Elsevier

Harmful Algae

Volume 103, March 2021, 102001
Harmful Algae

Original Article
To what extent can Ulva and Sargassum be detected and separated in satellite imagery?

https://doi.org/10.1016/j.hal.2021.102001Get rights and content

Highlights

  • Spectral noise quantified from Sentinel-3 and Sentinel-2 satellite images.

  • Simulations used to understand noise propagation in macroalgae detection and discrimination.

  • Detection and discrimination limits of sub-pixel algae coverage determined for Ulva and Sargassum.

  • Results serve as a reference to understand limitations and uncertainties in quantifying macroalgae.

Abstract

Blooms of floating macroalgae have been reported around the world, among which are recurrent blooms of Ulva prolifera and Sargassum horneri in the Yellow Sea and East China Sea. While satellite remote sensing has often been used to estimate their distributions and abundance as well as to trace their origins, because the algae mats are often much smaller than the size of an image pixel, it is unclear to what extent they can be detected and discriminated from each other in satellite imagery. Using data collected from laboratory experiments and by the Sentinel-3 OLCI (Ocean and Land Colour Instrument) and Sentinel-2 MSI (Multi Spectral Instrument) satellite instruments, we conduct simulated experiments to determine the lower detection limit and discrimination limit for these two macroalgae in different water environments and under different atmospheric conditions. For OLCI, the detection limit for both macroalgae is about 0.5% of a pixel, while the discrimination limit varies between 0.8% for clear water and 2% for turbid water. For MSI, the detection limit is about 2%, while the discrimination limit is about 6% for all water types. Below these two limits, detection and discrimination of macroalgae in these regions using the two sensors are subject to large uncertainties, thus requiring additional caution when interpreting algae areas and tracing algae origins.

Introduction

Various types of floating macroalgae have been detected in satellite imagery in the world's oceans (Smetacek and Zingone, 2013; Thornber et al., 2017; Bermejo et al., 2019; Qi et al., 2020). Of particular importance are the recurrent blooms of Ulva prolifera and Sargassum horneri in the western Yellow Sea (YS) (Liu et al., 2009; Hu et al., 2010; Qi et al., 2016) and East China Sea (ECS) (Qi et al., 2017; Xing et al., 2017), respectively, as they have posed various environmental and economic problems as well as ecological consequences. These macroalgae have been studied extensively using a variety of field, laboratory, and remote sensing techniques (Liu et al., 2010; Liu et al., 2018; Kong et al., 2018; Byeon et al., 2019; Min et al., 2017, 2019), as they may be harmful to the coastal environments (e.g., oxygen consumption, beach pollution, insect attraction, bad smells, etc.) although these algae do not contain toxins. While each technique has its own strengths and weaknesses, satellite remote sensing can provide synoptic and frequent as well as historical data to study their distributions and temporal changes in order better understand their potential driving mechanisms and consequences (e.g., Qi et al., 2016, 2017; Xing et al., 2017, 2019; Zhang et al., 2020).

However, two inherent limitations exist in satellite remote sensing, which may bring uncertainties in estimating the algae distributions and areal coverage as well as tracing algae originations.

The first is the detection limit, i.e., at what sub-pixel coverage can these small features in satellite imagery be detected, as the algae patches are often scattered to cover a small portion of an image pixel. For example, using image statistics, Hu et al. (2017) showed that >99.5% of Ulva-containing image pixels (250-m resolution) do not have full pixel coverage, and > 90% of these pixels have sub-pixel Ulva coverage of < 5%. The sub-pixel coverage, χ (0.0% - 100%), represents the amount of algae (in m2) within a pixel of finite size, where the amount can be from many scattered patches within a pixel. Using simulations, Hu et al. (2015) showed that, under typical conditions, the detection limit (i.e., the smallest χ) only depends on the sensor's signal-to-noise ratios (SNRs) of the red and near-infrared (NIR) bands that are used to quantify the pixel's red-edge reflectance (in practice, through an Floating Algae Index or FAI, Hu, 2009). For example, with SNRs of 200:1, Hu et al. (2015) estimated a detection limit of χdet= 1% when FAI was used to detect floating vegetation. This estimate was later confirmed by Wang and Hu (2016) using data collected by the Moderate Resolution Imaging Spectroradiometer (MODIS) 1-km data – with SNRs of 1000:1 the subpixel detection limit from image statistics was estimated to be χdet=0.2%. So, the detection limit is inversely proportional to SNRs. Once below the detection limit, the floating vegetation within the pixel is difficult or impossible to detect. However, these are for typical clear atmosphere and clear waters (e.g., for Sargassum in the Atlantic Ocean). For turbid atmosphere and/or turbid waters, the detection limit may be different.

