There are several assumptions used in the nearest-neighbor atmospheric correction and spectral unmixing (Eq. 4). Violations of these assumptions will cause errors in the derived RFM spectra. For example, if the atmosphere over the floating-matter pixel is different from over the nearby water, the nearest-neighbor atmospheric correction may not be applicable. In practice, however, because the target and reference pixels are very close (<1 km), such a violation is unlikely. In Step 4, the water within the floating matter (FM)-containing pixel is assumed to be the same as the nearby water. Because of the close proximity of the two pixels, this assumption should be valid for most cases unless the FM-containing pixel is at an ocean front where different water masses converge. The departure of RFM(754) from the assumed 0.3 will also lead to errors in the estimated χ (and therefore RFM). However, as long as RW (i.e., RR) in Eqs. (4) and (5) is ≪RFM, the shape of RFM is still retained, although the magnitude departs from the “truth” in proportion to the departure of RFM(754) from 0.3. Indeed, the condition of RW≪RFM can be satisfied for λ >600 nm for most floating matters unless the water is extremely turbid. Even for turbid waters, for certain floating matters where RFM is elevated at λ>530 nm (e.g., red Noctiluca, brine shrimp cysts, ice), the shape of the derived RFM should still be valid for λ>530 nm. Indeed, when RW is ≪RFM, even a simple subtraction of Rrc or top-of-atmosphere radiance between the target pixel and reference pixel, as demonstrated in Gower et al. (2006), may retain the spectral shapes of floating matters.

Another uncertainty source can come from the assumption of linear mixing between floating matters and water (Eq. 3). For macroalgae, linear mixing up to the reflectance saturation level has been shown in laboratory experiments (Hu et al., 2017; Wang et al., 2018). As long as the macroalgae stay on the very surface of the water (as opposed to being submerged under the surface), this assumption should be valid not just for macroalgae but for all floating matters. For the same reason, if certain portions of kelp are submerged in water, large uncertainties may result from the linear unmixing scheme. Under high-wind conditions, the strong mixing may result in submerged algae (especially for microalgae), thus violating the linear mixing rule. However, the cases presented in Figs. 2–5 were selected very carefully to avoid high wind speed (>5 ms-1, where wind speed was obtained from the National Centers for Environmental Prediction). Therefore, such mixing-induced uncertainties are unlikely.

Additional uncertainties may come from the HICO radiometric calibration, which affects Rt and all derivative products. Through the use of the Marine Optical Buoy (MOBY) and other clear-water sites, HICO has been calibrated vicariously (Ibrahim et al., 2018), which has resulted in significant improvements in the retrieved Rrs over water as compared with data without vicarious calibration. However, after the vicarious calibration, while the spectral shape of Rrc over water appears correctly, the shape of ΔRrc over land appears to be biased low at λ>800 nm. Without vicarious calibration, the opposite is observed. This is possibly due to the non-linear effects in the detector response to incoming light, and currently there appears no reliable way to address this issue (Amir Ibrahim, personal communication, 2021​​​​​​​). Similarly, calibration for λ<450 nm may be subject to larger errors than for λ between 450 and 800 nm. Therefore, RFM in the range of 800–900 nm is omitted here, and interpretation of 400–450 nm also requires more caution. Similarly, the spectral wiggling between 700 and 800 nm (e.g., Fig. 3b) appears to come from residual errors in correcting water vapor absorption and oxygen absorption in the atmosphere. Therefore, although the spectral wiggling does not affect the overall shape of the red-edge reflectance, it may not be used for algorithm development to discriminate floating-matter types.

Indeed, with all these possible sources of uncertainty, such HICO-derived RFM can still be used for spectral discrimination of different floating matters without ambiguity, as shown below.

