*2.1. Spectral Data Acquisition and Processing*

The kelp submergence experiment took place on a marina dock in Victoria BC, on a sunny, cloudless day in September 2020. The Secchi depth during the time of the experiment was 7.5 m, showing relatively clear water, similar to general conditions for the coastal waters of the Salish Sea at the same time of year with low influence from riverine discharge, and low levels of total suspended-matter, chlorophyll-a, and colored dissolved organic matter present in the water column [28]. While the ranges of both Nereocystis and Macrocystis overlap on the British Columbia coast, only Nereocystis is found around the southern tip of Vancouver Island where this study occurred. The location and timing of the experiment allowed for the control of four criteria that we required: (1) controlled sea-state; with the dock acting as a shelter from any slight breezes, thereby minimizing variability in glint or light refraction due to ripples or waves on the water [29]; (2) platform stability, which minimized the potential errors during spectral acquisition due to the movement of both kelp and sensor that might occur in situ from a boat; (3) maintained environmental conditions expected in situ during peak biomass for local kelp, such as the inherent optical properties of water and optical constituents within the water column that would be difficult to reproduce in vitro; (4) a water depth (12 m) greater than the Secchi depth to minimize the influence of substrate reflectance on the above-water reflectance signal [30,31].

The experiment consisted of four separate trials. For each trial, a sample of *Nereocystis* was attached to a black frame made of high-density polyethylene (a plastic with low reflectance across the visual and near-infrared wavelength ranges), which was submerged from the surface to 100 cm in 10 cm increments on the sunlit side of the dock (Figure 1). Before each trial, radiance measurement of a Spectralon white-reference panel (Lspec(λ)) and an internal dark-current reading were taken to calculate reflectance (Table 1; Equation (1)) and reduce noise in the spectral data [32]. During each trial, ten individual above-water hyperspectral radiance measurements (LT(λ)) of kelp were collected at each incremental depth. Two of the four trials used the *Nereocystis* bulb, and two trials used the *Nereocystis* blades. Therefore, in total, 20 measurements of LT(λ) were collected for each kelp structure (bulb or blades) at each depth. After each trial, 10 radiance measurements were taken of the sky (Lsky(λ)) to be used in sky glint corrections [33]. Additionally, a total of 60 LT(λ) measurements were taken of water with no kelp within the field of view as a baseline for comparison with submerged kelp.

**Table 1.** Spectral parameters used to calculate above-water reflectance, as per Equation (1). All spectral measurements were collected using a calibrated ASD Fieldspec Handheld2 spectroradiometer with a one-degree fore optic (full viewing-angle), which detects a wavelength range from 325–1075 nm at 1 nm increments.


**Figure 1.** Side view of submergence experiment showing the geometry of acquisition for spectroradiometer and angle of zenith for the sun. Inset shows nadir view of the experiment with the azimuthal angle between spectroradiometer and sun and kelp blades inside the black frame. Diagrams are not to scale.

The solar elevation angle during the experiment was 46◦, which ensured sun-glint did not contaminate the spectra based on our geometry of acquisition [33,34]. LT(λ) measurements were taken at 5◦ from a nadir viewing angle to avoid reflection of the white spectroradiometer in the field of view on the water surface, and a sensor-sun azimuthal angle of 135◦ was used to minimize specular reflection in the field of view (FOV) [33]. Lsky(λ) measurements were taken at 5◦ from zenith at the same azimuthal angle as LT(λ). The spectroradiometer was held one meter above water, giving a footprint ranging from about 1.6 cm at the surface to 3.8 cm when the target was 100 cm deep. This small footprint was meant to ensure that the LT(λ) measurements contained 100% kelp, avoiding mixed pixel considerations [35].

$$R(\lambda)\_{0+}(\%) = \left(\frac{(\mathcal{L}\_{\rm T}(\lambda))}{\left(\mathcal{L}\_{\rm spec}(\lambda)\right)} - \frac{\left(\rho' \cdot \mathcal{L}\_{\rm sky}(\lambda)\right)}{\left(\mathcal{L}\_{\rm spec}(\lambda)\right)}\right) \times 100\tag{1}$$

Here, *ρ* was the proportionality factor of 0.0211, which relates the radiance measured directly from the sky to the estimated amount of sky radiance reflected off the sea surface based on wind, cloud cover, and geometry of acquisition [33]. *R*(*λ*)0+(%) for kelp at the surface (0 cm) was not subjected to the sky glint correction. Hereafter, *R*(*λ*)0+(%) values for kelp on the surface, submerged kelp, and water with no kelp are referred to as R0+ for brevity.

