1. Introduction
Intertidal ecosystems are productive and dynamic environments located between low and high tide levels. These environments play various essential ecological and biogeochemical roles that benefit both terrestrial and marine ecosystems by sequestering carbon and nitrogen and regulating biogeochemical cycles [
1,
2]. With the increase in population, industrial development, and over-exploitation of resources over the past decades, coastal and littoral environments are among the most vulnerable ecosystems in the world [
1,
3,
4,
5,
6]. The anthropogenic pressures associated with coastal ecosystems and the increase in sea levels linked to climate change contribute to the loss of 20 to 70% of these ecosystems [
7,
8]. In addition, global warming causes a rise in the number and intensity of winter storms and reduces the coastal ice cover in cold temperate regions and higher latitudes, leading to rapid changes in coastal environment functions and structures [
8]. Management activity relies on the provision of spatial and explicit information on coastal vegetation distribution, making it more important than ever [
9].
Satellite remote sensing technology is increasingly used to map vegetation cover of habitats, varying from forests to coasts and their temporal evolution [
10,
11]. However, more in situ observations are needed to obtain accurate mapping and to orient the development of satellite-based classification algorithms [
12,
13]. Optical remote sensing images provide essential information on vegetation cover, such as the spatial extent presence and the dominant vegetation type, the biomass estimate, the percentage cover, and the leaf area index (LAI), [
14,
15,
16]. Recent advances in satellite sensors and algorithms now provide the ability to monitor coastal vegetation at relatively high spatial (<10 m pixel size) and temporal (weekly) scales [
17]. Spaceborne sensors commonly used to map coastal vegetation ecosystems or habitats include Landsat-8 [
18], Sentinel-2 [
17,
19], RapidEye [
19], and PlanetScope [
20].
Mapping coastal vegetation using satellite remote sensing imagery relies on the assumption that the dominant vegetation type presents unique spectral signatures to distinguish from others [
9,
21]. However, reflectance spectra of coastal vegetation frequently overlap, change rapidly over the season, and are not always distinguishable by using multispectral sensors alone due to the low number of broad spectral bands. Understanding vegetation spectral reflectance variability is essential and can be used to map different vegetation species using remote sensing images [
22,
23]. In addition, understanding the seasonal evolution, i.e., the phenology [
24], of the reflectance spectra of key vegetation types can provide additional clues for distinguishing them from space [
25,
26]. It can help, for example, to identify the season when the difference between the band is at its greatest.
The first objective of this study was to document with in situ measurements the seasonal evolution of dominant vegetation reflectance spectra, along with biophysical properties, in a cold temperate intertidal system. Analysis of spectral measurements is crucial to determine appropriate spectral resolutions and a classification scheme. The four prevalent vegetation types included eelgrass (Zostera Marina), macroalgae (Ascophyllum nodosum, Fucus vesiculosus), saltmarsh cordgrass (Spartina alterniflora), and creeping saltbush (Atriplex prostrata). The second objective was to assess the potential of combining multispectral sensors onboard satellite constellations offering high spatial and temporal resolution (i.e., Sentinel-2 Multispectral Instrument (MSI), Landsat-8 operational land imager (OLI), RapidEye (RE), and PlanetScope (PS)) to quantify the phenological change of intertidal vegetation. Finally, we identified the best seasonal window to classify and map coastal ecosystems during the relatively short growing season (<6 months) in the studied system.
4. Discussion
In this study, we showed that spectral seasonal evolution in coastal and intertidal vegetation types can be detected by in situ and spaceborne sensors. This evolution is related to the biophysical metric of the vegetation, such as the LAI, but it depends on the vegetation type and the pigment composition. A better understanding of the vegetation phenology provides key insights for the coastal mapping based on spectral classification, with strengths and limitations discussed in the following sections.
