Open Access
Add to BibSonomy
Abstract
Seagrass, a submerged flowering plant, is widely distributed in coastal shallow waters and plays a significant role in maintaining marine biodiversity and carbon cycles. However, the seagrass ecosystem is currently facing degradation, necessitating effective monitoring. Satellite remote sensing observations offer distinct advantages in spatial coverage and temporal frequency. In this study, we focused on a marine lagoon (Swan Lake), located in the Shandong Peninsula of China which is characterized by a large and typical seagrass population. We conducted an analysis of remote sensing reflectance of seagrass and other objectives using a comprehensive Landsat satellite dataset spanning from 2002 to 2022. Subsequently, we constructed Seagrass Index I (SSI-I) and Seagrass Index II (SSI-II), and used them to develop a stepwise model for seagrass detection from Landsat images. Validation was performed using in situ acoustic survey data and visual interpretation, revealing the good performance of our model with an overall accuracy exceeding 0.90 and a kappa coefficient around 0.80. The long-term analysis (2002-2022) of the seagrass distribution area in Swan Lake, generated from Landsat data using our model, indicated that the central area of Swan Lake sustains seagrass for the longest duration. Seagrass in Swan Lake exhibits a regular seasonal variation, including seeding in early spring, growth in spring-summer, maturation in the middle of summer, and shrinkage in autumn. Furthermore, we observed an overall decreasing trend in the seagrass area over the past 20 years, while occasional periods of seagrass restoration were also observed. These findings provide crucial information for seagrass protection, marine blue carbon studies, and related endeavors in Swan Lake. Moreover, our study offers a valuable alternative approach that can be implemented for seagrass monitoring using satellite observations in other coastal regions.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Seagrass are flowering plants that thrive in tropical, subtropical, and temperate marine waters, playing a vital role in marine ecosystems [1]. They serve multiple ecological functions, such as sediment trapping and stabilization, leading to improve water clarity [2]. Furthermore, seagrass provides essential food sources, habitat, and sanctuary for a variety of marine vertebrate and invertebrate species [3–5]. Notably, seagrass contributes to carbon absorption through photosynthesis and is recognized as a significant carbon sink in coastal ecosystems [6,7]. They provide the necessary substratum for the growth of epiphytic, product organic carbon by sequestration [8], and rank among the most productive ecosystems on earth [9]. Despite occupying a relatively small area in the global ocean, seagrass sequesters approximately 10-15% of the total marine carbon, positioning it as a critical carbon reservoir on a global scale [10,11]. However, seagrass has experienced a steady decline of 7% per year since 1990, due to anthropogenic activities and natural events [12,13]. Consequently, it is imperative to monitor seagrass, investigate its current state, distribution, and temporal dynamics, and explore potential causes of its degradation. These efforts will provide crucial information necessary for the protection and restoration of seagrass ecosystems.
In general, there are two primary methods for monitoring seagrass, field investigations and remote sensing technology [14,15]. Field investigation methods usually involve placing transects and capturing photographs [16–18]. While field investigations offer accurate seagrass discrimination, they are time-consuming and labor-intensive. Furthermore, the spatial and temporal coverage of field observations is often limited, which poses a challenge to obtaining a comprehensive understanding of seagrass distributions. In contrast, satellite remote sensing technology utilizing optical sensors provides significant advantages, such as extensive spatial coverage, frequent monitoring intervals, rich spectral information, and cost-effectiveness. Thus, remote sensing has emerged as a crucial method for monitoring and mapping seagrass [19,20], and some researchers have successfully mapped seagrass distributions using remote sensing imagery in recent years [21,22].
Optical sensors on satellites can capture radiances at various wavelengths, which are influenced by both the water column and the reflectance of bottom substrates in shallow waters [23,24]. Different seafloor substrates exhibit distinct reflectance spectra, making it feasible to identify seagrass based on spectral features. In general, three main approaches are commonly used to interpret seagrass information from remote sensing data, namely, the water column correction approach, classification techniques, and spectral index approach [14]. The water column correction approach includes empirical, semi-, and analytical algorithms to invert the reflectance spectra of the bottom substrate by correcting the effects of the water column based on the radiation transfer theory [25–27], and then identify seagrass from the reflectance spectral features. However, this approach often requires precise parameters such as the diffuse attenuation coefficient and water depth, which are typically obtained through labor-intensive field measurements. Consequently, the widespread use of the water column correction approach is limited in practice. Classification techniques encompass three methods: empirical image-based classification algorithms [28–31], object-based algorithms [20,32] and physics-based semi-analytical models [33,34]. These techniques can identify multiple seafloor coverage types and offer automation [35]. However, classification techniques often rely on subjective visual judgment for model training using large datasets, leading to subjective dependence and unclear optical basis. Additionally, these methods are computationally complex and inefficient.
In contrast, spectral index approaches utilize the distinctive spectral characteristics of reflectance to differentiate between seagrass and non-seagrass targets. Thus, these methods have a robust optical foundation and are typically straightforward to implement. Currently, spectral index approaches are widely employed for seagrass detection. For instance, Dierssen et al. (2019) [36] utilized the red edge of reflectance spectra from hyperspectral airborne imagery to map eelgrass beds in Elkhorn Slough; Wilson et al. (2020) [24] successfully mapped vegetated habitats, including seagrass and kelp, using multiple spectral indexes, such as the normalized difference vegetation index (NDVI), the reflectance at the blue band, and the red-to-green reflectance ratio, achieving an overall accuracy of 79%. Despite the widespread use of spectral index methods for seagrass detection in satellite images due to their convenience and high accuracy, certain potential issues should be acknowledged. Firstly, the spectral indexes proposed by previous studies are not universally applicable and may not directly apply to detect seagrass in specific regions. Additionally, the threshold of the spectral index plays a crucial role in distinguishing seagrass from other targets, and it may vary depending on the sensor used or the satellite imaging conditions [37,38]. Therefore, accurately determining the threshold is of fundamental importance for spectral index-based methods.
Swan Lake, located in Weihai on the Shandong Peninsula, China, is a semi-enclosed lagoon characterized by a significant seagrass ecosystem [39]. However, the current status and long-term variations of seagrass in Swan Lake remain unclear. In this study, taking Swan Lake as the study region, we first analyzed the spectral characteristics of seagrass and other water areas of Swan Lake using a comprehensive collection of long-term Landsat data. Subsequently, we developed a seagrass detection model by constructing two spectral indices, Seagrass Index I (SSI-I) and Seagrass Index II (SSI-II) with appropriate thresholds. Finally, we applied the developed model to the long-term Landsat data to investigate the spatio-temporal distribution changes of seagrass from 2002 to 2022 and discussed their potential driving factors.