The R0+ spectra were first smoothed using a mean filter with a window of 5 nm to reduce noise while maintaining spectral features, and all spectra were then manually inspected for quality control. All bulb spectra were highly consistent, however, some blade spectra showed deviations in both the blue-green and NIR regions; likely due to water movement between blades as kelp was submerged, causing opened gaps in the "canopy" of blades attached to the platform. These spectra likely did not contain 100% blades within the field of view and were therefore removed from further analysis (Table 2). Despite the

removal of some blade spectra, the smallest sample size at any depth after quality control was at 90 cm with *n* = 10 spectral samples. Therefore, we do not expect that these removals biased the results of this study.

**Table 2.** Total number of each class of spectra before and after quality control was performed.


*2.2. Simulation of Micasense and WorldView Band R0+ and Indices*

After sky-glint correction, smoothing, and quality control of the spectra, R0+ measurements were simulated into bands of the WorldView-3 (R0+WV3) and the Micasense RedEdge-MX (R0+MSRE) sensors [36,37]. These sensors were chosen because both have a relatively high spatial resolution (WV3: 1.84 m; MSRE: ~1–10 cm), which is ideal for mapping kelp canopy in nearshore regions where it is likely to be submerged by tides and currents [16]. The R0+ at the bands of these sensors were simulated using Gaussian functions to estimate the sensor's spectral response for each band, based on full-width half maximum values of each sensor's band (Figure 2; Table 3). For a direct comparison, only the VNIR bands shared by both sensors were used for simulations.

**Figure 2.** Relative spectral responses at each band according to Gaussian functions were used to simulate the shared bands of (**a**) WorldView-3 (WV-3) earth observation satellite and (**b**) Micasense RedEdge-MX (MSRE) uncrewed aerial vehicle sensors—from left to right: blue, green, red, red-edge, and near-infrared band locations are shown.

**Table 3.** The effective bandwidths of the overlapping bands for both WorldView-3 (WV3) and Micasense RedEdge-MX (MSRE) sensors.


#### *2.3. Normalized Vegetation Indices*

Once the hyperspectral data were simulated into the respective sensor bands, the R0+ at these bands were used to calculate normalized vegetation indices (VIn; Equation (2)), which are commonly used to enhance spectral features of interest and reduce sensitivity to environmental influences within remote sensing imagery [38,39]. We tested several band combinations for VIn as different band combinations may increase or decrease the separability between kelp and water in an image [40].

$$\text{VI}\_{\text{lt}} = \frac{\text{band 2} - \text{band 1}}{\text{band 2} + \text{band 1}} \tag{2}$$

Because naming conventions for different VIn combinations are not ubiquitous across published literature, here, we referred to each VIn as the order in which bands appeared in the numerator of the VIn equation, separated by an underscore (Table 4).


**Table 4.** Vegetation indices calculated from simulated multispectral data.

One of the most commonly used VIn for kelp mapping is NIR\_R, which was originally used to detect terrestrial vegetation because of the high NIR and low red signal [38], but has since been used for kelp canopy detection due to the similar spectral characteristics between kelp canopy and terrestrial vegetation [14]. More recently, NIR\_R has been positively correlated with both the areal extent and biomass of kelp canopy [15,18,19]. However, various other combinations of visible and NIR bands have been used for kelp canopy detection with multispectral sensors. For instance, Schroeder et al. (2019b) used NIR\_R and NIR\_G for kelp detection with the WorldView-2 imagery. The NIR\_G combination may be more accurate for detecting a wide range of chlorophyll levels [41] and has generally been found comparable with NIR\_R in the detection of both floating and submerged vegetation [26]. Stekoll et al. (2006) found that NIR\_B and NIR\_G both provided higher kelp canopy and water separability in aerial imagery than NIR\_R. Further, recent comparisons with multispectral UAV and satellite imagery have shown that RE indices can improve separability of Macrocystis canopy and water when compared with NIR based indices [17,42], although this improvement was not specifically attributed to improved detection of submerged portions of the kelp canopy in either study.

Here, we compared the statistical differences in NIR and RE-based VIn values. R0+MSRE and R0+WV3 bands were used to calculate NIR\_B, NIR\_G, NIR\_R, and RE\_B, RE\_G, and RE\_R for both bulb and blades separately, for each depth. The statistical analysis was comprised of (i) VIn values compared with one another at each depth from the surface to 100 cm, and (ii) VIn values for water (with no kelp) compared to one another. First, the dataset was tested for normality, and while quantile–quantile plots suggested reasonable normality of the data distributions, Levene's test showed nearly all groupings for comparison displayed heterogeneity in variance. Therefore a non-parametric test was used in the analysis [43]. The Welch's ANOVA test was used to determine whether significant differences between VIn existed at each depth, and the Games–Howell post hoc test was used to determine which indices were significantly different from one another [44,45]. As part of the analysis, we focused on the statistical results comparing the RE and NIR counterpart indices only (e.g., NIR\_R & RE\_R, or NIR\_B & RE\_B) at each depth.