We observed an important evolution of the vegetation biophysical metrics, with a growth phase from May to September, which was particularly evident for saltmarsh cordgrass (
Spartina alterniflora) and creeping saltbush. This observation was consistent with other studies [
53,
54,
55]. For eelgrass, the growth continued until the beginning of October, while the LAI showed an increase between May and June, remained constant until September, and was followed by a small decrease at the end of the season. Similar results were shown in other studies [
56,
57,
58].
Our vegetation
spectra measurement is consistent with other studies [
17,
59,
60,
61]. The in situ spectra analyses suggest that the general shapes of
are similar for all vegetation types (EG, MA, SC, and CS), but markedly different from the SD. Even if the shape is similar, they are different from each other and evolve over the season, as demonstrated using the SAM and the NDVI. This seasonal evolution is crucial for ecosystem remote sensing and is rarely mentioned and documented in other studies [
62]. The spectra from mid-June to the beginning of September often overlapped. At that time of the growing season, the distinction among vegetation types was less obvious. Furthermore, the evaluation of the pure vegetation spectra revealed that MA and EG were stable throughout the season under 100% cover. In contrast, CS and SC presented more variability during the growing season with similar spectra at the start (May and June) and at the end (September and October), but different in the middle of the season (July to August).
Sentinel-2 images were chosen to classify the intertidal and coastal vegetation types in the study area. Using the identical spectra for all images, the XGBoost algorithm (XGBoostnoSeason) was not able to distinguish the vegetation types in July and August. Nevertheless, good results were obtained in June and at the end of the growing season (September and October). The highest accuracy ( of 0.85) was obtained with the September 30th image at the onset of the senescence phase, suggesting that fully-grown vegetation is spectrally distinguishable from each other. We also demonstrate that accounting for vegetation phenology in the classification training dataset (XGBoostSeason) improved the classification accuracy by about 15%.
The NDVI was the most suitable proxy for the LAI among different VIs, and it was selected to evaluate the phenology from remote sensing using in situ spectra and multispectral images from four sensors (Landsat-8 OLI, Sentinel-2 MSI, PlanetScope, and RapidEye). The NDVI has been widely demonstrated to be a good descriptor of vegetation dynamics for many types of ecosystems, including wetlands [
17,
63,
64,
65,
66]. It is also widely used in satellite-based phenology monitoring, as it could be applied to almost any multispectral sensors on a wide range of platforms (in situ, drone, plane, satellite) [
67,
68]. Based on in situ NDVI measurements, SC and CS clearly showed a seasonal evolution, while MA, EG, and SD were more stable throughout the season. Frequent observations are needed to quantify the seasonal variability and evolution of the vegetation in a cold temperate environment due to a relatively short growing season (<6 months). Changes in coastal vegetation, such as growth, flowering, senescence, and shedding of leaves, take weeks and even months [
68,
69,
70]. For this reason, we recommend visiting the field as soon as the ice starts to melt (i.e., late March or beginning of April) [
69,
71]. In our study, we started the fieldwork in mid-May when the EG was already growing, with some stations showing 100% areal coverage.
By combining MSI and PS images, we could also track the vegetation phenology with a clearer signal for EG compared to the in situ. In general, satellite-derived phenology was consistent with the in situ NDVI values shown in this study and as presented by [
17,
60,
72,
73]. For example, Zoffoli et al. [
17] also used MSI images to document the phenology of intertidal seagrass (
Zostera noltii) in France. For comparison, these authors obtained 22 MSI images obtained at a low tide under cloud-free conditions for a ten-month period (March to December). With only six MSI images in our region, the combination with the PlanetScope constellation was necessary to fully capture the growing seasons and the following senescence. However, if our result demonstrates the feasibility, differences in spectral and spatial resolutions between sensors may complicate the combination of multisensor data for phenology monitoring.