2. Materials and methods
2.1 Study site
Swan Lake, located in Rongcheng, Shandong, China (37°20′N-37°22′N, 122°33′E-122°35′E) (Fig. 1), is a semi-enclosed lagoon covering an area of 4.8 km2 with a depth ranging from 0 to 2 m [39]. This lagoon harbors a typical seagrass bed in northern China, dominated by two seagrass species: Zostera marina and Zostera japonica. Zostera marina is the predominating seagrass species and mainly grows in the lower intertidal zone and subtidal shallow water areas, while Zostera japonica is predominantly found in the upper intertidal zone [40,41]. The lake’s substrate mainly consists of chalky sand with varying proportions of mud and silt, and a smaller fraction comprises mud (20-50%) and sand [39]. These environmental characteristics make Swan Lake an ideal study area for seagrass analysis.
2.2 Satellite data
Satellite images from various Landsat sensors, including Landsat-5 TM, Landsat-7 ETM+, Landsat-8 OLI, and Landsat-9 OLI2, were obtained from the United States Geological Survey (USGS, https://www.usgs.gov/) between 2002 and 2022. These sensors are equipped with similar visible and near-infrared (NIR) bands (Table 1). The acquired image data were level 1 Tier 1 products, specifically top-of-atmosphere (TOA) reflectance, which have undergone geometric and radiometric corrections and are projected in UTM with the WGS84 datum. To ensure accurate mapping of seagrass distribution, satellite images were carefully chosen based on criteria such as cloud-free conditions, absence of skylight, and low tide, using visual discrimination. Additionally, Landsat-7 ETM+ images with missing lines were excluded from the dataset. Following image selection, a total of 275 Landsat images were obtained, comprising 103 from Landsat-5 TM, 24 from Landsat-7 ETM+, 130 from Landsat-8 OLI, and 18 from Landsat-9 OLI2 sensors.
Table 1. The spectral band characteristics of various Landsat sensors.
Atmospheric correction was conducted on all Landsat images to obtain the remote sensing reflectance (Rrs(λ)) for pixels. This correction was performed using the dark spectrum fitting method implemented in ACOLITE software (Python version v20220222.0), which is specifically designed for small inland water bodies [42,43]. Additionally, the land areas and bare pixels during low tides were masked out using the Normalized Difference Moisture Index (NDMI). The NDMI is calculated as a ratio between NIR and short-wave infrared (SWIR) spectral bands and is commonly employed for shoreline extraction [44].
2.3 Accuracy assessment
We evaluated the accuracy of the seagrass remote sensing model using the confusion matrix method [45,46]. In brief, the confusion matrix provides various metrics, including the overall accuracy (OA), kappa coefficient (K), precision, recall, and F1-score. OA represents the proportion of correctly identified pixels (including all pixels). The kappa coefficient indicates the level of agreement between identified and true pixels, with values ranging from 0 to 1. Precision denotes the ratio of correctly identified pixels in a predicted class, while recall represents the ratio of correctly identified pixels in an actual class. The F1-score is the harmonic mean of precision and recall. In our study, we classified the Swan Lake substrate into three categories: seagrass, sand, and other water areas (excluding seagrass and sand). These were further grouped into two classes: seagrass and non-seagrass (comprising sand and other water areas). To assess the accuracy, we obtained true seagrass information from field surveys mentioned in the literature and through visual interpretation.
3. Development of the seagrass detection model
3.1 Rrs spectral characteristics of seagrass and non-seagrass targets
We conducted a comprehensive investigation of the Rrs(λ) spectral characteristics of three representative targets, namely seagrass, sand, and other water areas (Fig. 2). Through visual interpretation, we collected over four thousand samples from different seasons spanning a period of 20 years. The spectral characteristics of each target were analyzed across the blue (483 nm) to SWIR (1609 nm) bands. Seagrass exhibited a prominent peak in the green band (561 nm) and a minor absorption valley at 655 nm. Furthermore, a gradual decrease in Rrs(λ) was observed from the red band (665 nm) to the NIR band (865 nm) and SWIR band (1609 nm), with a small protrusion observed at 865 nm (Fig. 2(A)). The Rrs(λ) spectra of other water areas demonstrated a peak in the green band, followed by a sharp decrease from the red band to the NIR band, and a gradual decrease from the NIR band to the SWIR band (Fig. 2(B)). For the pixels with sand substrate, the Rrs(λ) exhibited relatively high overall spectral values and a flat trend from the red band to the NIR band (Fig. 2(C)).
Fig. 2. Rrs spectra of seagrass (A), other water areas (B), and sand (C). Comparisons of the mean Rrs spectra of seagrass with that of sand and other water areas (D).
We further calculated the mean Rrs(λ) for the three target types mentioned above (Fig. 2(D)). The disparities in the mean Rrs(λ) spectra between seagrass and other water areas were primarily observed in the near-infrared band (865 nm), with seagrass displaying a small protrusion that is not present in other water areas. When comparing the mean Rrs(λ) spectra of seagrass and sand, the main discrepancies were observed in the spectral variations from green to red band. Specifically, seagrass spectra exhibited a sharp decline, while sand spectra exhibited a gradual decline (Fig. 2(D)). These distinctive spectral characteristics serve as the optical foundation for seagrass detection from Landsat Rrs(λ) images.
3.2 Seagrass remote sensing model
Based on the analysis of Rrs(λ) spectral characteristics of seagrass, sand, and other water area pixels, we propose a remote sensing model for seagrass detection that incorporates two spectral indices. We constructed the first spectral index, termed “Seagrass Index I (SSI-I)”, which is calculated as the difference between Rrs(λ) in the NIR band and a baseline established between the Rrs(λ) in the red and SWIR bands. The formula for SSI-I is as follows:
Firstly, we examined the distribution of SSI-I values for seagrass and other water areas, comprising 7,857 and 8,697 samples, respectively (Fig. 3(A)). The SSI-I values for seagrass pixels ranged from -0.0049 to 0.0132, with a mean value of 0.0004 and a standard deviation (Std) of 0.0018. Conversely, the SSI-I values for other water areas varied from -0.0202 to -0.0020, with a mean of -0.0084 and Std of 0.0024. Differentiation between the SSI-I value histograms of seagrass and other water areas was evident. Additionally, we analyzed the SSI-I values of 4,162 sand samples and found that the SSI-I distribution of sand (-0.0077 to 0.0195) overlapped more with that of seagrass, making it challenging to distinguish between them (Fig. 3(A)). This indicates the need to develop another spectral index to effectively differentiate sand from seagrass in the subsequent steps.
