastline” is mainly adopted by the remote sensing community. Given the dynamic nature of idealized boundaries, various coastline indicators have been utilized in coastal studies. Therefore, the term “coastline” is used synonymously in this review article.

From the perspective of industry management and data acquisition, the coastline is divided into three categories in this study: instantaneous coastline, corrected coastline, and probabilistic coastline. The instantaneous coastline refers to the boundary between ocean and land at the moment of remote sensing imaging. It is usually instantaneous, and it changes with the tidal heights when imaging. The corrected coastline is obtained using a digital elevation model (DEM) and tide data on the basis of the instantaneous coastline, and can also be corrected for the effects of coastal processes. The corrected coastline is very stable and the MHHW, the mean low water surface of spring tides (MLLW), and 0 m mean high waterline can be determined according to its stability. A probabilistic coastline is a coastline predicted by a mathematical model based on a comprehensive understanding of the spatial position of the coastline and its changing rules. The probabilistic coastline is important in the analysis of coastal spatial resource evolution under the influence of global climate change and human activities.

3. Coastline products and performance evaluation

3.1. Coastline products

A benchmark product not only helps evaluating and verifying the performance of coastline extraction algorithms, but also performs a key role in the promotion of coastline research (Zhu et al. 2022). In recent decades, several coastline products (Table 1) have been published, and are used for coastal aquatic, ocean remote sensing, coastal zone management, coastal zone ecosystems and coastal geomorphology (Chen et al. 2023, 2023; Gens 2010; Li et al. 2023; Liang et al. 2023; Mahdavi et al. 2018).

Table 1. Comparison of the coastline production.

Name	Year	Study area	Spatial resolution/scale	Affiliation	References
WDB II (World Data Base II)	1977	Global	Approximately 1:3,000,000	CIA (Central Intelligence Agency), USA	(Central Intelligence Agency 1977)
WVS (World Vector Shoreline)	1990		Approximately 1:250,000	DMA (Defense Mapping Agency), USA	(Defense Mapping Agency 1988; Soluri and Woodson 1990)
DCW (Digital Chart of the World)	1992		Approximately 1:1,000,000	ESRI (Environmental Systems Research Institute), USA	(Digital Chart of the World, 1992)
GSHHS (Global Self-consistent, Hierarchical, High-resolution Geography Database)	1996 (first version), 2017 (version 2.3.7)		Approximately 100 m	University of Hawaiʻi, USA	(Wessel and Smith 1996 and 2017)
GSV (Global shoreline vector)	2018		Approximately 30 m	U.S. Geological Survey, USA	(Sayre et al. 2019)
GMSSD_2015 (global multiple scale shorelines dataset)	2019		Approximately 1 m	CAS (Chinese Academy of Sciences), China	(Liu et al. 2019)
Coastline of Southeast Asian islands	2000, 2015	Southeast Asian islands	Approximately 30 m	CAS (Chinese Academy of Sciences), China	(Zhang and Hou 2020)
Coastline of Antarctic	2005, 2010, 2017	Antarctic	Approximately 30–75 m	Beijing Normal University, China	(Yu et al. 2019)
Coastline of the United States	1930–1940 and 1970	the United States	The 1939–40 measurements were determined with a recording instrument on the largest-scale charts and maps available at that time. For the Great Lakes, shoreline lengths were measured in 1970 by the International Coordinating Committee on Great Lakes Basic Hydraulic and Hydrologic Data.	NOAA (National Oceanic and Atmospheric Administration), USA	(Ankrah, Monteiro, and Madureira 2022; Gens 2010)
	2003		Approximately 1:70,000	NOAA, USA	(Graham, Sault, and Bailey 2003)
	2003		The product accuracy is consistent with Shuttle Radar Topography Mission Water Body Data, Approximately 30 m.	NOAA, USA	(NASA/NGA 2003)
Coastline of Australia	1988–2019	Australia	Approximately 30 m	National Earth and Marine Observations Branch, Geoscience Australia, Australia	(Bishop-Taylor et al. 2021)
Coastline of China	1940, 1960, 1990, 2000, 2010, 2014	China	Topographic map with scale 1:250,000, 1:100,000, 1:50,000, remote sensing images with 30 m spatial resolution	CAS (Chinese Academy of Sciences), China	(Hou et al. 2016)
	1991–2015		Approximately 30 m	Tsinghua University, China	(Xu and Gong 2018)
	1975, 1990, 2000, 2010, 2015		Approximately 30 m	Beijing Normal University, China	(Wang et al. 2019b)

The earliest global coastline products can be traced back to 1990 (Soluri and Woodson 1990). In 1992, the United States, Canada, Australia, and the British Navy collaborated to publish a 1:1,000,000 scale digital map of the world (Digital Chart of the World, DCW), and the US Geological Survey converted it into an online ArcInfo vector data in 1994 (Digital Chart of the World, 1992; Liu et al. 2019). The Environmental System Research Institute, Inc. of the United State (ESRI) also published the ESRI version of global coastline and island data. In order to understand global changes, the US National Oceanic and Atmospheric Administration (NOAA) and several other research institutes obtained openly accessible global coastline data. The global coastline data were published in 1996 based on combining the WVS (World Vector Shoreline with working scale approximately 1:100,000) and WDB (World Data Base II with working scale approximately 1:3,000,000) databases (Wessel and Smith 1996). These data have been continuously updated during the past three decades, and the most updated version, 2.3.7 has been available since 15 June 2017 (Wessel and Smith 2017). A new 30-m spatial resolution global shoreline vector (GSV) was developed from annual composites of 2014 Landsat satellite imagery. Polygon topology was applied to the GSV, resulting in a new characterization of the number and size of global islands, and three size classes of islands were mapped: continental mainlands (5), islands greater than 1 km² (21,818), and islands smaller than 1 km² (318,868) (Sayre et al. 2019). Since the advent of Google Earth, the academic community has put forward an agenda to improve the spatial resolution of the global island and coastline data from high resolution satellite images. A high-resolution global coastline and island (isles, rocks) product with a background based on open Google Earth images for 2015 showed that the total area of the world of the global continental mainland is 137,126,029.55 km², the total area of global islands (isles, rocks) is 10,367,433.22 km², the global continental mainland coastlines are 734,739.56 km long, and those of islands are 1766,013.29 km long (Liu et al. 2019).