Spectral discrimination can be performed through either visual inspection or the use of a certain type of similarity index (e.g., SAM, Eq. 6). Here, results of the SAM analysis are presented in Table 1, followed by descriptions of visual inspection to interpret the spectral similarity or difference. Because nearly all floating algae show typical red-edge reflectance, discrimination of different algae types is focused on wavelengths <670 nm. To discriminate floating algae from non-living floating matters (e.g., marine debris), on the other hand, the inclusion of 670 nm is critical. Furthermore, because HICO data are noisy for wavelengths <450 nm, the SAM calculation was restricted to 450–670 nm for most RFM spectra of Figs. 2–4.

Table 1 shows the SAM results for three types of macroalgae (Sargassum, Ulva, kelp), three types of microalgae (Trichodesmium; Microcystis; red Noctiluca scintillans, or RNS), and one type of organic matter (brine shrimp cysts, or BSCs). Here, unless noted, Sargassum refers to Sargassum fluitans and S. natans (dominant pelagic type in the Atlantic Ocean) and Ulva refers to Ulva prolifera (dominant pelagic type in the Yellow Sea). For the same floating matter, if field-based RFM is available, then it is used as the reference; otherwise the mean HICO-derived RFM is used as the reference. For the SAM between different floating matters, all HICO-derived RFM values from both types are used (e.g., 4 Sargassum RFM values of Fig. 2a and 3 Ulva RFM values of Fig. 2b are used to calculate 12 SAM values), with their means and standard deviations listed in Table 1.

For each type of floating matter, HICO-derived RFM is very similar to either field-measured RFM or the floating matter's mean RFM, with SAM <4.6∘. In contrast, the SAM between different floating matters is always >9.9∘. These results suggest that, if these floating matters represent all that can be found in natural waters, they can be differentiated through spectroscopy analysis without any other ancillary information (e.g., knowledge of local oceanography or dominant floating algae type). This is despite the possible uncertainties in their reflectance magnitude, as discussed above. In the natural environments, however, there may be other types of floating algae whose spectral shapes may be similar to Sargassum fluitans and S. natans (e.g., either Sargassum horneri in the East China Sea or other brown algae) or Ulva prolifera (e.g., other green algae). Therefore, some form of ancillary information in addition to spectroscopy is still required in order to differentiate floating algae types.

The results from the SAM table can also be explained through visual inspection and interpretation of the spectral shapes, as discussed below.

From Fig. 2, it is clear that although the three types of macroalgae all share the same red-edge reflectance in the NIR, they have different spectral shapes in the visible wavelengths. Unlike the Ulva reflectance with a local peak around 560 nm, the spectral shapes of Sargassum reflectance resemble those of typical brown algae where the local reflectance trough around 625 nm is induced by chlorophyll c absorption and the low reflectance below ∼520 nm is due to carotenoid pigment absorption. These characteristics make it easy to distinguish Sargassum from Ulva (SAM>27∘, Table 1). On the other hand, it appears more difficult to spectrally discriminate Sargassum from kelp because they both have reflectance peaks around 600–645 nm and because they also share a common reflectance trough around 625 nm. However, considering Sargassum is moving in the ocean while kelp is fixed in location, they can be separated using sequential images. Even from a single image, when most visible wavelengths are used, Sargassum and kelp can still be spectrally discriminated (SAM>13∘, Table 1). Within the group of Sargassum spectra (Fig. 2a), there is some variability in the magnitude between 560–700 nm. It is unclear what caused such variability, although it could be due to changes in the carbon-to-chlorophyll ratio in Sargassum of different environment, as observed from kelp (Bell et al., 2015). Such a variability, however, would not impact the spectral discrimination of Sargassum from other floating matters, as the SAM between Sargassum spectra is <5∘, much lower than between Sargassum and any other types of floating matters (Table 1).

