Multi-Sensor Remote Sensing of Intertidal Flat Habitats for Migratory Shorebird Conservation

Abstract

Many species of shorebirds migrate long distances from their overwintering grounds in the southern hemisphere to breeding grounds in the northern hemisphere. The coastal intertidal zone, consisting of sand and mud flats exposed at low tide and covered at high tide, is heavily used as a migratory stopover or overwintering habitat. Understanding the spatial distribution of sediment types at these stopover sites is a critical step for understanding habitat use by shorebird species. Due to their importance as overwintering and stopover habitat for the imperiled western Atlantic subpopulation of the shorebird, the red knot (Calidris canutus rufa), as well as other migratory shorebirds, the northern coast of Brazil between Pará and Maranhão, and Bahía Lomas in northern Tierra del Fuego, Chile, were selected for further investigation as to the applicability of remotely sensed characterization of the intertidal flat habitats. Examination of the Landsat 8 multispectral reflectance and Sentinel-1 SAR backscatter reveals that sand and mud represent endmembers at opposite ends of a continuous gradient in feature space. While remotely sensed data can be used to discriminate between mud and sand intertidal types, the spectral relationships varied between the two very different geographic locations. The inclusion of both multispectral and radar sensing imagery can lead to important insights about the physical properties of the sediment that would be omitted by using one data source alone. Spectral unmixing techniques in Google Earth Engine were used to map the intertidal zone into general sediment classes spanning the gradient (i.e., mud, sandy mud, muddy sand, and sand). Comparison of the mapped outputs with field reference data suggests that mapping of mud- vs. sand-dominated areas can be accomplished with reasonable accuracy (overall accuracy of 75%).

1. Introduction

Many species of shorebirds migrate long distances from overwintering locations in the southern hemisphere to breeding grounds in the northern hemisphere [1,2,3,4,5]. Species rely on numerous stops along the migration route as staging or launch points for the next step of the migration [6,7,8]. The intertidal flats, consisting of sand and/or mud exposed at low tide and covered at high tide, are used extensively by migratory shorebirds as foraging habitat at coastal stopovers [1,9,10,11]. The grain size of the intertidal mud-sand sediment is a primary factor that determines the distribution of the benthic invertebrate fauna that the shorebirds feed upon [12,13,14,15,16,17]. Different species utilize different intertidal microhabitats, with species preferentially foraging on different substrates or at different times of the tidal cycle [12,18,19,20,21]. As intertidal habitats around the globe are changing rapidly because of rising sea levels [22,23,24], climate change [25], and human development pressure [26,27], understanding the spatial and temporal dynamics of intertidal habitat availability at critical stopover locations is key to developing effective conservation planning along major migratory flyways.

Several remote sensing platforms and analytical approaches have been used to map the intertidal zone and to characterize sediment type. Murray et al. [28,29,30] developed an effective method of isolating the intertidal zone across continental scale regions using Landsat multispectral remote sensing data, the normalized difference water index, and differencing between high-tide and low-tide images. Murray et al. [27] modified this approach with multi-temporal data and random forest machine learning to analyze over 700,000 satellite images to map the global extent and change in tidal flats over the span of 33 years (1984–2016). Others have refined these approaches, as well as harnessed the power of Google Earth Engine [31], to further improve the delineation of tidal flat areas [32,33]. While extremely successful at designating the binary extent of intertidal regions, these methods do not classify the intertidal zone into different sediment grain size classes. Research elsewhere has demonstrated the feasibility of remotely sensed discrimination of general sediment size classes. Optical and shortwave infrared reflectance data collected using airborne sensing platforms were classified by establishing spectral-end members of mud and sand, using linear mixture modeling to distinguish along the mud–sand gradient [34]. Landsat 5 satellite imagery was analyzed using both regression-based and spectral mixture methods to determine sediment size [35]. Ryu et al. [36] observed a positive linear relationship between Landsat near- and short-wave infrared image data and the percentage of coarser grained sediments.

These and other studies of multispectral remote sensing of the intertidal zone have noted the low spectral reflectance of mud in the visible and near-infrared spectrum relative to sand, typically caused by higher interstitial water content and organic matter retention in the finer grain-sized muds that increases absorption of the visible and near infrared wavelengths [37,38]. Using a bidirectional spectral reflectance model and laboratory-measured multi-angular spectra of soil samples obtained from tidal flats, Gao et al. [38] were able to model how the soil moisture content (interstitial water content) affected the photometric properties. The amount of residual surface water content can also be a complicating factor. Employing IKONOS imagery, Choi et al. [39] observed the opposite relationships of higher reflectance with the finer sediment grain size that they attributed to scarce remnant surface water on the mud flats in their study site. The timing of image acquisition in relation to the time since the last flooding tide, and thereby the amount of drying time, can also come into play. For example, Rainey et al. [40] reported that sand can show a lower reflectance than mud when intertidal sediments are completely saturated, suggesting that images collected after the longest time of intertidal exposure provide the best basis for differentiating sediment grain size by minimized interstitial and surface moisture.