Fig. 3. Frequency distributions of SSI-I values for seagrass, sand, and other water area samples (A) of all Landsat sensor data. Frequency distributions of SSI-I values for seagrass and other water area samples of Landsat-5 TM and Landsat-7 ETM+ data (B), and of Landsat-8 OLI and Landsat-9 OLI2 data (C).
Furthermore, we observed that although SSI-I can generally differentiate seagrass from other water area samples, there is a partial overlap of their SSI-I values ranging from -0.0049 to -0.002, referred to as “fuzzy” targets. Approximately 16.06% of seagrass samples fell within this overlap, which undoubtedly impacts the accuracy of distinguishing between seagrass and other water area samples. We hypothesized that this result might be attributed to differences in the band settings of the Landsat sensors. Notably, Table 1 reveals no variation between Landsat-5 TM and Landsat-7 ETM+ in the NIR and SWIR bands (i.e., characteristic bands), while Landsat-8 OLI and Landsat-9 OLI2 display greater differences. Therefore, we attempted to separate the SSI-I values into two groups based on the sensor band characteristics: one group comprising TM and ETM+ sensors and the other group consisting of OLI and OLI2 sensors (Figs. 3(B) and 3(C)).
For the TM and ETM+ group, SSI-I values for seagrass pixels ranged from -0.0049 to 0.0114 (Mean = -0.0002, Std = 0.0016), while other water area samples ranged from -0.0202 to -0.0025 (Mean = -0.0080, Std = 0.0022). Although the range of “fuzzy” targets in the TM and ETM+ group showed negligible changes, the proportion of seagrass samples decreased from 16.06% to 11.38% (Fig. 3(B)). The SSI-I values for seagrass pixels in the OLI and OLI2 group distributed within the range of -0.0021 to 0.0132 with an average of 0.0008 (Std = 0.0018), while other water area pixels changed from -0.0184 to -0.0020 (Mean = -0.0086, Std = 0.0025). Notably, the overlapping SSI-I values between the two categories disappeared, eliminating any cross-mixing of pixels (Fig. 3(C)). This finding suggests that grouping Landsat data according to the sensors improves SSI-I's ability to distinguish seagrass from other water area samples. Based on Figs. 3(B) and 3(C), we determined that the threshold of SSI-I for distinguishing seagrass using Landsat-5 TM and Landsat-7 ETM+ is -0.0034 (the midpoint of the overlapping SSI-I area), while the threshold for Landsat-8 OLI and Landsat-9 OLI2 is -0.0021 (the minimum SSI-I value for seagrass).
In addition, it is important to emphasize that although SSI-I, with an appropriate threshold for each group of Landsat data, enables the discrimination between seagrass and other water area pixels and also eliminates a significant portion of sandy substrates, a small fraction of sand samples were still incorrectly classified as seagrass, as depicted in Fig. 3(A). Consequently, we endeavored to develop a second Rrs(λ) spectral index to further distinguish between seagrass and sand.
Through the analysis of the Rrs(λ) spectral variations between seagrass and sand samples, we observed a distinct disparity in the green band (561 nm) and red band (655 nm), where seagrass exhibited a sharp decline while sand displayed a gradual decrease from the green to red band (Fig. 2(D)). Building upon this observation, we developed the second spectral index, designated “Seagrass Index II (SSI-II)”, with the purpose of discriminating between seagrass and sand. The calculation of SSI-II is defined as follows:
Similar to the analysis of the SSI-I, we calculated the SSI-II values for seagrass and sand samples across all Landsat sensor data, and their frequency distributions are depicted in Fig. 4(A). Seagrass samples exhibited SSI-II values ranging from 0.1325 to 1.8944, with a mean of 0.4443 (Std = 0.1622), while sand samples spanned from -0.2138 to 0.2838 (Mean = 0.0258, Std = 0.0716). Within the “fuzzy” interval (0.1325 to 0.2838), approximately 17.43% of seagrass samples and 7.96% of sand samples were situated, indicating the challenge of accurately distinguishing between seagrass and sand. Consequently, we separately calculated the SSI-II values for the two groups of Landsat sensors: Landsat-5 TM and Landsat-7 ETM+, Landsat-8 OLI and Landsat-9 OLI2.
Fig. 4. Frequency distributions of SSI-II values for seagrass and sand samples of all Landsat sensors data (A), of Landsat-5 TM and Landsat-7 ETM+ data (B), and of Landsat-8 OLI and Landsat-9 OLI2 data (C).
For the TM and ETM+ sensors (Fig. 4(B)), the SSI-II values of seagrass ranged from 0.1325 to 0.1894, with a mean of 0.4851 (Std = 0.1991). In contrast, sand samples exhibited SSI-II values ranging from -0.1807 to -0.2838, with a mean of 0.0381 (Std = 0.0705). Likewise, within the “fuzzy” pixel range, the proportions of seagrass and sand samples decreased to 0.14% and 3.73%, respectively. For the OLI and OLI2 group (Fig. 4(C)), the SSI-II values of seagrass ranged from 0.1467 to 1.2826, with a mean of 0.4157 (Std = 0.1224), while sandy samples exhibited values ranging from -0.2138 to 0.2223, with a mean of 0.0088 (Std = 0.0697). Within the “fuzzy” pixel range (0.1467 to 0.2223), seagrass pixels accounted for 0.74% and sandy pixels for 0.17%, indicating a low-level mixture achieved by segregating Landsat data based on sensors. Subsequently, we determined the SSI-II thresholds for distinguishing seagrass and sand as 0.1627 for Landsat-5 TM and Landsat-7 ETM+, and 0.2071 for Landsat-8 OLI and Landsat-9 OLI2, by calculating the middle of the overlapping SSI-II area. So far, combining the two indexes of SSI-I and SSI-II, a new spectral-based model has been proposed for seagrass detection in Swan Lake using Landsat images. The flowchart outlining the development of this model is provided in Fig. 5.