China’s “blue land” covers 3 million km². The total coastlines measure 32,000 km, with the mainland coastline spanning 18,000 km and with the island coastline measuring 14,000 km (Hou et al. 2016; Wu, Hou, and Xu 2014). China’s coastal zones are one of the fastest developing and most economically dynamic regions in the world. Since the 1950s, the coastline of mainland China has undergone drastic changes and has been continuously and significantly enhanced (Wang, Hou, and Wang 2017; Xu and Gong 2018). Based on multi-temporal topographic maps, remote sensing images and field surveys covering the entire coastal zone of mainland China, the coastlines of six periods since the early 1940s were extracted and analyzed in terms of their structure, fractals, change rates, land-sea patterns, and bay areas (Hou et al. 2016; Wang et al. 2019b; Wang, Hou, and Wang 2017; Xu and Gong 2018; Wang, Yan, and Su 2021). The results obtained by the visual interpretation-based method showed that the length of the coastline changed from 18.10 × 10³ km in the 1940s to 19.70 × 10³ km in 2014, an increase of 8.84% (Hou et al. 2016). The extraction of the instantaneous coastline using a modified normalized water index based on Landsat data from 1975 to 2015, found that the overall length is 17,364.90 km in 2015, which is 8.90% longer than 1975 (Wang et al. 2019a, 2019b, 2021). In addition, many researchers pay attention to the coastline changes in key areas, such as bays (Cai et al. 2022; Fan et al. 2023; Li and Damen 2010), big cities (Hu et al. 2021; Liu et al. 2018), starting point of the Belt and Road Initiative (Duan et al. 2021), archipelago (Chen et al. 2022), these results can contribute to the control and prevention of environmental pollution, optimization of resource configuration and sustainable socio-economic development of coastal zones.

The coastal zone is the center of the world’s economic activities and the main site of development and construction for all countries (Gens 2010; Hou et al. 2016). The coastline, as one of the important elements of the coastal zone, contains rich marine resources and environmental information while dividing the sea from the land, and it is necessary to monitor coastline changes to achieve the sustainable development of coastal zone resources (Li et al. 2023). The coastline products of the United States of America (USA), Australia and other countries were acquired by remote sensing images (Bishop-Taylor et al. 2021, Zhang and Hou, 2020; Gens 2010; Zhang and Hou 2020). For the USA, a number of federal projects distribute coastlines in digital format since 2003 using airborne LiDAR, coastal airborne spectrographic imager (CASI), data acquired for the Shuttle Radar Topography Mission (SRTM), and the “Digital Coast” project is a Federal-State local partnership to develop an information framework for addressing coastal and ocean issues since 2007 (Ankrah, Monteiro, and Madureira 2022; Gens 2010). For Australia, Bishop-Taylor et al. combined sub-pixel waterline extraction with a new pixel-based tidal modeling method to seamlessly map almost 2 million km of tide-datum coastline along the entire Australian coast from 1988 to 2019, present a new continental dataset documenting three decades of coastal change across Australia (Bishop-Taylor et al. 2021). Zhang and Hou extracted and assessed coastline changes on Southeast Asian islands overall during 2000–2015 based on Landsat remote sensing images, and found that the coastline length increased by 532 km from 2000 (148,508 km) to 2015 (149,040 km), natural coastlines decreased by 2503 km, while artificial coastlines increased by 3035 km (Zhang and Hou 2020).

3.2. Evaluation criterion

Accuracy assessments evaluate the performance of coastline extraction algorithms and analyze the accuracy of the results. Accuracy assessment is carried out by qualitative and quantitative evaluation. A qualitative evaluation is the assessment of an image through visual analysis with the support of feature knowledge; a quantitative evaluation is the assessment of an image using accuracy indices. A qualitative evaluation is conducted mainly on two factors: location and shape. In terms of positional accuracy, producer’s accuracy (PA), user’s accuracy (UA), and overall accuracy (OA) are calculated based on the extraction and reference results. It should be noted that since the coastline is a linear element, it is necessary to establish a buffer zone with 1 or 2 times the spatial resolution of the remote sensing image as the radius before positional accuracy evaluation. For non-positional accuracy, the root mean square error (RMSE) of the lengths of the coastline in the resulting images is calculated and compared with the length of the reference coastline.

3.2.1. Producer’s accuracy (PA)

PA refers to the proportion of correctly identified coastlines in the total number of reference pixels and is calculated as:

(1) $PA=\frac{O^{\prime }\cap O}{O},$(1)

where PA refers to producer’s accuracy and O’ and O are the extracted results and reference data (usually, benchmark product or visual interpretation results), respectively. The OE (errors of omission) refers to the reference sites omitted from the corrected class in the classified map that can be calculated according to formula 1-PA using PA.

3.2.2. User’s accuracy (UA)

UA refers to the proportion of correctly identified coastlines in the total number of pixels of the extraction result:

(2) $UA=\frac{O^{\prime }\cap O}{O^{\prime }},$(2)

The CE (errors of commission) refers to the reference sites that were misclassified from the correct class in the classified map that can be calculated according to formula 1-UA using UA.

3.2.3. Overall accuracy (OA)

Overall accuracy is the probability that an individual will be correctly classified by a test; that is, the sum of the true positives plus true negatives divided by the total number of individuals tested (Chen et al. 2019; Hou et al. 2022). The equation is as follows:

(3) $OA=\frac{TP+TN}{TP+TN+FP+FN},$(3)

where TP is the number of pixels that are the coastline in reference data and the coastline in extracted results, FP is the number of pixels that are the non-coastline in reference data and the coastline in extracted results, TN is the number of pixels that are the non-coastline in reference data and the non-coastline in extracted results, FN is the number of pixels that are the coastline in reference data and the non-coastline in extracted results.