Similarly to the macroalgae, the microalgae scums also show elevated NIR reflectance (Fig. 3), and their spectral shapes in the visible wavelengths makes it straightforward to distinguish between kinds (SAM>17∘) and also straightforward to distinguish them from macroalgae (SAM>9.9∘). One exception may be the cyanobacterial scums (blue-green algae blooms) (Fig. 3b) as they show a reflectance peak around 550 nm, similarly to Ulva (Fig. 2b). However, reflectance around 550 nm is nearly symmetric for cyanobacterial scums but asymmetric for Ulva. There is also a local reflectance trough around 625 nm for cyanobacterial scums due to absorption of phycocyanin, but such a trough is lacking in the Ulva spectra. Such a characteristic makes it possible to differentiate between the two even without a priori knowledge of the ocean or lake environment, as the SAM between the two groups is ∼16.8∘ (Table 1). What is interesting is that within each class, either Trichodesmium or Microcystis, although the spectral shape is nearly identical from different spectra (SAM<5∘), there is substantial variability in the magnitude in the visible wavelengths, which might be due to changes in their carbon-to-chlorophyll ratios (Behrenfeld et al., 2005). Furthermore, the spectral-wiggling features between 450 and 660 nm in Fig. 3a are due to Trichodesmium-specific pigments such as phycourobilin, phycoerythrobilin, and phycocyanin that absorb light strongly at 495, 550, and 625 nm, respectively (Navarro Rodriguez, 1999). These features are unique to Trichodesmium scums, which makes it straightforward to develop classification algorithms once certain spectral bands are available to capture these features (e.g., Hu et al., 2010).

Of all the microalgae scums of Fig. 3, the spectral shapes of red Noctiluca (Fig. 3c) appear different from all others, but they show the same characteristics as those reported from the limited field measurements (Van Mol et al., 2007): a sharp, featureless increase from ∼520 to ∼600 nm. This unique spectral shape makes RNS different from all other floating matters (SAM>9.9∘, Table 1). The difference within this group is that the spectra from Sagami Bay off Japan show reflectance troughs around 670 nm. Because red Noctiluca is known to feed on other algae, it is speculated that the 670 nm trough is due to chlorophyll pigments of the consumed algae. Once more hyperspectral data are available in the future to test this hypothesis using field data, this characteristic may be used to study how red Noctiluca interacts with other algae. On the other hand, once more hyperspectral data are available in the future, it is also possible to test whether other algae (e.g., Mesodinium rubrum; Dierssen et al., 2015), once they have formed surface scum, have similar spectral shapes to those of red Noctiluca.

The non-algae floating matters in Fig. 4 show spectral characteristics different from both macroalgae and microalgae; for example they lack the typical red-edge reflectance of vegetation and lack typical spectral variations in the visible wavelengths due to pigment absorption. Within this group, the organic matter of BSCs (Fig. 4a) and emulsified oil (Fig. 4b) show some degrees of similarity as they also have monotonic reflectance increases from a wavelength between 500–560 to at least 740 nm. The difference between them is that BSC reflectance always starts to increase at ∼560 nm with an inflection wavelength of ∼640 nm, while reflectance of oil emulsions start to increase at variable wavelengths without any inflection between 560–740 nm. Indeed, the infection at ∼640 nm appears to be a common feature between BSC slicks and coral spawn slicks (Yamano et al., 2020). In contrast, depending on the oil emulsion state, oil emulsion may have different spectral characteristics (Lu et al., 2019), suggesting that there is no fixed “endmember” spectra for oil spills.

The inorganic “particles” (i.e., water bubbles, ice) also have distinctive spectral shapes. The examples in Fig. 4c indicate that submersed bubbles from ship wakes are similar in terms of spectral shapes, but all others are nearly identical in their lack of any narrow-band spectral features. Rather, foams, whitecaps, and ice all show flat reflectance spectral shapes between 400–800 nm that are consistent with in situ measurements of foams (Dierssen, 2019). The lack of narrow-band spectral features is similar to marine debris (Garaba and Dierssen, 2020). Such a similarity will make detection of marine debris very difficult, especially around ocean fronts because these are where surface materials tend to aggregate and foams also tend to form.