Microwave synthetic aperture radar (SAR) backscattering can be affected by both surface roughness and the dielectric properties of the soil because of the soil moisture retention [41,42]. Typically, rough surface textures will have higher returns than smooth surfaces and very moist surfaces will have higher returns than dry surfaces. As such, microwave backscatter has been successfully used to classify mud and sand composition. Van der Wal et al. [43] correlated low backscatter from the ERS-1 and ERS-2 C-Band SAR satellite sensors with smoother surface textures and higher proportions of mud, while higher backscatter was associated with larger sediment sizes and rougher surface texture. Gade et al. [44] had similar conclusions using a combination of satellite platforms operating with a number of different radar bands (L, C, and X bands from ALOS PALSAR, ERS SAR, Radarsat-2 and ENVISAT ASAR, and TerraSAR-X, respectively). In both of these studies, the authors suggested that soil moisture content was consistently high across the study sites, making surface roughness the primary predictor of sediment size. Using TerraSAR-X imagery for a visual analysis and classification of intertidal habitats, Adolph et al. [45] noted that an essential aspect of the interpretation of SAR images of tidal areas is the varying presence of residual surface water cover. Finally, a fusion of image data from multiple platforms has also been used, combining the strengths of both optical/shortwave infrared’s response to moisture content and SAR’s response to surface roughness to successfully model sediment size using regression analysis [37].

In many of the remote sensing studies reviewed above, the research was focused on specific flats, estuaries, or coastlines, limiting the patterns observed to the specific regions examined. More recently, great strides have been made in harnessing the power of Google Earth Engine in identifying and mapping intertidal areas. However, there has been less progress in further characterizing the various types of intertidal habitat, in particular the dominant sediment type. This is critical baseline information useful for monitoring and informing management decisions concerning the quality and availability of shorebird foraging and roosting habitat. In this paper, two research questions are posed:

(1)

How well can a combination of multispectral reflectance and SAR backscatter and spectral unmixing techniques as implemented in Google Earth Engine be used to discriminate between mud and sand intertidal types?

(2)

Is the relationship between multispectral reflectance, SAR backscatter, and sediment type applicable across two very different geographic locations?

2. Methods

2.1. Study Sites

Due to their importance as overwintering and stopover habitat for the imperiled western Atlantic subpopulation of the shorebird, the red knot (Calidris canutus rufa) as well as other migratory shorebirds [1,2], the northern coast of Brazil between Pará and Maranhão, and Bahía Lomas in the Strait of Magellan, Tierra Del Fuego, Chile, were selected for further investigation as to the applicability of remotely sensed characterization of the intertidal zone. The northern coast of Brazil (Figure 1) within the states of Pará and Maranhão (between the Baia de Marajo on the west and Baia de Sao Marcos on the east) encompasses two Ramsar sites: the Maraja archipelago at the mouth of the Amazon River; and the Reentracias Maranhenses. This coast consists of “complex estuarine system [s] of extensive islands, bays, coves, and rugged coastline covered mainly by mangrove forest” and is considered critical migratory shorebird habitat by both RAMSAR (Ramsar Convention, https://www.ramsar.org/wetland/brazil (accessed on 8 August 2022)) and the Western Hemisphere Shorebird Reserve Network (WHSRN) (https://whsrn.org/whsrn_sites/reentrancias-maranhenses/ (accessed on 8 August 2022)). Current conservation protection is largely structured around the 12 Brazilian Marine Extractive Reserves (MER) that span this coast (Figure 1). The primary intention of the MERs is to protect the fisheries and livelihood of the resident fishermen through regulations on fishing practices, catch limits, or exclusion zones [46,47]. The intertidal mud- and sand-flat habitats that are critical to shorebirds also fall within the MERs’ boundaries [10,48]. However, the effectiveness of MERs as areas of protected shorebird habitat is not necessarily clear. Identifying and characterizing intertidal habitat is a critical first step in understanding how these MERs currently protect shorebird species, making future shorebird conservation decisions, and still fulfilling the MERs’ original mission.

The Bahía Lomas study site is located on the northern coast of the Chilean section of Tierra del Fuego (Figure 1). Bahía Lomas features the largest intertidal flats in Chile, fronting a 69 km long beach and several salt marshes and is both a RAMSAR site https://rsis.ramsar.org/ris/1430 (accessed on 8 August 2022) and a Hemispheric Importance Site of the Western Hemisphere Shorebird Reserve Network (WHSRN) https://whsrn.org/whsrn_sites/Bahía-lomas/ (accessed on 8 August 2022). Bahía Lomas hosts almost 50% of the western Atlantic red knot (Calidris canutus rufa) population, making the area among the most critical wintering ground for this species in South America [2,49]. Similarly characterizing and mapping intertidal habitat is critical baseline information useful for monitoring and informing management decisions that might affect shorebird foraging and roosting habitat.