As for Landsat-8 and Sentinel-2—both EO missions are dedicated to land cover monitoring and are widely used for phenology assessment. They both acquire images at lower temporal and spatial resolutions compared to PlanetScope. Even if the images are not available at a rapid rate due to tidal height and cloud cover restrictions, we can expect monthly images for the Sentinel-2 sensor, but much less for Landsat (one or two per year). Sentinel-2 images are easily usable and provide much better spectral resolution compared to the very-frequency acquisition of PlanetScope. Therefore, Sentinel-2 images allow further mapping capability by applying a classification algorithm using a high number of bands ( 10 bands). Still, Landsat is a very useful sensor, offering long-term time series for detecting and relatively good spectral capability for coastal mapping [
74,
75,
76]. In addition, the temporal resolution will improve with the recent launch of Landsat-9 in September 2021, opening the door for data fusion from MSI and OLI [
73,
77,
78]. The fusion of Landsat and Sentinel-2 images has tremendous potential to improve the ability to detect vegetation change and to cover all key periods of the vegetation phenology. However, in our case, combining the data was not necessary because the Sentinel-2 time series was already covering all of the vegetation phenology.
The PlanetScope sensor constellation is relatively new and offers many advantages, including high temporal resolution obtained at high spatial resolution compared to Landsat and Sentinel-2. Those images appear to be promising for extracting natural resources information, including intertidal ecosystem estimates and potentially even vegetation diversity estimates [
20,
79,
80]. As seen in our results, the sensor provides many images covering most of the vegetation key. The spatial resolution of PlanetScope images provides many details, but it is not without issues. Indeed, the variable radiometric quality, inconsistent radiometric calibration across multiple platforms, and low spectral resolution are central challenges for marine and coastal applications, such as vegetation classification and ecosystem monitoring [
20,
81,
82,
83]. The low spectral resolution of the first Dove generation limits its use for the vegetation type classification. Furthermore, the noise level is high, and the radiometric quality and inconsistency are low on clusters of pixels, especially in homogeneous pixels. This indicates the low signal-to-noise ratio of PlanetScope images [
20,
84]. This issue was encountered during the atmospheric correction, and even after the correction, noise can still be detected in the NDVI images. The next generation of PlanetScope sensors launched in January 2022 with a greater number of bands (eight bands) and a radiometric signal that may have helped resolve most of the limitations of the early sensor fleets.
The presence of water overlying the vegetation at the time of in situ data acquisition or in the images highly affected the spectral reflectance due to its high absorption in the red and the infrared. With high water levels within the vegetation, the species identification can be difficult, and further processes will be needed to use the images and spectra [
85,
86,
87,
88]. For this reason, the state of the tidal level could significantly influence the spatiotemporal distribution of remotely-sensed parameters, such as vegetation NDVI, which uses red and infrared wavelengths. For example, the bathymetric map combined with water level measurements during the data acquisition could be used to correct the water column interference to retrieve the bottom reflectance [
88]. However, estuarine water masses of the St. Lawrence are characterized by high concentrations of colored dissolved organic matter (CDOM) and suspended sediments that severely limit the light penetration, even in the visible bands [
89], and impair water column correction. Due to the loss of spectral information in the NIR, submerged vegetation indices need to be based on visible bands only [
40,
88], or located in the red-edge portion of the spectra [
59]. In conclusion, we recommend selecting images at the lowest tides possible to maximize the area to be mapped and to minimize the effect of water on the vegetation. The maximum tidal height that allows the vegetation to be mapped using multispectral data requires prior knowledge of the area. Furthermore, the mapped area will mainly depend on the area characteristics, such as the bathymetry/elevation and location of the vegetation.