Fig. 5. Flowchart of the remote sensing model for seagrass detection. SSI-I and SSI-II denote the seagrass spectral index I and II, respectively. A and B are the thresholds of SSI-I and SSI-II, respectively.
4. Model evaluation and application
4.1 Performance of the seagrass detection model
To evaluate the accuracy of our developed seagrass detection model, we compared the satellite detection results with in situ surveys and visual interpretation. The in situ data were obtained from acoustic surveys conducted by Xu (2019) [39] in Swan Lake during September 2017 and August 2018, reporting seagrass areas of 231.7 ha and 199.1 ha, respectively. Therefore, we selected satellite images from these two time periods and applied our remote sensing model to detect seagrass. Firstly, we employed visual interpretation to validate the model's performance. As shown in Fig. 6, the distributions of seagrass detected from the satellite images exhibited overall consistency with the pseudo-color images (NIR, red, and green band combinations). The values of OA and F1-score exceeded 0.90, and the kappa coefficients surpassed 0.82 (Table 2). Meanwhile, our model’s results revealed seagrass distribution areas of 214.1 ha and 169.7 ha in September 2017 and August 2018, respectively, demonstrating good agreement with the field measurements by Xu (2009) [39]. In addition, we compared our model's performance with existing models. However, to the best of our knowledge, there is no spectral-based seagrass detection model specifically designed for Swan Lake. Thus, we implemented the approach of Wilson et al. (2020) [24] for the coastal waters of Atlantic Canada, which uses NDVI, reflectance at the blue band, and a red-to-green reflectance ratio for seagrass detection. The comparisons demonstrated that our model exhibited higher accuracy (Table 2). These findings indicate that our proposed seagrass detection model exhibits good performance.
Fig. 6. Comparisons of the pseudo-color image (NIR, red, and green band combinations) of Landsat-8 OLI with seagrass distribution detected using the remote sensing model of this study in September 2017 (A and B) and August 2018 (C and D), respectively.
Table 2. Comparison of the performance between our seagrass detection model and the present model in September 2017 and August 2018.
Note that the aforementioned model validation conducted in September and August, during the peak growth phase of seagrass, facilitated relatively easy remote sensing detection due to the extensive seagrass distribution. However, this study aims to detect and analyze the spatial and temporal changes of seagrass in Swan Lake, including its growth stages during seeding and shrinkage when seagrass is less abundant. Thus, it is crucial to verify whether the model can achieve satisfactory performance under these varying seagrass conditions. To accomplish this, we selected satellite images from four different seasons for each group of Landsat sensors: Landsat-5 TM and Landsat-7 ETM+, Landsat-8 OLI and Landsat-9 OLI2. These images were utilized to detect seagrass distributions, and the accuracy indicators were computed based on visual interpretation data to evaluate the model's performance (Table 3). The results indicated that, except for one period, the OA values for the other periods exceeded 0.90. Accordingly, the majority of kappa coefficients were approximately 0.80, and the F1-score values were above 0.85. These findings demonstrate that the developed model exhibits robust performance across different seagrass growth stages, even when seagrass distribution is limited. This further enables us to investigate the spatial and temporal variations of seagrass.
Table 3. Performance of the remote sensing detection model at different growth stages of seagrass.
4.2 Variations of seagrass distribution from 2002 to 2022
We utilized the newly developed model to analyze the spatial-temporal variations of seagrass distribution in Swan Lake over the past two decades (2002-2022) using Landsat data. A total of 255 cloud-free, no skylight, and low tide Landsat images were employed for seagrass detection. The monthly average seagrass area in Swan Lake was computed by aggregating data over twenty years (Fig. 7). Significant seasonal variations in the seagrass area were observed. From January to August, there was a consistent increase in the seagrass area, followed by a decline from August to December. More specifically, the seagrass area exhibited a gradual increase from March to April, and a rapid expansion from April to July, reaching its peak in August. From September to November, the seagrass area gradually decreased, and during winter (December-February), it maintained a relatively low level. These observations indicate that the growth of seagrass in Swan Lake can be characterized by four stages: seeding during late winter and early spring (January-April), growth in late spring and early summer (May-July), maturation in mid-summer (around August), and shrinkage in autumn (September-November). This growth pattern agrees with those observed in other regions, as reported in previous studies [40,48,49].
Fig. 7. Seasonal variation in the 20-year (2002-2022) averaged monthly seagrass area in Swan Lake from January to December was derived from Landsat images.
We further conducted an analysis of the monthly spatial frequency distribution of seagrass in Swan Lake by examining the number of seagrass pixels present in each location across all Landsat images spanning the past two decades. This allowed us to investigate the spatial variations in seagrass distribution from January to December and gain insights into the primary areas of seagrass growth in Swan Lake. The results, as depicted in Fig. 8, reveal distinct patterns in seagrass distribution from January to December. In February and March, seagrass was sparsely distributed throughout Swan Lake. However, in April, noticeable sprouting of seagrass was observed, particularly in the central and northern regions. From May to August, the coverage of seagrass expanded significantly, primarily in the central area of Swan Lake, with the seagrass growth exhibiting an outward pattern. By August, the seagrass coverage had reached its maximum extent, with the central and northern areas exhibiting the highest occurrence of seagrass throughout the entire lake.
Fig. 8. Monthly frequency distribution of seagrass in Swan Lake from January to December over the past 20 years (2002-2022).
Subsequently, the seagrass distribution range gradually decreased, and from October to November, a significant reduction in seagrass occurrence was observed in the central and northern regions, although seagrass coverage persisted. Concurrently, a recession was noticeable along the entire edge of Swan Lake, particularly the southern edge. From November to January, seagrass gradually faded, eventually becoming sporadically present in January. Interestingly, it was observed that the central area emerged as both the primary distribution area and the location where seagrass first appeared and last disappeared (Fig. 8).