3.2.4. Root mean square error (RMSE) and relative RMSE

The root mean square error (RMSE) is a popular metric used in machine learning and statistics to measure the accuracy of a predictive model (Chen et al. 2022; Cho et al. 2020; Sun et al. 2021). Relative root mean squared error (rRMSE) is a variation of the RMSE that helps evaluate the accuracy of a predictive model relative to the range of the target variable (Han et al. 2023). RMSE and rRMSE provide a clear understanding of the model’s performance, with lower values indicating better predictive accuracy (Kim et al. 2021; Shin et al. 2020). The equation of RMSE and rMSE are as follows (Han et al. 2023):

(4) $RMSE=\sqrt{\frac{1}{N}\sum _{i=1}^N{(y_i-\stackrel{\wedge }{y_i})}^2},$(4)

(5) $rRMSE(\mathrm{\%})=\sqrt{\frac{\frac{1}{N}\sum _{i=1}^N{(y_i-\stackrel{\wedge }{y_i})}^2}{\sum _{i=1}^N{(\stackrel{\wedge }{y_i})}^2}}\times 100=\frac{RMSE}{\stackrel{-}{y}}\times 100,$(5)

where y_i and $\stackrel{\wedge }{y_i}$ are the length of reference data and extracted results in ith section, respectively, $\stackrel{-}{y}$ is the mean value of extracted results, and N is the number of samples.

4. Data sources and methods

4.1. Data sources

A variety of data sources are available to examine the position and attributes of coastlines, and the choice of what data to use at a specific site is generally determined by data availability (Boak and Turner 2005). Historical land-based photographs, coastal maps and charts, which provide general background information, have become important data sources of coastline extraction. However, the limited information available on scale or ground control points limit their application to the quantitative mapping of coastlines. Over the last decade, remote sensing images have been the main data sources for large-scale coastline extraction, increasing the timeliness and reducing the cost of the assessments.

Several remote sensing data sources are commonly used to detect and monitor coastlines, including optical images, SAR (synthetic aperture radar) images and airborne LiDAR (light detection and ranging). Medium-resolution optical images have been widely available since the launch of the Landsat satellite in 1972. Recently, a number of commercial optical satellites with high spatial resolution such as IKONOS (Hamylton and East 2012), Quickbird (Ford 2013), RapidEye (Duarte et al. 2018), GeoEye (Dai et al. 2019), WorldView (Smith et al. 2021) and Gaofen-2 (Liang et al. 2021). The spectral continuity of optical images has the advantage of providing detailed spectral information to accurately distinguish surface objects with different characteristics. Optical images with high spatial resolution and suitable spectral characteristics can not only determine the position of coastlines, but also accurately identify their attributes. With the launch of the Seasat satellite in 1978, spaceborne SAR images became an option for coastline extraction, and Radarsat, ERS, Envisat, ALOS, and Gaofen-3 have provided good application results. The SAR signal echo depends heavily on the roughness of the imaged ground objects and a continuous coastline can be found with an edge-tracing algorithm. Compared to optical remote sensing images, SAR images can overcome adverse meteorological conditions, but their quality is affected by geometric distortion and speckle noise, making image interpretation more challenging. LiDAR is an airborne laser ranging technique that collects highly accurate point measurements along a dense profile (Brock et al. 2002). The coastlines derived from LiDAR data are referenced to the statistically established tidal datum surface. Therefore, LiDAR help overcome the problems associated with identifying coastlines with the use of wet/dry beach lines on aerial photographs. LiDAR data also provide detailed information about near-shore bathymetry and beach topography over a broad region. The cross-shore profile method fits LiDAR points along foreshore profiles, and water levels are intersected with the regression line to identify the coastline, and the points from profiles are then linked to represent a coastline (Fabris 2021; Morton, Miller, and Moore 2005; Gens 2010). Alternatively, the contouring method subtracts the desired tidal datum from the DEM generated by LiDAR data and values of zero are obtained for the water line, and these are contoured to derive the tidally referenced coastline (Robertson et al. 2004; Gens 2012).

4.2. Methods

Depending on the specific platform used, derived coastlines may be based on the use of visually discernible coastal features, digital image-processing analysis, or a specified tidal datum. Therefore, the methods can be divided into three categories: methods for instantaneous coastline, methods for corrected coastline, and methods for the probabilistic coastline. In principle, the different types of coastlines are obtained by different techniques/methods/models, and the accuracy of the coastline extraction depends on the spatial resolution of the data source. A comparison of the coastline extraction methods using remote sensing technology is described in Table 2, according to the research topic, technique/method/model, theories, data source, pros, cons, and accuracy.

Table 2. Comparison of the coastline extraction methods using remote sensing technology.

Type	Technique/method/model	Theories	Data source	Pros	Cons	Accuracy	References
Instantaneous coastline	Segmentation-based methods (region growth, snake model, edge detector, OTSU, Segnet, DeepUnet, histogram segmentation)	Image segmentation is used to distinguish land and sea based on the spectral differences between ground objects.	Aerial photography, optical satellite imagery, SAR imagery, Airborne LIDAR	Good spatial resolution Rapid and large-extent mapping In expensive Regularly updatable Can recognize the attributes of the coastline	Hard to be synchronized with tidal conditions Methods/models not necessarily applicable to complex coastline structures	Dependent on the spatial resolution of remote sensing imagery	(Bishop-Taylor et al. 2021; Buono et al. 2014; de Vries et al. 2021; Dewi et al. 2016; Kanwal et al. 2019; Kelly and Gontz 2018; Li and Damen 2010; Li and Niu 2022; Li et al. 2018; Meng et al. 2020; Pradhan, Rizeei, and Abdulle 2018; Sheng et al. 2012; Wang, Zhang, and Ma 2010; Wu, Liu, and Wu 2017; Yasir et al. 2020; Zhang 2011b; Zhang et al. 2013, 2021)
	Classification-based methods (decision tree classification, unsupervised/supervised classification, object-oriented classification)	The boundary between sea and land area is considered as a coastline according to the target characteristics.
	Hybrid methods (Object-based region growth integrated edge detection method, multiple segmentation methods, edge detection and supervised classification method, OMI-based method)	Combine the advantages of various technologies and methods to accurately distinguish land and sea.
Corrected coastline	Tidal correction-based method	Obtained by calculating the horizontal distance between the waterline at different satellite imaging moments and the waterline at the MHWS tides using the tidal datum and multi-temporal remote sensing images.	Tidal data, DEM	Synchronize with tidal conditions Clear meaning (coastal management)	Inconsistencies in tidal datum caused regional limitation	Dependent on the tidal model and DEM	(Liu et al. 2013; Ghosh, Kumar, and Roy 2015; Meng et al. 2019; Nandi et al. 2016; Zhao et al. 2017)
	Combined method of tide model and DEM	Obtained by calculating the horizontal distance between the waterline at the satellite imaging moment and the waterline at the MHWS tides using tidal data and DEM.
Probabilistic coastline	Probabilistic model	Based on the criterion, the trend of the coastline is predicted using the probabilistic model.	Land use/cover change product, socioeconomic data, sea level height data, meteorological/climatic data	Be beneficial to historical research Be beneficial to understand erosion and expansion in future	Limited precision and accuracy	Dependent on the probabilistic model and auxiliary data	(Amani et al. 2020; Dastgheib et al. 2022; Kroon et al. 2020)