2.2. Field Data Collection

As part of an effort to better understand habitat use in the two migratory/overwintering stopover sites, several combined shorebird and intertidal habitat surveys were conducted. Qualitative habitat observations were collected during point counts of wading shorebirds in the winter of 2016 and 2017 in northern Brazil, and a similar survey was conducted in January 2018 in Bahía Lomas (Figure 1). The surveys consisted of 250 m fixed radius plots spaced 500 m apart along walking transects in the intertidal zone during low to mid tide. Center points were recorded with GPS units and qualitative habitat observations recorded for each, focusing on the predominant sediment composition within the center of the 250 m radius plot, including the approximate sediment size class, color, and firmness. Approximately 630 survey points in Brazil and 175 points in Bahía Lomas were collected.

Sediment type was distinguished into four classes, mud, sandy mud, muddy sand, and sand, corresponding to the sediment grain size continuum outlined by Folk [50] and based on tactile observations using the “texture by feel” approach [51]. A more rigorous determination of sediment composition using laboratory analytical techniques was not feasible because of the remote field setting and the lack of access to laboratory facilities or the ability to transport sediment samples. Pictures were taken at most sites and surrounding landscape from the central point in conjunction with field notes to corroborate the classification of the dominant sediment type of each survey point.

Mud in both locations (Figure 2) was typified by a pocked surface texture because of invertebrate (crab and worm) holes and bird tracks. Unique to Bahía Lomas were extensive areas of dried mud, with large plates or flakes and a brighter or lighter appearance (Figure 2). Sand sediment exhibited a wide range of colors across the northern Brazil study site, ranging from a yellow-orange to light gray. Sand in Bahía Lomas was typically much darker and had a more consistent gray color in appearance, though both locations often had a rippled or washboard-like surface (Figure 3).

2.3. Landsat Reflectance Data Acquisition/Preparation

A flowchart illustrating the various steps undertaken in the satellite imagery processing and analysis can be found in the Appendix A. Landsat 8 Surface Reflectance Code (LaSRC) imagery products were accessed for the entirety of the study areas from the USGS archive through http://earthexplorer.usgs.gov/ (accessed on 23 January 2019). LaSRC imagery, originally produced by the US Geological Survey, are processed to account for atmospheric interference caused by scattering or absorption, resulting in normalized surface reflectance values for each image. Due to the equatorial location of the Brazilian study site, a large number of available images had significant cloud cover (>40%), severely limiting image choice. As such, no preference was given to date or time of year for each image, only for cloud percentage and tidal stage. To compensate, multiple high-tide and low-tide images for each scene were acquired, totaling 20 high-tide and 27 low-tide images from 2013 to 2018 for Brazil. Clouds and cloud shadows were removed using the Landsat 8 surface reflectance values and top-of-atmosphere thermal data via the methods outlined by Martinuzzi et al. [52]. In short, thresholds were used to create cloud cover masks, which were shifted and expanded to cover corresponding shadows. The clouds and shadows were masked by the combined cloud and shifted-expanded shadow rasters. For each scene, the cloud/shadow masked images were composited by filling in cloud/shadow gaps of one image, with the cloud/shadow free regions of the other images. Pixels that were cloud or cloud shadows in all available imagery were coded as NA. Composite images were created for each scene in the Brazilian study region, then mosaicked together, to create a site-wide image for both high-tide and low-tide imagery for subsequent analysis.

The Bahía Lomas study area proved much less troublesome, being smaller in size, encompassed by one Landsat scene. In addition, cloud cover was much less problematic, with one cloud-free high tide (2017) and one cloud-free low tide (2016) image acquired for the study region.

For all images, tidal stage was determined by the methods outlined by Murray et al. [28], using Oregon State University’s tidal model [53], the date and time of each image’s acquisition, and a user-established reference point for each scene. The Oregon State University tidal model is a generalized inverse model of barotropic ocean tides, with versions designed at both global and select regional scales (http://volkov.oce.orst.edu/tides/ (accessed on 9 May 2018)). One of the regions with a specifically calibrated tidal model was the Amazonian drainage basin, where the Brazil study is located. A second tidal model had been calibrated by OSU for the Patagonian Shelf, where the Bahía Lomas study is located. Where possible, extreme tidal-stage imagery (i.e., higher high tides and lower low tides, sometimes referred to as spring tides) was used.

2.4. Sentinel-1 SAR Data Acquisition/Preparation

Low-tide vertical–vertical (VV) and vertical–horizontal (VH) images from the SENTINEL-1 C-SAR Level-1 ground range detected products (https://sentinels.copernicus.eu/web/sentinel/home (accessed on 18 September 2019)) by the European Space Agency. SENTINEL-1 captures the “C band” portion of the electromagnetic spectrum in the microwave range of frequencies, ranging from 4.0 to 8.0 gigahertz, and wavelengths of 3.9 to 7.5 cm. The tidal stage was determined using the same methods outlined above, with three images acquired per scene, totaling 15 images for Brazil and 3 for Bahía Lomas. A 3 × 3 mean filter was used for noise reduction and then images were coarsened to 30 m pixels to match the resolution of the Landsat data. For each scene, the mean pixel value of the three processed images was used to account for minor variations in the tidal stage, geo-registered to the Landsat 8 data, then mosaicked together resulting in a low-tide vertical–vertical image and vertical–horizontal image for the study region.