To investigate the annual changes in seagrass in Swan Lake, we computed the average distribution area of seagrass using Landsat images from July to September, which corresponds to the period of extensive seagrass coverage during the growth season. We excluded data from 2007 and 2012 due to poor image quality caused by missing lines in Landsat-7 ETM+ images. Additionally, the data from July to September 2019 was also removed due to cloud cover. The analysis revealed an overall declining trend in the seagrass distribution area between 2002 and 2022, although fluctuations were present (Fig. 9). Between 2002 and 2004, a significant increasing trend was observed, reaching a peak value of 357.6 ha in 2004. Subsequently, the seagrass area remained relatively stable from 2005 to 2011, with mean values hovering around 288 ha. During this period, slight fluctuations were also observed, with a minimum area of 233.7 ha in 2009 and a maximum area of 338.9 ha in 2011. Starting in 2011, an obvious decreasing trend in the seagrass area was observed until 2018, especially in 2013 and 2015. However, there was a slight increase in seagrass area from 2019 to 2022. It is worth noting that although the distribution area of seagrass exhibited fluctuations over the 20-year period, the peak values within each rising period (2002-2004, 2009-2011, and 2019-2022) gradually decreased, measuring 357.6 ha, 338.9 ha, and 242.1 ha, respectively. This observation further supports the overall declining trend in seagrass distribution.
Fig. 9. Annual (July–September) mean distribution area of seagrass in Swan Lake between 2002 and 2022 derived from Landsat images. The dotted line is the linear regression curve representing the variation trend of the seagrass area during the past 20 years.
5. Discussion
In this study, we have developed a stepwise model for the detection of seagrass from Landsat satellite images. The model incorporates two spectral indexes of remote sensing reflectance Rrs(λ), i.e., SSI-I and SSI-II, derived from the analysis of spectral variations between seagrass and non-seagrass targets observed across different Landsat sensors over an extended period. This ensures that our model is built on a robust optical foundation, and validation based on field observation and visual interpretation has demonstrated the good performance of the model, with OA values exceeding 0.90. However, it is important to emphasize that accurately capturing the spatial and temporal variations of seagrass using long-term data relies on the appropriate determination of threshold values for the spectral indexes. In this study, considering the unique characteristics of different Landsat sensors, we divided the Landsat data into two groups: Landsat-5 TM and Landsat-7 ETM+ group, and Landsat-8 OLI and Landsat-9 OLI2 group. This division enabled us to identify optimal thresholds for SSI-I and SSI-II. As depicted in Figs. 3 and 4, the results confirmed that this grouping approach enabled more accurate threshold determination, thereby improving the accuracy of seagrass detection.
Meanwhile, the stability of the thresholds for SSI-I and SSI-II is another crucial aspect influencing the accuracy of seagrass detection using our model with long-term Landsat data. Therefore, we conducted tests to evaluate the validity of the spectral index thresholds determined in this study. We selected satellite images from different seasons and years for each Landsat sensor group and analyzed the frequency distributions of SSI-I and SSI-II for each image. As shown in Fig. 10, the frequency distributions of SSI-I values for seagrass and other water areas exhibited clear distinctions across all images and time periods. More importantly, the threshold determined in this study for seagrass detection is generally located around the valleys between the SSI-I peaks of seagrass and other water samples. This implies that our thresholds have minimal impact on seagrass detection during different seasons and years. Regarding SSI-II, a slight overlap between seagrass and sand features; even so, the thresholds used in this study still yielded satisfactory results for most periods, with only minor misclassifications occurring in winter when seagrass presence is sporadic. The misclassification of minor seagrass pixels as other water pixels may be attributed to the weaker activity and internal structure of seagrass during winter, resulting in low reflectance in the NIR band and a correspondingly lower SSI-I value than the determined threshold. A small amount of sand misclassified as seagrass may be related to high water turbidity, which weakens the contributions of the sand substrate to the Rrs and results in relatively low values in the red band, leading to an SSI-II value higher than the determined threshold. Nevertheless, the misclassification rate of seagrass on December 23, 2016, was less than 7%. These findings demonstrate the overall robustness and applicability of the SSI-I and SSI-II thresholds established in this study for long-term seagrass detection. This may be attributed to the fact that the thresholds were determined based on a comprehensive analysis of a large collection of long-term Landsat images.
Fig. 10. Frequency distribution of SSI-I (A) and SSI-II (B) in different seasons and years. The red line denotes the threshold of SSI-I or SSI-II used for seagrass detection. The data in the first row of figures (A) and (B) were from Landsat-5 TM and Landsat-7 ETM+ observations in spring, summer, autumn, and winter, and the data in the second row of figures (A) and (B) were from Landsat-8 OLI and Landsat-9 OLI2 observations in four seasons.
Utilizing long-term Landsat images, we could obtain and examine the spatial and temporal variations in seagrass distribution in Swan Lake. Our findings revealed distinct seasonality and cyclicity in seagrass growth, wherein they exhibited gradual expansion from spring to summer, reaching maturation around August, followed by degeneration during autumn and winter. Previous studies have emphasized the significance of water temperature as a crucial factor influencing seagrass growth [23,50]. Generally, seagrass thrives within a temperature range of 0-30°C, with optimal growth occurring between 15-20°C. When the water temperature exceeds 30°C, seagrass growth is impeded, and even the root system may suffer damage or perish. To explore the relationship between seagrass variation and water temperature, we obtained monthly climatological sea surface temperature (SST) data from ERA5, the fifth-generation ECMWF reanalysis for global climate and weather, spanning the period from 2002 to 2022. Our analysis revealed an average SST of approximately 7.73°C during spring, reaching a maximum of 24.48°C in summer, followed by a decline during winter to 6.16°C (Fig. 11(A)). This variation in SST closely corresponds to the seasonal changes in seagrass areas, with growth occurring as temperatures warm from spring to summer, and regression leading to disappearance from autumn to winter. These results suggest a plausible link between the observed seasonality of seagrass in Swan Lake and water temperature fluctuations. This was confirmed by the correlation analysis, showing a strong relationship (R2 = 0.96, p < 0.0001) between SST and seagrass area (Fig. 11(B)). However, it should be noted that in addition to temperature, other water environmental factors, such as pH, salinity, and dissolved oxygen, maybe also possible causes related to seagrass growth [48,51]. Further investigation is warranted to comprehensively explore these factors and their potential contributions to seagrass dynamics.
Fig. 11. Variation of climatological monthly average SST in Swan Lake from January to December over the past 20 years (2002-2022) (A) and the relationship with monthly average seagrass area (B).