4.2.1. Instantaneous coastline

Early coastline extraction efforts mainly involved photogrammetric methods, including visual interpretation (BaMasoud and Byrne 2011; Bini, Casarosa, and Luppichini 2021; Duarte et al. 2018; Hereher 2015; Jayakumar and Malarvannan 2016; Meyer, Matzke, and Williams 2015; Murali et al. 2015; Qiu et al. 2021; Suo and Zhang 2015; Xu, Gao, and Ning 2016; Zhang 2011a; Zhang and Hou 2020) and GPS survey (Moussa et al. 2019; Saranathan et al. 2011; Virdis, Oggiano, and Disperati 2012). However, the accurate visual interpretation of coastlines requires advanced expertise, and the process is time-consuming and labor-intensive. With the development of remote sensing technology, the available datasets include aerial photos and multi-source satellite remote sensing data from Landsat (Mondal et al. 2020), SPOT (El-Asmar and Hereher 2011), and other satellite imagery. For example, in Billa and Pradhan (2011), a Lee Sigma filter and a Sobel and linear edge detector were applied to SAR data for edge enhancement. Then, the shoreline was obtained through threshold segmentation. As the SAR imaging of coastal areas is affected by waves, some researchers have discussed the influence of wave growth on coastline extraction from SAR images (Sletten and Hwang 2011), and their results indicate that the detected waterline position is the sensitive function of radar frequency. Shore-based LiDAR has also been used in shoreline studies using geometric data such as instantaneous slope depths to extract shorelines (Xhardé, Long, and Forbes 2011, Liu et al. 2013; Abessolo Ondoa et al. 2016; Bracs et al. 2016; Al-Nasrawi, Hamylton, and Jones 2018; Lin et al. 2019; Ribas et al. 2020).

4.2.1.1. Segmentation-based methods

Image segmentation divides an image into several distinct areas according to image characteristics, such as grayscale, color information, spatial texture, and geometric shapes. These features should have maximum consistency or similarity in the same area while having obvious differences between different areas. Segmentation-based methods for coastline extraction can be categorized into 1) threshold-based methods, 2) edge detection-based methods, and 3) region-based methods.

4.2.1.1.1. Threshold-based methods

In threshold-based methods, a threshold for a specific image feature (including but not limited to gray value) is set, and the pixels are divided into different categories (usually water and land in coastline extraction research) by comparing them with the threshold. Therefore, feature and threshold selection represents a key process in threshold-based methods for image segmentation. Classic threshold segmentation methods include the maximum between-class variance method (OTSU), maximum entropy method, minimum error method, and cross-entropy method. Histogram-based threshold selection is another method (Chen et al. 2019; Sheng et al. 2012; Zhang 2011b). The valley point (Raju and Neelima 2012) between the dual peaks of an image gray histogram can be used as a segmentation threshold, and the image can then be binarized to extract the coastline. Kelly and Gontz (2018) calculated seven types of spectral indices and adopted the OTSU algorithm to determine the threshold of water and land separation. Objects on the surface of the Earth have characteristic spectral signatures due to their physical properties and their interactions with electromagnetic radiation, spectral indices derived from multispectral remote sensing products give some insights into different surface processes, such as the Normalized Difference Water Index (NDWI) (Buono et al. 2014; Li et al. 2018), Modified Normalized Difference Water Index (MNDWI) (Dewi et al. 2016; Kanwal et al. 2019; Karsli, Guneroglu, and Dihkan 2011; Pradhan, Rizeei, and Abdulle 2018; Wang, Zhang, and Ma 2010; Wu, Liu, and Wu 2017), Difference Vegetation Index (DVI) (Cenci et al. 2013; Murray et al. 2012), and tasseled cap transformation (Cenci et al. 2013; Chen et al. 2019; Zollini et al. 2019). Therefore, on the basis of calculating the spectral indices, image segmentation can be performed using different thresholds (determined by the OTSU algorithm or directly set to 0) to obtain a binary image and convert it into a vector file to complete the coastline extraction (Viaña-Borja and Ortega-Sánchez 2019).

4.2.1.1.2. Edge detection-based methods

Edge detection-based segmentation is a method for finding the edge pixels of water and land in an image where the pixel value change rate is the largest and is expressed by a derivative (gradient, $\mathrm{\nabla }f$) (Dharampal 2015). The edge can be specified with an edge detector or energy function-based method as following equations (Dharampal 2015; Ziou and Tabbone 1998). The equation is shown as following:

(6) $\mathrm{\nabla }f=\left[\frac{\mathrm{\partial }f}{\mathrm{\partial }x},\frac{\mathrm{\partial }f}{\mathrm{\partial }y}\right]$(6)

(7) $\theta (x,y)={tan}^{-1}(\frac{\mathrm{\partial }f}{\mathrm{\partial }x}/\frac{\mathrm{\partial }f}{\mathrm{\partial }y})$(7)

(8) $M(x,y)=\sqrt{{(\frac{\mathrm{\partial }f}{\mathrm{\partial }x})}^2+{(\frac{\mathrm{\partial }f}{\mathrm{\partial }y})}^2}$(8)

where $\mathrm{\nabla }f$ represents the direction of the maximum rate of change of f at position (x, y), $\frac{\mathrm{\partial }f}{\mathrm{\partial }x}$and $\frac{\mathrm{\partial }f}{\mathrm{\partial }y}$ are the derivatives in the X-axis and Y-axis, respectively, M(x, y) and $\theta (x,y)$ are the value and angle of the gradient $\mathrm{\nabla }f$, respectively.