2.5. Intertidal Mask

An intertidal zone-only image mask was developed for both locations using methods outlined by Murray et al. [28]. A Landsat normalized difference water index (NDWI = (green − near infrared)/(green + near infrared)) was calculated using the mosaicked Landsat 8 imagery, enhancing the presence of surface water. Water tends to have higher reflectance in the green wavelengths than in the near infrared; on land this is often reversed. Where water is present, the difference between green and near infrared will be positive and, as a result, the NDWI will be positive. A threshold was determined by visual inspection to designate all NDWI values above the threshold as water, all values below the threshold as land, for both high-tide and low-tide images. The high-tide-classified NWDI land image was then differenced from the low-tide NDWI land image, isolating the intertidal zone (i.e., those areas exposed at low tide but inundated at high tide; Figure 4). This binary intertidal zone map was used to mask the stacked images comprised of the cloud-free Landsat 8 surface reflectance and smoothed Sentinel-1 backscatter data for subsequent analysis at both locations. The above preprocessing steps were accomplished locally using ERDAS Imagine software. The resulting cloud-screened, intertidal-masked image products were uploaded to the Google Earth Engine (GEE; [31]; https//earth.engine.google.com (accessed on 1 April 2021)) for further analysis.

2.6. Spectral Pattern Analysis

An exploratory data analysis was conducted to compare the similarity or difference in spectral response for the different sediment types across the two study areas. A 250 m buffer was generated around each field reference survey point and the intertidal-masked remote sensing data extracted (Figure 4). A 250 m buffer was used in this analysis to correspond with the approximate area for which the birds were observable and countable from the center survey point. The mean pixel value for each band was calculated for each plot using ArcGIS’s spatial analyst, zonal-statistics-to-table tool. These means were compared via Tukey HSD tests to evaluate whether the habitat classes (mud, mixed, and sand) were significantly different at the two study sites. To compare the classes between locations, (e.g., sand in Brazil vs. sand in Bahía Lomas), the mean pixel values were normalized as a percentage of the mean of the band values of that study site (e.g., (mean near-IR pixel for plot [i] in Brazil/(mean of all Brazil’s near-IR plots)), resulting in values that essentially represent the relative reflectance values rather than absolute reflectance. Absolute reflectance values can be different based on sensor characteristics (i.e., sensor azimuth) between dramatically different latitudes [54], so relative reflectance more accurately captures the different patterns between locations. Statistics and relevant figures were produced with Rstudio software (R Core Team 2013) and the ggplot2 package.

2.7. Unmixing Analysis and Classification

As the intertidal flats often present a gradient from sand to mud with mixtures in between, we conceptualized individual pixels arrayed along this gradient (i.e., some locations had pure sand or mud but elsewhere a mixture of the two). When working with mixed image pixels, a spectral unmixing approach is commonly used to estimate the proportions of the component ground covers [55]. A constrained linear spectral unmixing algorithm coded in GEE using the UNMIX function (https://developers.google.com/earth-engine/guides/image_transforms (accessed on 1 June 2021)) was applied to determine the ground cover composition of the intertidal pixels. The reflectance at each pixel of the image is assumed to be a linear combination of the reflectance of each ground cover present within the pixel. In this case, the spectral endmembers were “pure” sand at one extreme and “pure” mud at the other in the spectral feature space (Figure 5). The unmixing algorithm determined the relative proportion of each ground cover (spectral endmember) in each pixel. A linear model was chosen based on our analysis of the spectral response, as well as on prior work using unmixing techniques to characterize intertidal habitats [34,35,36]. The endmember spectral values were chosen by extraction from a small sample of field reference sites, which were then plotted in feature space (Figure 5) to determine the endmember values (i.e., toward the outer envelope of the feature space for “pure” sand and “pure” mud). A constrained model was employed to limit the percentages between 0 and 100%.

The resulting unmixed ground cover proportion maps were then categorized to create a discrete habitat map with four general classes (mud, sandy mud, muddy sand, and sand) in Bahía Lomas and three classes (mud, mixed, and sand) in the northern Brazil study area to match the field reference classification schema. For example, sandy mud represented between 25–50% percent sand (as estimated by the spectral unmixing). An accuracy assessment was undertaken by comparing the field survey observation data (the reference data independent from any calibration training data) and the remotely sensed determination for each field reference point. Some of the field reference points were excluded because of cloud cover or other anomalies in the imagery data. The unmixed outputs were also categorized into a two-class map (mud vs. sand-dominated) to examine the efficacy of identifying the dominant endmember components.

3. Results

3.1. Spectral Pattern Analysis

As evidenced in Figure 5, the continuous gradient in sediment composition is paralleled with a near-linear response in spectral reflectance. Overall, as sediment grain size increased in the northern Brazil study area (mud to sand), reflectance increased (Figure 5 and Figure 6). This situation was reversed in Bahía Lomas, as sediment grain size increased (mud to sand) reflectance decreased (Figure 5 and Figure 6). SAR backscatter appeared to respond to sediment grain size in a more consistent fashion with backscatter decreasing as sediment size increased in both study locations, though it was more pronounced in the Bahía Lomas study area.