This study also examines the annual variations of seagrass in Swan Lake and reveals an overall declining long-term trend with fluctuations from 2002 to 2022. Several factors, including anthropogenic activities and natural disturbances, may contribute to this phenomenon. The low seagrass area in Swan Lake in 2002 may potentially be attributed to sea cucumber farming and the presence of dams. From 2002 to 2004, the seagrass area in Swan Lake exhibited a clear increasing trend, which then stabilized (Fig. 9). This variation pattern may be linked to human activities, such as the removal of the dam and fish ponds near the inlet, as well as in the western and northern coastal areas of Swan Lake [52]. However, there were significant decreases in the seagrass area in 2013 and 2015, which may be associated with the discharge of the Jinshui River. Starting in 2011, freshwater from the Jinshui River was directly released into Swan Lake without effective treatment, probably leading to intensified sandbank erosion processes in the lake, which in turn may have a negative impact on the growth of seagrass [53].
Additionally, from 2009 to 2011, a small increase in the seagrass distribution area was observed, likely attributed to artificial seagrass cultivation efforts undertaken by the local government. Especially, it should be mentioned that a transplantation experiment of eelgrass Zostera marina was conducted in Swan Lake in 2017 [54]. This may lead to a large-scale restoration of seagrass and contribute to a significant increase in seagrass areas in recent years (from 2019 as shown in Fig. 9). The aforementioned restoration activities indicate an increased awareness of the importance of seagrass among humans, and the government's commitment to seagrass conservation. However, it is important to acknowledge that there has been an overall decrease in the seagrass distribution area in Swan Lake between 2002 and 2022. Therefore, greater attention and constructive efforts are required to protect and restore seagrass in this region. In addition to human activities and natural disasters, some researchers propose that global warming may also be a contributing factor to the changes in seagrass areas [13,55]. At this stage, it is important to recognize that this study represents only a preliminary investigation into the potential influencing factors of seagrass. Future studies should focus on comprehensively examining the driving mechanisms by utilizing available datasets encompassing both human activities and the marine environment, which will contribute to a deeper understanding of seagrass dynamics.
6. Conclusions
This study developed a stepwise model for seagrass detection from satellite observation utilizing two spectral indexes, namely SSI-I and SSI-II, that were constructed based on the analysis of spectral characteristics of remote sensing reflectance of large amounts of Landsat images. Evaluations conducted using the in situ data and visual interpretation demonstrated good performances of the detection model, with overall accuracy greater than 0.90 and a kappa coefficient around 0.80. Moreover, the thresholds determined for SSI-I and SSI-II are proved to be generally robust. When applying the model to long-term Landsat data spanning from 2002 to 2022, notable spatial and temporal variations in the distribution area of seagrass were observed in Swan Lake. The central Swan Lake exhibited the earliest growth of seagrass and sustained its presence for the longest duration. The temporal variation of seagrass showed a distinct seasonality pattern, with seeding occurring in early spring, growth during late spring and early summer, maturation in the middle of summer, and shrinkage in autumn. The analysis of the long-term changes between 2002-2022 indicated an overall decreasing trend in the seagrass distribution area; however, fluctuations, especially small increases, were also observed during certain short periods, which could potentially be attributed to human interventions such as active protection and restoration efforts.
Overall, the spectral-based approach for seagrass detection proposed in this study can serve as a proof-of-concept template for developing site-specific models in other coastal regions. By utilizing long-term satellite observations, this study provides novel insights into the spatial and temporal dynamics of seagrass in Swan Lake. Moreover, it presents a preliminary analysis of the potential driving factors behind these dynamics. The findings of this study offer valuable insights for seagrass management and protection, marine blue carbon studies, and related fields. Future research should focus on investigating the mechanisms driving seagrass changes and implementing the new proposed approach to different satellite sensors and other regions for seagrass detection.
Funding
National Natural Science Foundation of China (42176181, 42001240, 42176179, 42106176); Natural Science Foundation of Jiangsu Province (BK20211289, BK20210667); Open Fund of State Key Laboratory of Remote Sensing Science (OFSLRSS202103); Open Fund of Key Laboratory of Coastal Zone Exploitation and Protection, Ministry of Natural Resources (2021CZEPK02); Zhejiang Provincial Natural Science Foundation of China (LY20C030004).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. J. L. Olsen, P. Rouzé, B. Verhelst, et al., “The genome of the seagrass Zostera marina reveals angiosperm adaptation to the sea,” Nature 530(7590), 331–335 (2016). [CrossRef]
2. T. van der Heide, E. H. van Nes, M. M. van Katwijk, H. Olff, and A. J. P. Smolders, “Positive Feedbacks in Seagrass Ecosystems – Evidence from Large-Scale Empirical Data,” PLoS One 6(1), e16504 (2011). [CrossRef]