Edge detectors

Filter templates are used to calculate the partial derivatives of gradients and are typically called edge detectors. The Sobel detector (Yu et al. 2019), linear detector (Billa and Pradhan 2011), and instantaneous coefficient of variation are common first-order differential operators while the Canny detector (Cenci et al. 2013; Liu and Jezek 2004) is a second-order differential operator. d_x and d_y are the common templates of the Sobel detector for horizontal and vertical edge detection, respectively.

(9) $d_x=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]d_y=\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right]$(9)

In addition, threshold-based segmentation is always combined with edge detection to provide a binary image for convenient edge determination (de Vries et al. 2021). In existing research, a threshold is selected on the basis of the global threshold constant false alarm rate (Bayram et al. 2013; Yu et al. 2019), histogram (Sheng et al. 2012), and adaptive threshold selection methods. For example, the software e-Cognition (Zhang et al. 2013) and FIT-COAST (Liu et al. 2020) provide object-oriented multiscale segmentation methods and mainly use the spatial geometric features of images (Kass, Witkin, and Terzopoulos 1988; Zhang et al. 2015).

Energy function-based algorithms

In coastline extraction research, energy function-based methods mainly refer to an active contour model and the developed algorithms based on such a model. The basic idea is to use a continuous curve to depict a target edge (such as a coastline) and make the curve gradually approach the actual target boundary under the guidance of the decreasing energy of a predefined energy function. The position of the curve at which the energy reaches the minimum is where the target contour is. These methods could be categorized into the parametric active contour model and geometric active contour model according to the different forms of the curves. In the parametric active contour model, the curve is represented by multiple parameters; the snake model (Caselles, Kimmel, and Sapiro 1997) and geodesic active contours (Alonso et al. 2010) are typical examples (Xu and Prince 1997).

Specifically, the snake model could not extract the position information of a concave edge, whereas the gradient vector flow (GVF) (Osher and Sethian 1988) is able to solve it. However, the iteration of the evolution equation of the GVF-based snake model depends on the performance of the edge detection operator and is unstable in some local areas. Hence, Sheng et al. (2012) proposed a controllable GVF for coastline extraction. The disadvantage of the parametric active contour model is that it is difficult to use when dealing with changes in curve topology, such as the merging or splitting of curves. Thus, a geometric active contour model (Silveira and Heleno 2009) with a level set function was proposed and has been successfully employed in shoreline extraction (Ouyang, Chong, and Wu 2010; Shu, Li, and Gomes 2010). Given its heavy computational burden, threshold-based coarse segmentation is first conducted, and a narrowband level set method is used to refine the results and thereby speed up coastline extraction (Modava and Akbarizadeh 2017). However, for areas with uneven pixel grayscale, the extraction effect is poor. To increase the calculation speed, a study proposed the changeable moving step method based on the level set algorithm and the resolution depressed method based on the level set algorithm for finding the variable moving step square (Kuleli 2010). In Crawford et al. (2020), a coastline extraction method using spatial fuzzy clustering and level sets was proposed, and the results indicated the method’s capability of achieving a precise positioning of coastlines on full-resolution images and avoiding mapping errors.

Region-based methods

The principle of region-based methods is to divide an image into different regions according to similarity criteria. That is, pixels that possess similar attributes are grouped into unique regions. The regional growth method starts with a group of seed pixels representing different areas and then gradually merges similar pixels around the seed to expand the area until merging pixels under predefined criteria or a stopping growth threshold is no longer possible. The similarity in region-based methods is usually measured in terms of gray-level intensity, variance, color, and multispectral features, Zhang (2011b) calculated the spectral Euclidean distance and then used the regional growth method to obtain the coastline in particular (Zhang et al. 2013).

4.2.1.2. Classification-based methods

Water features appear darker due to the strong absorption by the particular wavelength, and coastline positions can be mapped using classification-based methods (Hereher 2011; Sekar, Kankara, and Kalaivanan 2022; Zhang 2011b). Classification-based methods involve assigning different labels according to target characteristics and considering the boundary between sea and land area as a coastline, according to classification unit, which can be divided into two categories: pixel-based classification methods and object-oriented classification methods (Baselice and Ferraioli 2013; Gašparović and Jogun 2018; Ghassemian 2016).

Pixel-based classification methods

Pixel-based classification methods widely used methods that classify pixels into distinct categories based solely on the spectral and ancillary information within that pixel (Aly et al. 2012; Amani et al. 2020; Dervisoglu, Bilgilioğlu, and Yağmur 2020). With the advantages of simple operation, pixel-based classification methods is widely used in many fields including coastline extraction (Maiti and Bhattacharya 2011). Alesheikh, Ghorbanali, and Nouri (2007) built classification rules on the basis of spectral information, texture information, and other necessary statistical information of selected samples to separate sea and land, thereby achieving the purpose of extracting coastlines. In Dong et al. (2016), the coastline was extracted based on the classification of water and land by the maximum likelihood classifier using all bands of the Landsat satellite series (MSS, TM, ETM+), except for the thermal band, and further, the erosion and accretion on the coast of Cape Rosetta in Egypt from 1973 to 2008 were studied. In Du et al. (2016), the unsupervised fuzzy c-means classification (FCM) was proposed for coastline extraction, and the effect of the gradual transition between water and land was eliminated. Su and Gibeaut (2017) used hyperspectral data obtained by an unmanned aerial vehicle to compare the effects of ISODATA and three other supervised classification methods (i.e. maximum likelihood classification, random forest, and SVM) with texture and color information on coastline extraction. Cao et al. (2020) and Zimmerman et al. (2020) applied the Iterative Self-Organizing Data Analysis Technique (ISODATA) to classify water and non-water, and the boundary between the two groups was regarded as the coastline.

Pixel-based classification methods uses many classifiers, such as minimum distance classifier, maximum likelihood classifier, K-nearest neighbor classifier, K-means classifier, ISODATA classifier and SVM classifier etc., which have been successful in different fields or application scenarios (Lee et al. 2023; Li et al. 2022; Macarringue, Bolfe, and Pereira 2022; Yoo et al. 2019). However, it is well known that as arbitrary objects, pixels do not necessarily represent the pure landscape that is being characterized. In the absence of contextualization, pixel-based classification results tend to contain significant noise as no neighborhood information is being considered, which will result in false information appearing in the classification results, and thus reduce the accuracy of coastline extraction (Chen et al. 2019, 2022; Fisher 1997). Moreover, both the landscape structure and the resolving power of the sensor contribute to our ability to identify and map objects of interest in a map, but the presence of mixed pixels will affect the judgment of coastline position.