The reflectance of mud (Figure 7A) and sand (Figure 7B) classes for all Landsat 8 bands were significantly different between Brazil and Bahía Lomas (

Table 1. p-values from Welch’s t-test comparing normalized sand values between Brazil and Bahía Lomas, and normalized mud values between Brazil and Bahía Lomas, with p-values < 0.05 highlighted in gray.

	Band	Normalized Sand	Normalized Mud
Landsat 8	Blue	<0.001	<0.001
	Green	<0.001	<0.001
	Red	<0.001	<0.001
	Near IR	<0.001	<0.001
	SWIR 1	<0.001	<0.001
	SWIR 2	<0.001	<0.001
Sentinel-1	Vertical-Vertical	0.553	0.082
	Vertical-Horizontal	0.011	0.638

). This was not the case with Sentinel-1 backscatter data, with sand vertical–vertical not being significantly different between locations (Table 1), and mud not being significantly different for both vertical–vertical and vertical–horizontal.

3.2. Spectral Unmixing and Classification

Figure 8 and Figure 9 illustrate the mean and range in sediment end-member composition values for mud, mixed, and sand field reference plots. The estimated percentage of mud decreases and sand increases as one compares the mud vs. mixed vs. sand-dominated reference plots. The % endmember composition was translated into nominal sediment class categories for a more traditional accuracy assessment. Overall accuracy of the three or four class maps were moderate at 59% and 45% for the northern Brazil and Bahía Lomas, Chile, sites, respectively (

Table 2. Error matrix for the northern Brazilian study site’s accuracy assessment. Cells filled with gray are where both the classified and reference land covers agree. Overall accuracy and Cohen’s kappa coefficient are listed in the bottom right.

(A) 3 Class Sand-Mixed-Mud Map (n = 475).
		Reference
		Mud	Mixed	Sand	Sum	User’s
Classified	Mud	86	27	68	181	48%
	Mixed	8	9	76	93	10%
	Sand	9	7	185	201	92%
	Sum	103	43	329	475
Producer’s		83%	21%	56%	Overall	59%
					Kappa	0.33
(B) 2 Class Sand vs. Mud-Dominated Map (n =232).
		Reference
		Mud		Sand	Sum	User’s
Classified	Mud	60		58	118	51%
	Sand	3		111	114	97%
	Sum	63		169	232
Producer’s		95%		66%	Overall	74%
					Kappa	0.48

and

Table 3. Error matrix for the Bahía Lomas, Chile, site’s accuracy assessment (n = 174). Cells filled with gray are where both the classified and reference land covers agree. Overall accuracy and Cohen’s kappa coefficient are listed in the bottom right.

(A) 4 Class Sand-Mixed-Mud Map
	Reference
Classified	Mud	Sandy Mud	Muddy Sand	Sand	Sum	User’s
Mud	5	3	2	0	10	50%
Sandy Mud	18	7	6	1	32	22%
Muddy Sand	3	10	16	3	32	50%
Sand	2	19	28	51	100	51%
Sum	28	39	52	55	174
Producer’s	18%	18%	31%	93%	Overall	45%
					Kappa	0.23
(B) 2 Class Sand vs. Mud-Dominated Map
			Reference
			Mud	Sand	Sum	User’s
Classified	Mud		33	9	42	78%
	Sand		34	98	132	74%
	Sum		67	107
Producer’s			49%	92%	Overall	75%
					Kappa	0.44

). When considering only endmembers (mud and sand) and excluding the mixed class, all accuracy ratings, as well as the kappa coefficient, were higher with an overall accuracy of 74% and a kappa of 0.48 for the Brazil study site and an overall accuracy of 75% and kappa of 0.44 for the Chile study site (Table 2 and Table 3).

3.3. Habitat Characterization

Examination of the results of the field surveys conducted in Brazil (2016) and Chile (2018) shows the variation in habitat usage by the dominant species of migratory shorebirds (Figure 10A,B). At the Brazil study site, sanderlings (SAND), red knots (REKN), and semipalmated plovers (SEPL) were observed feeding with the highest frequency in sand-dominated habitats. The whimbrel (WHIM) was more of a generalist, while ruddy turnstones (RUTU), black-bellied plovers (BBPL), and semipalmated sandpipers were observed feeding more frequently on mud-dominated habitat. At the Chile study site, red knots (REKN) were observed most frequently feeding on the mixed sand–mud flats while the white-rumped sandpiper (WRSA), and to a lesser extent the Hudsonian godwit (HUGO), was observed feeding on the sand-dominated and mixed flats. The WRSA was the only one of the three species observed feeding in the mud-dominated areas.

Classification of the unmixing analysis reveals that the Brazilian study area is composed primarily of mud-dominated sediment type with 48% mapped as mud, 29% as mixed mud-sand, and 23% as sand-dominated (

Table 4. Intertidal land cover across the Brazilian study site (total) broken down into the area (ha) and relative composition (% of MER) within the marine extractive reserves, and the area (ha) and relative composition (% of out) outside the marine extractive reserves.