3. R. Costanza, R. D’Arge, R. de Groot, S. Farber, M. Grasso, B. Hannon, K. Limburg, S. Naeem, R.V. O’Neill, J. Paruelo, R.G. Raskin, P. Sutton, and M. van den Belt, “The value of the world's ecosystem services and natural capital,” Ecol. Econ. 25(1), 3–15 (1998). [CrossRef]
4. J. E. Duffy, “Biodiversity and the functioning of seagrass ecosystems,” Mar. Ecol.-Prog. Ser. 311, 233–250 (2006). [CrossRef]
5. R. Unsworth, L. Nordlund, and L. Cullen-Unsworth, “Seagrass meadows support global fisheries production,” Conserv. Lett. 12(1), e12566 (2019). [CrossRef]
6. E. Mcleod, G. Chmura, S. Bouillon, R. Salm, M. Björk, C. Duarte, C. Lovelock, W. Schlesinger, and B. Silliman, “A blueprint for blue carbon: Toward an improved understanding of the role of vegetated coastal habitats in sequestering CO2,” Front. Ecol. Environ. 9(10), 552–560 (2011). [CrossRef]
7. V. X. H. Phang, L. M. Chou, and D. A. Friess, “Ecosystem carbon stocks across a tropical intertidal habitat mosaic of mangrove forest, seagrass meadow, mudflat and sandbar,” Earth Surf. Process. Landf. 40(10), 1387–1400 (2015). [CrossRef]
8. P. I. Macreadie, M. E. Baird, S. M. Trevathan-Tackett, A. W. D. Larkum, and P. J. Ralph, “Quantifying and modelling the carbon sequestration capacity of seagrass meadows – A critical assessment,” Mar. Pollut. Bull. 83(2), 430–439 (2014). [CrossRef]
9. C. M. Duarte and J. Cebrián, “The fate of marine autotrophic production,” Limnol. Oceanogr. 41(8), 1758–1766 (1996). [CrossRef]
10. J. W. Fourqurean, C. M. Duarte, H. Kennedy, N. Marbà, M. Holmer, M. A. Mateo, E. T. Apostolaki, G. A. Kendrick, D. Krause-Jensen, K. J. McGlathery, and O. Serrano, “Seagrass ecosystems as a globally significant carbon stock,” Nat. Geosci. 5(7), 505–509 (2012). [CrossRef]
11. A. Knudby, C. Newman, Y. Shaghude, and C. Muhando, “Simple and effective monitoring of historic changes in nearshore environments using the free archive of Landsat imagery,” Int. J. Appl. Earth Obs. Geoinf. 12, S116–S122 (2010). [CrossRef]
12. M. Waycott, C. M. Duarte, T. J. B. Carruthers, R. J. Orth, W. C. Dennison, S. Olyarnik, A. Calladine, J. W. Fourqurean, K. L. Heck, A. R. Hughes, G. A. Kendrick, W. J. Kenworthy, F. T. Short, and S. L. Williams, “Accelerating loss of seagrasses across the globe threatens coastal ecosystems,” Proc. Natl. Acad. Sci. U.S.A. 106(30), 12377–12381 (2009). [CrossRef]
13. S. Xu, Y. Zhang, Y. Zhou, S. Xu, S. Yue, M. Liu, and X. Zhang, “Warming northward shifting southern limits of the iconic temperate seagrass (Zostera marina),” Iscience. 25(8), 104755 (2022). [CrossRef]
14. M. S. Hossain, J. S. Bujang, M. H. Zakaria, and M. Hashim, “The application of remote sensing to seagrass ecosystems: an overview and future research prospects,” Int. J. Remote Sens. 36(1), 61–114 (2015). [CrossRef]
15. Y. Li, J. Bai, L. Zhang, and T. Li, “Review on the research of seagrass optical remote sensing monitoring,” J. Tropical Oceanograph. 40(06), 1–13 (2021). [CrossRef]
16. S. Misbari and M. Hashim, “Change Detection of Submerged Seagrass Biomass in Shallow Coastal Water,” Remote Sens. 8(3), 200 (2016). [CrossRef]
17. P. Wicaksono, M. A. Fauzan, I. S. W. Kumara, R. N. Yogyantoro, W. Lazuardi, and Z. Zhafarina, “Analysis of reflectance spectra of tropical seagrass species and their value for mapping using multispectral satellite images,” Int. J. Remote Sens. 40(23), 1–24 (2019). [CrossRef]
18. P. Wicaksono and W. Lazuardi, “Assessment of PlanetScope images for benthic habitat and seagrass species mapping in a complex optically shallow water environment,” Int. J. Remote Sens. 39(17), 5739–5765 (2018). [CrossRef]
19. F. Immordino, M. Barsanti, E. Candigliota, S. Cocito, I. Delbono, and A. Peirano, “Application of Sentinel-2 Multispectral Data for Habitat Mapping of Pacific Islands: Palau Republic (Micronesia, Pacific Ocean),” J. Mar. Sci. Eng. 7(9), 316–321 (2019). [CrossRef]
20. J. Marcello, F. Eugenio, J. Martín, and F. Marqués, “Seabed Mapping in Coastal Shallow Waters Using High Resolution Multispectral and Hyperspectral Imagery,” Remote Sens. 10(8), 1208–1228 (2018). [CrossRef]
21. N. Ha, M. Manley-Harris, T. Pham, and I. Hawes, “Detecting Multi-Decadal Changes in Seagrass Cover in Tauranga Harbour, New Zealand, Using Landsat Imagery and Boosting Ensemble Classification Techniques,” Isprs Int. J. Geo-Inf. 10(6), 371–391 (2021). [CrossRef]
22. K. L. Wilson, M. A. Skinner, and H. K. Lotze, “Eelgrass (Zostera marina) and benthic habitat mapping in Atlantic Canada using high-resolution SPOT 6/7 satellite imagery,” Estuarine Coastal and Shelf Science. 226, 106292 (2019). [CrossRef]
23. B. K. Veettil, R. D. Ward, M. D. A. C. Lima, M. Stankovic, P. N. Hoai, and N. X. Quang, “Opportunities for seagrass research derived from remote sensing: A review of current methods,” Ecol. Indic. 117, 106560 (2020). [CrossRef]
24. K. L. Wilson, M. C. Wong, and E. Devred, “Branching Algorithm to Identify Bottom Habitat in the Optically Complex Coastal Waters of Atlantic Canada Using Sentinel-2 Satellite Imagery,” Front. Environ. Sci. 8, 579856 (2020). [CrossRef]
25. D. R. Lyzenga, “Passive remote sensing techniques for mapping water depth and bottom features,” Appl. Opt. 17(3), 379–383 (1978). [CrossRef]
26. D. R. Lyzenga, “Remote sensing of bottom reflectance and water attenuation parameters in shallow water using aircraft and Landsat data,” Int. J. Remote Sens. 2(1), 71–82 (1981). [CrossRef]
27. T. Sagawa, E. Boisnier, T. Komatsu, K. B. Mustapha, A. Hattour, N. Kosaka, and S. Miyazaki, “Using bottom surface reflectance to map coastal marine areas: a new application method for Lyzenga's model,” Int. J. Remote Sens. 31(12), 3051–3064 (2010). [CrossRef]
28. D. Poursanidis, K. Topouzelis, and N. Chrysoulakis, “Mapping coastal marine habitats and delineating the deep limits of the Neptune's seagrass meadows using very high resolution Earth observation data,” Int. J. Remote Sens. 39(23), 8670–8687 (2018). [CrossRef]