Object-oriented classification methods

Unlike pixel-based classification methods, object-oriented classification technology sets adjacent pixels as objects to identify interested spectral elements, and makes full use of spatial, texture and spectral information of high-resolution panchromatic and multi-spectral data to segment and classify, and outputs high-precision classification results or vectors (Kass, Witkin, and Terzopoulos 1988; Kotaridis and Lazaridou 2021). The object has more features than a pixel, and these features include color, size, shape, direction, proximity, and scale topology relations, which help to distinguish different types of feature more accurately. For object-oriented classification methods, the study area of remote sensing data is divided into multiple scales, and the optimal segmentation scale of objects in the study area is selected based on the result of multiple scales, the objects containing coastline is classified, and the coastline is extracted by vectorization or edge detection. In the previous study, many scholars used object-oriented classification methods to obtain coastlines, such as support vector machine (SVM) algorithm with a radial kernel, a rule-based object-oriented classification method (Liu et al. 2020; Pradhan, Rizeei, and Abdulle 2018).

Object-oriented classification generally includes segmentation, feature extraction, image classification, and accuracy evaluation (Aly et al. 2012). For coastline extraction using object-oriented classification methods, on the one hand, different objects have different scales in images, so it is difficult for a single segmentation scale to take into account both the macroscopic and microscopic features of images, and it is necessary to fully describe and express different types of ground objects at different scales. On the other hand, each object has the same or similar internal features, including spectrum, space, texture and shape, etc., it is difficult to determine the rules for accurately distinguishing particular objects.

4.2.1.3. Hybrid methods

Some studies combined methods to extract coastlines (threshold method, edge detection, or classification) (Cenci et al. 2013; de Vries et al. 2021; Kass, Witkin, and Terzopoulos 1988; Liu et al. 2020; Meng et al. 2020; Yu et al. 2019). For example, the SVM algorithm has been adopted to classify coastline pixels, and the object-based region growth integrated edge detection method was proposed to automatically extract coastlines (Zhang et al. 2013). More specifically, the eCognition software was used to conduct multiscale segmentation, and then a Canny edge detector was used to identify the edge pixels in the near-infrared band (the edges of the inland land and water were sharp). Thereafter, the ocean was identified using the area growth method with a new feature called the object merger index. Postprocessing was previously employed to delete unnecessary objects (lakes, ships, etc.) and optimize the coastline. In the work of de Vries et al. (2021), Landsat top-of-atmosphere reflectance data were used to calculate the NDWI, NDVI, and thermal infra-red. A Canny edge detector was used to distinguish vegetation, intertidal zones, water bodies, and subtidal silt. Then, the Otsu algorithm was implemented to automatically determine the threshold to achieve the separation of vegetation. The thermal infra-red threshold was employed to distinguish the water body from the intertidal zone. After obtaining several pure pixels of vegetation, water, and intertidal zone, linear spectral unmixing was performed to obtain the ratio of water, vegetation, and silt for each pixel and thereby achieve the boundary of subtidal mud banks. The linear spectral unmixing method assumes that the reflectance of each pixel is a linear combination of different objectives with different weight coefficients according to the proportion of the area occupied by each of them (Sun and Du 2019; Sun et al. 2020).

4.2.2. Corrected coastline

Either on the basis of visual interpretation and automatic interpretation technology, remote sensing images provide the instantaneous water boundary at the moment of sensor imaging. Many scholars have developed technologies to correct the coastlines to suit management applications (Liu et al. 2013; Zhao et al. 2017). For example, the remote sensing image at the time of the MHWS tides can be selected to extract water boundaries, but the process is difficult because of the regional differences in sea water movement processes (Nandi et al. 2016). In some cases, the tidal level value at the time of remote sensing imaging was identified, and then the slope relationship of the beach surface implied in the DEM and LiDAR was used (Ghosh, Kumar, and Roy 2015; Meng et al. 2019). The waterline position in remote sensing imaging can be calculated relative to the mean high water spring tide surface. The coastline correction methods can be divided into two categories: the tidal correction-based method and the combined tide model and DEM method.

4.2.2.1. Tidal correction-based method

The tidal correction-based method uses the tidal model and instantaneous waterline for coastline correction. First, the mid-perpendicular line of the instantaneous waterline is drawn to obtain a segmentation line, and the slope is calculated using the plane coordinates and height from the tidal datum plane. Second, the point position on the mid-perpendicular line at the MHWS tides can be calculated. Finally, connecting the points calculated on each mid-perpendicular line reveals the coastline at the MHWS tides. The diagram of the tide correction-based method is shown in Figure 2.

Figure 2. Diagram of the tidal correction-based method.

At the MHWS tides, the plane coordinates of a point at the mid-perpendicular line can be calculated by the following equation:

(10) $\left\{\begin{array}{c}X=\frac{H-h_{T1}}{h_{T1}-h_{T2}}(x_{T2}-x_{T1})+x_{T2}\\ Y=\frac{H-h_{T1}}{h_{T1}-h_{T2}}(y_{T2}-y_{T1})+y_{T2}\end{array}\right.$(10)

where X and Y are the coordinates of mean high water line, x_T1, y_T1 and x_T2, y_T2 are the coordinates of the intersection of the instantaneous waterline at times T1 and T2 and the mid-perpendicular line. h_T1 and h_T2 are the tidal levels at T1 and T2, respectively, H is the tidal level of the corrected coastline.

4.2.2.2. Combined method of tide model and DEM

Given the influence of tides, the horizontal distance from the waterline to the “coastline” is calculated according to the tidal height at the satellite imaging moment and the tidal height at the MHWS tides. The result is combined with the DEM data of the study area, and the corrected coastline (for the MHWS tides) is obtained.

The diagram of the combined tide and DEM method is shown in Figure 3.

Figure 3. Diagram of combined tide and DEM method. C₁ and C₂ are the waterlines obtained on the basis of the remote sensing data of different imaging times; the distance between C₁ and C₂ is set as △L.