Intertidal Sediment Type	Area Inside MER			Area Outside MER			Area Total
	Ha	% of MER	% of Total	Ha	% of Out	% of Total	Ha	% of Total
Mud	21,944	50.0%	19.5	32,686	47.5%	29.0%	54,630	48.5%
Mixed	10,658	24.3%	9.5	21,720	31.5%	19.3%	32,378	28.7%
Sand	11,253	25.7%	10.0	14,455	21.0%	12.8%	25,708	22.8%
Total	43,855		38.9%	68,861		61.1%	112,716

). This coastline is composed of long fingers of mangrove swamps interspersed with sheltered coastal bays and creek channels fringed by mud flats. The outer fringe of the shoreline is fronted by sandy barrier beaches and intertidal sand bars. The field surveys shows that migratory shorebirds feed on all three habitat types, though some species are found more frequently on one type vs. another (e.g., red knots appear to prefer the sand-dominated flats vs. semipalmated the mud-dominated flats) (Figure 10A). The MERs cover approximately 39% or the intertidal flat habitat with the same relative percentage of sediment class breakdown occurring within the MERs versus outside (Table 4).

The Bahía Lomas, Chile, study area is composed primarily of sand-dominated sediment with 53% mapped as sand followed by muddy sand and muddy sand, and it is only approximately 5% mud-dominated (

Table 5. Intertidal land cover across the Bahía Lomas, Chile, study area, broken down into the area (ha) and percent of study area (%).

Intertidal Sediment Type	Area (ha)	Area (%)
Mud	1664	5.4%
Sandy Mud	6129	20.0%
Muddy Sand	6740	22.0%
Sand	16,110	52.6%

). Bahía Lomas is much more open to wave action with large areas of sand-dominated flats, while mud-dominated flats occur only in the back bay zone and are flooded only during the highest tides. Our species of particular concern, the red knot, was observed most frequently on the mixed sand–mud flats (Figure 10B), which represent over 40% of the bay’s intertidal flat area.

4. Discussion

Our first research question focused on the efficacy of employing multispectral reflectance and SAR backscatter to discriminate between mud and sand intertidal habitat types in both northern Brazil and Bahía Lomas. While the endmembers of mud and sand appear to be sufficiently spectrally distinct, the spectral feature space represents a continuum of response (Figure 5 and Figure 6). The second research question examined the universality of the relationship between multispectral reflectance, SAR backscatter, and sediment size between the locations. Based on the normalized remote sensing data, Brazil and Bahía Lomas’ sand and mud classes are significantly different (p < 0.05) across the visual and infrared spectrum (Table 1). However, the vertical–vertical response was not significantly different between study sites (p > 0.05) for both land-cover types (mud and sand). These results are likely due to the nature of reflectance and backscatter and how they interact with the physical properties of the intertidal substrate. General observations between the two locations hint at similarities and differences of the physical properties of the sediment. Mud substrates appear relatively similar in surface texture between the two sites; both have a pocked, complex surface structure. Bahía Lomas does distinguish itself with the occasional cracked, dry, and plated mud surfaces (Figure 2). The muds of Bahía Lomas were lighter in color than those at the northern Brazilian study site, which were dark gray to brown in color, possibly because of higher organic matter content. The Bahía Lomas muds were consistently drier with no residual surface water. The surface texture of sand for both locations appears consistently similar, with rippled or washboard surfaces. However, the visible color of sand differs noticeably between the two sites, with Bahía Lomas consistently being darker and Brazil having some regions with a more orange or red tone (Figure 3). Two general patterns emerge: differences in substrate color and similarities in surface texture between both locations, with some exceptions in Bahía Lomas’ platey, cracked mud. These patterns continue to play out in the observed Landsat reflectance (color) and Sentinel backscatter (surface texture) data.

Typically, reflectance is influenced by soil moisture, with smaller grained soils (higher proportions of finer-grained silt and clay, i.e., mud) retaining a higher concentration of water than larger-sized soils (i.e., sand). Increased water retention results in lowered reflection across all visual and infrared bands. The Brazil study site followed the same patterns as previous studies [34,35,36,37] with reflectance increasing as sediment size increased (Figure 6). Saturated conditions did decrease the reflectance but, even so, sand-dominated flats were consistently brighter than mud. Bahía Lomas exhibited the inverse relationship, with reflectance decreasing as sediment grain-size increased (Figure 6). This trend is even more apparent when comparing normalized reflectance data, with Bahía Lomas consistently having higher relative mud-reflectance (Figure 7A), and lower relative sand-reflectance (Figure 7B). Rainey et al. [40] and Choi et al. [39] attributed the lower reflectance of sand they observed as compared to mud to the higher interstitial or remnant surface water content of the water-saturated sand flats. Field observations corroborated earlier work by Espoz et al., [56] that the intertidal areas of Bahía Lomas that were consistently flooded and wave-washed at high tide were sand dominated, while the lower-energy inshore areas that flooded only during the highest monthly tides were mud dominated. The highest-elevation mud-dominated areas regularly dried out, resulting in a lighter-toned surface along with desiccation cracking and polygonal contraction plates. Subsequent periods of extreme high tide or storm-driven flooding created erosion channels and further topographic heterogeneity. While one specific date of low-tide imagery was used in the spectral mixing analysis, visual examination of other low-tide images for the Bahía Lomas study site showed consistent patterns of lower reflectance for the sand-dominated flats vs. higher values for the mud-dominated flats.