Many atmospheric correction (AC) algorithms can be applied to multispectral images; this is a crucial step in the processing of remote sensing data for aquatic and coastal applications. Ideally, AC aims to separate the top-of-atmosphere observation by the satellite sensor into the signal from the atmosphere and the signal from the surface to retrieve surface reflectance [
90]. By using PlanetScope and RapidEye sensors, we are limited by the atmospheric correction algorithm due to a lack of spectral bands (as, for example, in the SWIR). Even though we can apply different atmospheric corrections, the algorithm applied to the image needs to be the same to compare the sensors together. Here, we adopted ACOLITE, as it allows the application of the dark spectrum fitting (DSF) atmospheric correction method to all imagery evaluated in this study. Furthermore, ACOLITE has been developed for the coastal environment and is currently widely used for aquatic-based applications, such as coastal water monitoring [
91,
92,
93,
94]. The sensitivity of satellite-derived NDVI phenology to AC could have been quantified, which was out of the scope of this study.
The area covered by our sampling sites was minimal (472,800 m) as we focused our effort on one EG meadow and intertidal vegetation section to maximize our frequent visits to the sites. This area was selected for its diversity of vegetation cover and easy access but was nevertheless quite representative of the entire coast of the region. The small size of the studied area helped us to develop excellent knowledge of the vegetation dynamics and the ecosystem structure. However, the sites were not optimal for the detection of small macroalgae patches (e.g., <1–2 m) with the limited satellite spatial resolution.
5. Conclusions
In this work, we assessed the seasonal dynamics of four typical intertidal vegetation types encountered in cold temperate coastal littoral, including macroalgae (Ascophyllum nodosum and Fucus vesiculosus), eelgrass (Z. marina), saltmarsh cordgrass (S. alterniflora), and creeping saltbush (A. prostrata). The seasonal evolution was determined based on biophysical characteristics (leaf area index), in situ reflectance spectra, vegetation indices, and classification of multispectral images. We identified a significant seasonal change in phenology of saltmarsh cordgrass and creeping saltbush. Even though some seasonal change could be observed for some vegetation types, no significant changes were observed in the in situ reflectance spectra for eelgrass and macroalgae. Moreover, we evaluated the potential of the NDVI to quantify the vegetation phenology from space. We demonstrated that the NDVI was the best vegetation index proxy to track the phenology that could be applied to multispectral cameras (including drones). Satellite-based NDVI, which strongly correlates with in situ values for saltmarsh cordgrass and creeping saltbush, were used to assess the potential of multispectral instruments to assess the phenology. By combining Sentinel-2 and planet imagery, we showed that the seasonal evolution of eelgrass NDVI was more evident than with in situ measures, likely because of the initial coverage of the quadrats (2500 cm). The extreme gradient boosted decision tree (XGBoost) algorithm was applied to a monthly time series of Sentinel-2 using in situ spectra as input spectral classes. The results indicate September as the best month of the year to classify coastal vegetation in our cold temperate environment, i.e., when the vegetation is fully grown and spectrally distinguishable.
Further work is required to monitor the vegetation species from this complex ecosystem located in a cold temperate climate with a relatively short growing season. We intend to extend the vegetation species mapping, especially for marshes that have a high plant diversity (
Salicornia maritima,
Spartina pectinata,
Spartina patens, etc.). Satellite remote sensing provides access to spatial scales, enabling the environment to be documented over vast areas. Widening the study area to cover all the coasts of the St. Lawrence maritime estuary and Gulf system would be interesting to know their conditions and interannual evolution. In addition, it will extend our knowledge of vegetated coastal ecosystems and their overall importance to the environment. Furthermore, it would be interesting to evaluate the carbon stock sequestration rates in coastal habitats (seagrass and marshes). Remote sensing tools are nowadays developed to quantify the extent of seagrasses and marshes, the species composition of these environments, and the above-ground biomass. In addition, some authors have demonstrated the possibility of estimating carbon stocks using empirical algorithms [
95,
96,
97]. With these types of data, it will be possible to document and monitor changes in carbon stocks and estimate emissions as functions of ecosystem degradation, conservation, and restoration. Finally, historical data could be used to assess the history of carbon stocks and emissions and the spatial distribution and changes of vegetation species.