Querying the tide survey data reveals that the satellite transit time and the mean high tide heights of the two remote sensing data are h1, h2, and H, respectively, where h2 > h1. Then, the correction distance L (relative to the waterline C2) of the coastline can be calculated by Equation (11)(11) $\left\{\begin{array}{c}L=\left(\begin{array}{cc}H-& h_2\end{array}\right)/tg\theta \\ \theta =arctg\left[\left(h_2-h_1\right)/L\right]\end{array}\right.$(11) .

(11) $\left\{\begin{array}{c}L=\left(\begin{array}{cc}H-& h_2\end{array}\right)/tg\theta \\ \theta =arctg\left[\left(h_2-h_1\right)/L\right]\end{array}\right.$(11)

where θ is the slope of the tidal flat, which can be calculated by the regional DEM.

For the corrected coastline obtained using the combined tide model and DEM method, the collected measured tidal level data should be used as input to the T-Tide Tidal Harmonic Analysis Toolkit (https://www.eoas.ubc.ca/~rich/) to obtain the tidal harmonic constants of different tidal sections (Pawlowicz, Beardsley, and Lentz 2002). Given the regional characteristics and spatial transformation differences of tidal levels, the main tidal sections in the study area are screened out. According to the tidal harmonic constants of the selected subtides, the tidal level values of the satellite imaging time and the mean high tide time are calculated on the basis of the mean sea level. Therefore, a key problem is to obtain accurate tidal harmonic constants while considering the density of measured tidal stations and the regional applicability of the tidal model.

Coastline corrections are mainly aimed at flat terrain areas and require high-precision, centimeter-level, DEM data to obtain slope information. Therefore, the acquisition of centimeter-level DEM data is another key problem in the realization of coastline tide correction.

4.2.3. Probabilistic coastlines

Probabilistic coastlines are a special type of coastline developed on the basis of waterlines. When the remote sensing image acquisition time varies, the coastlines extracted from various images, even those from the same area, are different. The land and water areas are divided according to a certain method for each scene image, and the probability of whether each pixel is a water area or land is determined. This method requires multiple consecutive observations of the same area.

Similarly to the waterline extraction process, a spectral index is widely adopted to distinguish water and land because of its practical features. For example, the NDWI was calculated from high spatial resolution remote sensing satellite datasets (Quickbird, WorldView-2, and WorldView-3) and then masked with an adopted threshold method to extract the coastline (Ford 2013). As the image acquisition time varies, the tide height also differs. Therefore, the extracted coastline also changes. Each pixel has a water probability, which refers to the ratio of the pixels classified as water to the total number of repeat satellite observations. If the probability of a pixel of water is greater than 50%, then it is classified as water, and the others are treated as land. The result is the final coastline. The same extraction process is also used, and differences may exist in the spectral index (such as MNDWI) or the threshold used to classify water and land (Amani et al. 2020).

Intertidal extent extraction is closely related to coastlines. The binary NDWI thresholding method was employed to delineate land and water for each interval of the observed tidal range, that is, the difference between the observed highest and lowest tidal heights (Alesheikh, Ghorbanali, and Nouri 2007). The aforementioned operations have been performed on 10 tidal height intervals, and a complete relative extent model was obtained. At the same time, the standard deviation of the NDWI (for each tidal height interval) was calculated, and then a confidence layer was obtained.

5. Evolution of coastal remote sensing

To visually display the changes in coastal remote sensing publications, we used the Publication Years module of Web of Science (WOS) in obtaining the annual publication of the core collection with the subject keywords “coastline” and “remote sensing” from 2000 to 2020. We also utilized the CiteSpace software to process the retrieved results. A total of 643 articles were obtained. The number of studies related to remote sensing in coastal zones indicates a change of “rising first, then falling, and then rising” since 2000, showing an overall upward trend, especially a significant increase after 2017, and reaching a maximum of 91 articles in 2020 (Figure 4). To reflect the proportion of coastal studies in the field of remote sensing, we retrieved the number of articles related to remote sensing under the same rule and obtained 77,174 articles in total. As shown by the yellow line in Figure 4, the proportion of remote sensing studies in coastal zones has remained at about 1% since 2000.

Figure 4. Annual publication of articles related to coastal zone remote sensing.

To obtain the development status of coastal zone remote sensing at each time stage, we used CiteSpace in analyzing the collinear atlas of keywords from 2000 to 2020. The node type was a keyword, the time slice was 1, and the selection criterion was TOP50%. That is, keywords were selected from the top 50% of the citation frequency of each time slice, the results as shown in Figures 5 and 6.

Figure 5. Frequency order of keywords in coastal remote sensing papers published between 2000–2020 (data from Web of Science).

Figure 6. Classification of keywords in coastal remote sensing papers published between 2000–2020 (data from Web of Science).

The keywords not only are the refinement of the research focus and direction of the publications, but also the summary of the core of the publications. As shown in Figures 5 and 6, coastal zone development can be divided into the following five stages according to the ranking and clustering of keywords and the analysis of hot spots in the literature year by year:

1. The keywords from 2000 to 2002 were “management” and “interferometry.” The research at this stage took SAR image data as the main data source. Some databases made through optical satellite remote sensing monitoring were applied to the protection of ecological systems and planning for various effective management strategies for managers.

2. The keywords from 2003 to 2008 were “GIS,” “mangroves,” “erosion,” and “beaches.” The research hotspots in this stage were mangroves and beaches in coastal ecosystems. Based on GIS technology’s multisource data integration capabilities, researchers were able to capture the decade-long decline and erosion processes of mangroves and beaches by using a large volume of information (Goffin et al. 2022; Mahdianpari et al. 2020; Rossetto 2023). Then, global forest development models, such as mangrove forests and coastline movement models, were proposed (Wang et al. 2019b; Yang et al. 2022 and 2022c).

3. The keywords from 2009 to 2013 were “GIS,” “dynamics,” “vegetation,” and “coastline change.” At this stage, few articles related to coastal remote sensing were published, and the research focus returned to the dynamic change of coastlines. Based on GIS, satellite long-time series data, field surveys, and other means, researchers constructed large-scale temporal and spatial dynamic models of estuaries and coastal estuaries (Hardy, Wu, and Peterson 2021; Li et al. 2019). Relative to those in the previous stage, the types and capabilities of modeling at this stage became rich and strong (Tien Dat Pham et al. 2021).