The two physical characteristics driving radar backscatter values in sediment are surface roughness and dielectric properties that are due to surface moisture content. Van der Wal et al. [43] found that surface roughness was the strongest correlate with sediment size compared to other variables, including surface moisture. Soil moisture content can significantly influence how much electromagnetic energy penetrates, with high levels of moisture preventing penetration and increasing backscatter, while low levels of moisture facilitate penetration and reduce backscatter [41,42]. Where soil moisture content is consistently high across the study sites, surface roughness is generally the primary predictor of radar backscatter. The similar patterns in backscatter between Bahía Lomas and Brazil (Figure 7A,B and Table 1) suggest that the locations have similar responses to the combined effects of surface roughness and moisture content. Van der Wal et al. [43] found that smooth surfaces were associated with mud and lower backscatter and rough surfaces with sand and higher backscatter. Our results differed from van der Wal’s, in that intertidal mud locations had higher backscatter values than intertidal sand in both study sites (Figure 6A,B). This would suggest that though the mud habitat has a finer sediment size, the surface roughness caused by pocking by tracks and worm/crab holes, as well as drying cracks and erosion channels, may result in a greater C-band radar backscatter than the rippling/washboard observed in the sandflats. While the C-band radar appears to have a more consistent relationship across field sites, the dual response to surface moisture and surface roughness can be difficult to disentangle.

The efficacy of linear spectral unmixing analysis in translating the continuum in spectral response into a sediment type/habitat map for conservation management purposes was assessed. Comparison of the field reference plots and the results of the spectral unmixing analysis shows a consistent decrease in the % of predicted mud endmember and conversely an increase in % sand as one progresses from the mud to the mixed to the sand reference plots in both the Brazil and Chile study sites (Figure 8 and Figure 9). While one potential output of the unmixing analysis can be maps of % sand or mud components (out of 100%), the results can also be segmented into discrete sediment type classes (i.e., for the three-class map, mud was defined as > 66% mud component). Assessment of the resulting classified map using a “traditional” accuracy assessment approach reveals that the predicted class and the field reference class only corresponded 45–59% of the time (Table 2 and Table 3). When the classification was simplified into only two classes, mud- vs. sand-dominated, the accuracy increased to approximately 75% but, as reflected in the middling kappa values (Table 2 and Table 3), these results can only be considered moderately accurate. Further examination of the user’s and producer’s accuracy did not show consistent patterns between the two study sites. The high user’s accuracy (>90%) for the sand class in the Brazil study site suggests the user can be confident that areas classed as sand were sand but that a large percentage of pixels classed as mud were actually sand dominated (Table 2). The high producer’s accuracy (>90%) for the sand class and low producer’s accuracy for the mud class for the Bahía Lomas site suggests that the remotely sensed mapping is misclassifying some areas of mud as sand (Table 3). The results suggest that while using multispectral reflectance and SAR backscatter can effectively discriminate between mud and sand spectral endmembers, disentangling the middle of the continuum remains challenging. In this study, only a comparatively simple two-endmember linear unmixing model was employed; application of a nonlinear model or inclusion of other spectral endmembers might improve the discrimination.

Our approach of field-determined sediment grain-size based on tactile observation left room for greater error because of possible observer-to-observer differences. Limited access to in-country soil labs because of the remote field locations and the difficulties around exporting soil samples back to the United States were the primary reasons for utilizing this more qualitative approach. While not as robust as systematic soil laboratory analysis, our approach had the benefits of low cost and rapid data collection, while still successfully classifying endmembers and identifying sediment-size/remote sensing trends. Yates et al. [35] employed more rigorous lab-determined sediment-size data but still struggled with classifying the mud–sand continuum into discrete classes. Ryu et al. [36] observed a strong response between remotely sensed response and a critical grain size of 0.25 mm (medium sand) but suggested that other techniques were needed to discriminate finer grain sizes (i.e., <0.0625 mm, silt).