In addition to the spectra of Figs. 2–4 that can be well recognized, HICO also showed reflectance spectra that are difficult to discriminate from spectroscopy alone, as shown in Fig. 5. Without a known reflectance library, one can only speculate what algae type could be responsible for the algae scum spectra from some ancillary information in the literature. For example, the often-reported blooms of Lingulodinium polyedrum and Akashiwo sanguinea in coastal waters off southern California and in Monterey Bay, respectively, may show spectral shapes of those in Fig. 5a when they are heavily concentrated in surface waters. Inference may also be made for other cases once similar ancillary information is available. Even when such information is absent, one can still rule out some possibilities simply based on the spectral shapes. For example, the reflectance spectrum in Fig. 5b from Lake Victoria cannot be from cyanobacteria that has been often reported in this lake (Fig. 3b), but it is most likely from a dinoflagellate bloom, as blooms of other algae types have also been reported in this lake (Haande et al., 2011). Likewise, the different spectra from the same Taganrog Bay in Fig. 5c and d suggest different algae types. Clearly, although cyanobacterial blooms have been reported in many lakes, without spectral diagnosis one cannot simply jump to the conclusion that a freshwater bloom is caused by a certain type of cyanobacterium.

Because HICO is a pathfinder sensor that collected only a limited number of scenes, not all reported floating matters have been captured. For example, no HICO scene appears to have captured pumice rafts, Sargassum horneri, sea snot, or marine debris. Therefore, the spectral reflectance dataset presented here is incomplete. The use of data from other similar pathfinders, for example the DLR Earth Sensing Imaging Spectrometer (DESIS) on the ISS (235 bands from 400–1000 nm, 30 m resolution, 2018–present) and the PRecursore IperSpettrale della Missione Applicativa (PRISMA, 237 bands from 400–2505 nm, 30 m resolution, 2019–present), may complement the spectral data using the same approach presented here (e.g., sea snot reflectance spectra derived from DESIS; Hu et al., 2022). Even in their present form, given the large variety of floating matters presented here, the spectral data may lead to several implications for current and future satellite missions.

First, although all current multi-band sensors can detect floating matters through its elevated NIR reflectance (Qi et al., 2020), the Sentinel-3 Ocean and Land Colour Imager (OLCI) appears to be the best at differentiating spectral shapes in the visible wavelengths because of its 21 spectral bands between 400 and 1020 nm, especially because of its 620 nm band that can be used to differentiate whether an algae scum appears greenish or brownish, thus providing extra information to discriminate algae type in the absence of hyperspectral data.

Second, for the same reason, although only four bands (blue, green, red, NIR) are available on the PlanetScope (Dove) constellation, the recent SuperDove constellation is equipped with four additional bands with one centered at 610 nm and thus may significantly enhance the capacity of the current high-resolution sensors (∼3–4 or 30 m) to differentiate greenish and brownish algae types.

Finally, the Ocean Color Instrument (OCI) on NASA's PACE mission, to be launched in 2023, will be the first of its kind to map global oceans with hyperspectral capacity (5 nm resolution between 340–890 nm, plus seven discrete bands from 940 to 2260 nm) with a nominal resolution of 1 km. Unlike HICO, OCI will cover global oceans and lakes every 1–2 d, thus providing unprecedented opportunities to detect, differentiate, and quantify various types of floating matters. The spectral reflectance data, derived from one sensor (HICO) with a stable calibration, may serve as a consistent dataset to help select the optimal bands for future applications once PACE data become available, for example, through the use of an SAM matrix as demonstrated in Table 1. Likewise, the SBG mission currently being planned by NASA is expected to have hyperspectral capacity between 380 and 2500 nm with a nominal resolution of 30 m (Cawse-Nicholson et al., 2021); such a mission will provide unprecedented opportunity to map various floating matters on a global scale, and the hyperspectral dataset developed here can help develop algorithms before its launch.