4. The keywords from 2014 to 2017 were “classification,” “impacts,” “GIS,” “sea level rise,” “extraction” and “Landsat.” As a result of the attention of the international community to the ecological environment, the research at this stage focused closely on the impact of human activities, such as economic development, on coastal wetland systems. Based on the development of Landsat 8 and other satellites and machine learning, the application of high spatial resolution images to supervise the classification and extraction of coastlines became a hot topic in this stage (He et al. 2022; Wulder et al. 2022).

5. The keywords from 2018 to 2020 were “impacts,” “climate change,” “sea level rise,” “GIS,” “Landsat” and “time series.” At this stage, the number of studies related to remote sensing in coastal zones gradually increased, and high-quality mangrove review papers were published. With the development of machine learning, such as neural networks and rich time phase data, coastal zone, mangroves, and port monitoring became research hotspots. The application of machine learning, cloud computing, and other technologies to remote sensing in coastal zones is expected to become a hot research topic in the future (Amani et al. 2020; Feng et al. 2022; Sornette and Wu 2023).

6. The keywords are classified into 10 classes (Figure 6). As can be seen from Figure 6 that there is a large overlap area among the clusters, indicating that they are closely, the top five are Landsat, accretion, GIS, random forest, and movement. Landsat ranked first, thus indicating that remote sensing in coastline extraction is closely related to Landsat satellites. The research directions mainly include coastline change, evolution erosion, impact, and management, the research objects mainly include mangrove, coastal zone, shoreline, and vegetation, the impact factors include climate change.

6. Development and future research direction

Coastal zone monitoring is an important task in sustainable development and environmental protection, and the coastline is one of the most important linear features on the earth’s surface, which has a highly dynamic nature. Over the last decade, a range of airborne, satellite, and land-based remote sensing techniques have become more generally available for coastline extraction. Multiple remote sensing platforms, multiple types of sensors, and advanced algorithms facilitate fine coastline extraction.

6.1 Application of multisource data fusion to coastline information extraction

Potential data sources for coastline extraction include historical photographs, coastal maps and charts, aerial photography, beach surveys, and remote sensing imagery derived from satellite platforms, and shore-based HD cameras, covering visible, multispectral, hyperspectral, SAR, LiDAR (Boak and Turner 2005; Meng et al. 2018). At present, coastline extraction leans toward the comprehensive application of multisource data, and methodologies have been developed to detect robust, fast, and objective coastline features (Chen et al. 2019, 2019; Gens 2010). Fusion technology can integrate spectral features, geometric features, texture features, scattering features, temperature features and terrain features effectively, enhance the interpretation of remote sensing images and suppress possible ambiguity, incompleteness, uncertainty, and errors (Li et al. 2022; Zhang 2010). The integration of optical, thermal infrared, radar, and LiDAR data can improve the spatial resolution, distinguish the boundaries between water bodies and other ground objects, and extract the accurate positions of coastlines. In addition, the fusion of multisource data will provide a great convenience for accurate coastline extraction in specific application fields, including marine, shipping, geography, economy, and environment.

6.2 Application of advanced algorithms to coastline extraction

Deep learning is a research hotspot in the field of artificial intelligence that utilizes multilayer convolutional neural networks (U-Net, U²-Net, SegNet, SeNet) to realize the abstraction and learning of target features and constructs a multilevel semantic hierarchical model from pixels to features and then to attributes (Kang et al. 2022; Tsiakos and Chalkias 2023). Deep learning methods have also been applied to coastline extraction (Seale et al. 2022; Zou et al. 2022). With the increase of data sources, the complexity of application scenarios leads to challenges in coastline extraction. Deep learning, driven by data, extracts useful information from data through multilayer learning and objectively mines possible relationships between data to improve data processing efficiency and accuracy and bring new opportunities for the intelligent analysis and mining of coastline information in complex geographical environments.

6.3 Challenges of large-scale and rapid coastline mapping

Many scholars have focused on coastline extraction and the spatio – temporal evolution analysis of local regions and have thus proposed several methodologies (Mutanga and Kumar 2019). However, traditional methodologies are often difficult to use when dealing with the collection, storage, and processing of massive data in global coastline mapping. Remote sensing cloud computing platforms, such as the Google Earth Engine with massive remote sensing data and powerful data computing capabilities, and deep learning technologies with powerful feature expression capabilities facilitate rapid global coastline mapping.

7. Conclusions

With the rapid development of remote sensing satellites with “high spatial, high temporal, and high spectral” resolution, it has become the development trend of coastline extraction quickly and accurately over large areas and over long periods. In the study, we reviewed and critically compared methodologies from different points of view, with particular emphasis on the definition of coastlines, coastline products and evaluation criterion, data sources and implementation of methodologies, and description of research focus.

Due to the combined effects of intense human activities and global climate change, coastline have undergone dramatic changes in recent decades, affecting the entire population living in coastal areas, accounting for approximately 60% of the world’s population living in coastal areas. Although a number of techniques are available to accurately delineate the coastline from a variety of remote sensing data, but only regionally. However, comprehensive use of multi-source remote sensing data to carry out large-scale, long-time series and high-precision coastline mapping is still the direction of future efforts.

Disclosure statement

No potential conflict of interest was reported by the authors.

Data availability statement

The data, models, or codes generated or used in this paper are available upon reasonable request.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China [Grants No. 42122009, 42171311, 42271340, 41971296, 41801256, 42171326, 42176174], Zhejiang Province “Pioneering Soldier” and “Leading Goose” R&D Project [Grant No. 2023C01027], Zhejiang Provincial Natural Science Foundation of China [Grants No. LR19D010001, LQ18D010001], Fundamental Research Funds for the Provincial Universities of Zhejiang [Grant No. SJLZ2022002], Public Projects of Ningbo City [Grant No. 2021S089], Ningbo Science and Technology Innovation 2025 Major Special Project [Grants No. 2021Z107, 2022Z032], Open Fund of State Key Laboratory of Remote Sensing Science [Grant No. OFSLRSS202218], Chongqing Technology Innovation and Application Development Special Project [Grant No. cstc2020jscx-msxmX0193]. The anonymous reviewers provided valuable comments to help improve this manuscript, whose effort is also appreciated.