29. D. Traganos, B. Aggarwal, D. Poursanidis, K. Topouzelis, N. Chrysoulakis, and P. Reinartz, “Towards Global-Scale Seagrass Mapping and Monitoring Using Sentinel-2 on Google Earth Engine: The Case Study of the Aegean and Ionian Seas,” Remote Sens. 10(8), 1227–1240 (2018). [CrossRef]
30. D. Traganos and P. Reinartz, “Machine learning-based retrieval of benthic reflectance and Posidonia oceanica seagrass extent using a semi-analytical inversion of Sentinel-2 satellite data,” Int. J. Remote Sens. 39(24), 9428–9452 (2018). [CrossRef]
31. D. Traganos and P. Reinartz, “Mapping Mediterranean seagrasses with Sentinel-2 imagery,” Mar. Pollut. Bull. 134, 197–209 (2018). [CrossRef]
32. P. Wicaksono, P. A. Aryaguna, and W. Lazuardi, “Benthic Habitat Mapping Model and Cross Validation Using Machine-Learning Classification Algorithms,” Remote Sens. 11(11), 1279–1302 (2019). [CrossRef]
33. T. Kutser, J. Hedley, C. Giardino, C. Roelfsema, and V. E. Brando, “Remote sensing of shallow waters – A 50 year retrospective and future directions,” Remote Sens. Environ. 240, 111619 (2020). [CrossRef]
34. L. Mckinna, P. Fearns, S. Weeks, J. Werdell, M. Reichstetter, B. Franz, D. Shea, and G. Feldman, “A Semianalytical Ocean Color Inversion Algorithm with Explicit Water Column Depth and Substrate Reflectance Parameterization,” J. Geophys. Res. Oceans 120(3), 1741–1770 (2015). [CrossRef]
35. T. Yamakita, F. Sodeyama, N. Whanpetch, K. Watanabe, and M. Nakaoka, “Application of deep learning techniques for determining the spatial extent and classification of seagrass beds, Trang, Thailand,” Bot. Marina. 62(4), 291–307 (2019). [CrossRef]
36. H. M. Dierssen, K. J. Bostrom, A. Chlus, K. Hammerstrom, D. R. Thompson, and Z. Lee, “Pushing the Limits of Seagrass Remote Sensing in the Turbid Waters of Elkhorn Slough, California,” Remote Sens. 11(14), 1664–1687 (2019). [CrossRef]
37. A. Mora-Soto, M. Palacios, E. Macaya, I. Gómez, P. Huovinen, A. Pérez-Matus, M. Young, N. Golding, M. Toro, M. Yaqub, and M. Macias-Fauria, “A High-Resolution Global Map of Giant Kelp (Macrocystis pyrifera) Forests and Intertidal Green Algae (Ulvophyceae) with Sentinel-2 Imagery,” Remote Sens. 12(4), 694–713 (2020). [CrossRef]
38. C. Hu, “A novel ocean color index to detect floating algae in the global oceans,” Remote Sens. Environ. 113(10), 2118–2129 (2009). [CrossRef]
39. S. Xu, A Study on Sonar Detection of Typical Zostera Marina Seagrass Beds in the Yellow-Bohai Seas and Eelgrass Seed Storage (University of Chinese Academy of Sciences, 2019).
40. J. Liu, W. Li, L. Qin, and X. Zhang, “Spatio-temporal variations in benthic and macrobenthic mol-luscs in Swan Lake, Shandong, China,” Marine Sci. 41(1), 113–122 (2017). [CrossRef]
41. L. Tan, Life history strategies of eelgrass (Zostera marina L.) in Swan Lake, Shandong Peninsula, China (Ocean University of China, 2015).
42. Q. Vanhellemont, “Adaptation of the dark spectrum fitting atmospheric correction for aquatic applications of the Landsat and Sentinel-2 archives,” Remote Sens. Environ. 225, 175–192 (2019). [CrossRef]
43. Q. Vanhellemont and K. Ruddick, “Atmospheric correction of metre-scale optical satellite data for inland and coastal water applications,” Remote Sens. Environ. 216, 586–597 (2018). [CrossRef]
44. E. H. Wilson and S. A. Sader, “Detection of forest harvest type using multiple dates of Landsat TM imagery,” Remote Sens. Environ. 80(3), 385–396 (2002). [CrossRef]
45. S. Wang, J. Li, B. Zhang, E. Spyrakos, A. N. Tyler, Q. Shen, F. Zhang, T. Kuster, M. K. Lehmann, Y. Wu, and D. Peng, “Trophic state assessment of global inland waters using a MODIS-derived Forel-Ule index,” Remote Sens. Environ. 217, 444–460 (2018). [CrossRef]
46. R. G. Congalton, “A review of assessing the accuracy of classifications of remotely sensed data,” Remote Sens. Environ. 37(1), 35–46 (1991). [CrossRef]
47. L. Qi, C. Hu, Q. Xing, and S. Shang, “Long-term trend of Ulva prolifera blooms in the western Yellow Sea,” Harmful Algae. 58, 35–44 (2016). [CrossRef]
48. Y. Zhang, X. Wang, and J. Lu, “The temporal and spatial variation characteristics of key environmental factors in Zostera marina bed and bare sand areas in Swan Lake, Rongcheng,” Chin. J. Marine Environ. Sci. 41(2), 253–266 (2022).
49. Y. Sun, Z. Song, and P. Liu, “Rhizosphere microbial community structure and driving mechanism of Zostera marina and Zostera japonica in Swan Lake, Weihai,” Acta Microbiologica Sinica. 61(09), 2675–2692 (2021). [CrossRef]
50. S. M. Trevathan-Tackett, C. D. H. Sherman, M. J. Huggett, et al., “A horizon scan of priorities for coastal marine microbiome research,” Nat. Ecol. Evol. 3(11), 1509–1520 (2019). [CrossRef]
51. Y. Zhang, J. Li, Z. Wu, S. Yuan, W. Li, and P. Zhang, “Effects of Different Prolonged Light Durations on Survival, Growth and Physiology of the Eelgrass Zostera marina,” Front. Environ. Sci. 10, 893377 (2022). [CrossRef]
52. D. Yang, X. Yin, and L. Zhou, “Seagrass distribution changes in Swan Lake of Shandong Peninsula from 1979 to 2009 inferred from satellite remote sensing data,” Satell. Oceanogr. Meteorol. 3(1), 301–311 (2018). [CrossRef]
53. F. Liu, Study on ecological protection and restoration of Rongcheng Swan Lake wetland landscape (Harbin Institute of Technology, 2016).
54. Q. Zhang, J. Liu, P. Zhang, Y. Liu, and Q. Xu, “Effect of silt and clay percentage in sediment on the survival and growth of eelgrass Zostera marina: Transplantation experiment in Swan Lake on the eastern coast of Shandong Peninsula, China,” Aquat. Bot. 122, 15–19 (2015). [CrossRef]
55. H. Tan, R. Cai, J. Du, and W. Hu, “Climate change and marine ecosystems: impacts, adaptation, and vulnerability,” Trans Atmos Sci. 45(04), 489–501 (2022). [CrossRef]