Common to many remote sensing applications, temporal mismatch between when the remote sensing image was taken and when the in-situ data were collected can lead to conclusions that do not accurately reflect ground conditions [54]. Several of the intertidal mapping studies employed airborne imagery [34,40] to lessen the time lag between image acquisition and field data collection. Residual surface water and interstitial water content vary on short time scales (i.e., minutes to hours). While the great number of dates of imagery from a variety of platforms that are accessible through the GEE image catalog increases the utility of satellite-based mapping and the monitoring of intertidal areas, it does not fix the temporal mismatch problem. Remote sensing of the intertidal zone is especially problematic as cloud-free imagery is needed to correspond with both extreme low-tide and high-tide conditions. Landsat 8 image selection over the northern Brazil study site was more limited by significant cloud cover than the more semi-arid Bahía Lomas study site. This required compositing several cloud-removed images for each scene by acquiring multiple low-tide images over several years and calculating the average pixel value. While single-date imagery was available over Bahía Lomas, there was still several years’ difference in time between the field and image data collection. Clouds were not problematic for the Sentinel-1 radar imagery acquisition because microwave radar penetrates cloud cover, though acquisition during the needed tidal stage can still be limiting. Even if extreme high- and low-tide imagery is obtained, a temporal mismatch of several years can result in differences that are due to changes in the ground conditions in this very dynamic intertidal zone (i.e., as a result of storm-driven over-wash and sediment erosion and deposition). An undetermined portion of the inaccuracy of our results is likely due to this temporal mismatch.

The results suggest that the spectral unmixing analysis using multispectral reflectance and SAR backscatter can effectively discriminate between mud- vs. sand-dominated sediment types and provide intertidal habitat maps of sufficient quality to inform management and conservation. Analysis of the resulting maps revealed that the approximate composition of the intertidal area found within the Brazilian MERs is proportional to the area that is protected; in other words, no intertidal habitat type is preferentially protected nor excluded by the MERs (Table 4). While this is reassuring, it is also clearly the first step in understanding how MERs might serve to protect the intertidal habitats and the species that rely upon it. There was no “inside” vs. “outside” comparison for the Bahía Lomas study area, as the entire bay is nominally protected. However, the classified unmixed imagery does provide an areal estimate and a mapping of the spatial distribution of the general classes of intertidal sediment types that are important as foraging habitat. Alternatively, rather than estimating the dominant sediment class as a precursor to further classification, other studies have employed a machine learning approach to identify and map specific intertidal habitat features of interest. Though not concerned with bird habitat per se, Zhang et al. [32] employed multi-temporal feature extraction, machine learning classification using GEE, and morphological post-processing to map intertidal vs. supratidal barren vs. vegetated flats in China’s eastern coastal zone. Henriques et al. [57] combined Sentinel-1 SAR and Sentinel-2 multispectral imagery and a digital elevation model (DEM), along with random forest machine learning techniques, to map intertidal habitat types of importance to migratory shorebirds and their benthic prey at a stopover location along the east Atlantic flyway in West Africa. Mapped habitat types included rocks, shell beds, and macroalgae and distinguished between areas of sandy and mixed bare sediment and areas occupied by fiddler crabs. A random forest classification approach, as compared to the linear unmixing approach we employed, is especially appropriate where different habitat types occur as discrete patches rather than smoothly intergrading.

5. Conclusions

The first step to informing the management and conservation of shorebird foraging habitat is isolating and mapping the spatial extent of the intertidal zone. A critical next step is a characterization of the areal amount and spatial distribution of the different intertidal habitats. Similar to previous research, we confirmed significant relationships between sediment size and visual, near/mid-infrared, and radar remote sensing. The mud–sand sediment continuum is represented by a continuous gradient in spectral feature space, with mud and sand endmembers readily distinguished, but middle, mixed classes less so. Our results show that spectral unmixing analysis of medium resolution Landsat and Sentinel-1 SAR imagery provides an effective means of quantifying this continuous gradient between sand and mud sediment types on coastlines with extensive areas of intertidal habitat. Other medium resolution multispectral sensors such as Sentinel-2 should be equally applicable. The resulting mapping can also be segmented into qualitative habitat types if that is desired to inform conservation management. Google Earth Engine provides ready access to an immense catalog of multispectral (both Landsat and Sentinel-2) and Sentinel-1 SAR imagery [31,58], as well as the spectral unmixing tools to make this approach feasible for any coastline in the world, provided that field reference data are available for initial training and validation.

For some conservation and management decisions, the global tidal flat maps produced by Murray et al. [27] and available on GEE may be sufficient. However, as all intertidal areas are not used equally (i.e., do not have the same value as habitat), more detailed habitat maps may be needed to inform conservation management decisions concerning specific shorebird species. While our results suggest that the highly accurate discrimination of discrete classes of sediment type using medium-resolution remotely sensed imagery is challenging, a broad-scale mapping of the spatial distribution of intertidal habitat types is feasible. The resulting maps can then serve as input to shorebird species distribution models, or, alternatively, the remotely sensed reflectance values can be used as direct inputs, rather than classifying remotely sensed imagery into intertidal habitat types as was done here [59]. Shorebirds have demonstrated clear resource partitioning across different intertidal sediment types [12,18] and can have their distributions shifted by changes in sediment composition that are due to channel dredging and beach management practices [60,61,62,63] or to complete loss altogether [26,27]. Spaceborne remotely sensed mapping provides a feasible approach to provide baseline mapping as shown here, as well as to monitor changes in intertidal habitat composition that may significantly affect the quality and availability of shorebird foraging habitat.