The second is the discrimination limit, i.e., above what sub-pixel coverage can these small features be differentiated to be a certain type of floating vegetation as opposed to others. Without a priori knowledge of the local environment, the discrimination can only rely on the spectral characteristics of the pixel. Hu et al. (2015) used simulations to show that in order to detect the chlorophyll-c absorption feature around 625 nm from Sargassum-containing pixels, a subpixel fraction of χ=25% is required for a sensor with SNRs of 200:1. Clearly, discriminating floating algae type is more demanding than detecting floating algae presence. Furthermore, depending on the algorithms used to discriminate algae type, the discrimination limit may also be different. Indeed, the 25% requirement is based on the assumption that no other information is available on the local environment and all types of floating matters (macroalgae, microalgae, weathered oil, dead seagrasses, garbage patches, etc.) may coexist in the same region. In reality, such situations are rare, and the discrimination is often between two types of floating algae for certain regions, for example between Sargassum (either S. flutains or S. natans) and Trichodesmium for the tropical Atlantic Ocean, or between Ulva and Sargassum horneri in the YS and ECS (Xiao et al., 2020). In such cases, discrimination between two types of floating algae may be much easier, with relaxed requirements on subpixel coverage. However, to date, such requirements have not been quantified. Specifically, to what extent can Ulva and Sargassum horneri be separated in satellite imagery?

The objective of this study is to address the two questions on the detection and discrimination limits of Ulva and Sargassum horneri using multi-sensor measurements, noise propagation theory, and simulations. The ultimate goal is to provide a general guide on the capacity of satellite sensors in mapping Ulva and Sargassum, while at the same time demonstrate an approach to quantify such capacity for other floating matters in other regions.

Throughout this study, the subpixel coverage (and therefore detection limit) is expressed as a percentage (χ from 0.0% to 100%) of the pixel size rather than a physical area (m2), although they are equivalent for any satellite sensor with a finite pixel size or spatial resolution. Once χ is known, the corresponding physical area is simply a product of χ and pixel size. In presenting algae distribution maps with gridded data, the algae concentration in each grid is also expressed in χ, representing algae density within the grid (Qi et al., 2016, 2017; Wang and Hu, 2016), equivalent to m2 (algae) per km2 (area). This is regardless of whether the algae is in a single patch or multiple patches within the pixel. When the field or laboratory determined calibration constant is available to convert algae area (m2) to biomass, χ can also be easily converted to algae biomass density (kg algae per pixel or per grid), as shown in Hu et al. (2017) and Wang et al. (2018). Although the calibration constant is not applicable above a certain “saturation” point (Fig. 6 of Hu et al., 2017), because algae density is way below the “saturation” point in most waters and this is especially true regarding the detection limit, the “saturation” has no impact on this work. For this reason, χ is used in this study to represent subpixel algae coverage or areal density.

Section snippets

Laboratory and field data

Reflectance data of Ulva and Sargassum horneri from water tank experiments were taken from Hu et al. (2017) and Huang et al. (2018), respectively (Fig. 1a). They are used in this study to represent typical spectral endmembers of floating macroalgae with 100% subpixel coverage (i.e., χ = 100%).

Reflectance of three water types over the YS and ECS were taken from Lee et al. (2016). They range from clear water, moderately turbid water, to turbid water, respectively (Fig. 1b). In this study, they

SNRs and data product noise

A sensor's sensitivity is often defined as its SNRs at typical radiance input (Ltyp, mW cm−2 μm−1 sr−1), where SNR of a specific band isSNR=Ltyp/σ.

Here, sensor noise in the above equation, σ (mW cm−2 μm−1 sr−1), is defined as the standard deviation (i.e., a scalar quantity) of a univariate Gaussian distribution of random noise errors. This is similar to the definition of measurement uncertainties that is used to define the probability of a measured (or estimated) property in approaching the

Data product noise and uncertainty

Because σ(Rλ) depends on both water and atmosphere conditions, OLCI and MSI images collected over variable conditions were used to estimate σ(Rλ), specifically from relatively clear waters, moderately turbid waters, and highly turbid waters in the YS and ECS, and from clear and turbid atmospheres (aerosol optical thickness at 865 nm changed between ~0.06 and ~0.25). Fig. 4a and b shows some representative locations where σ(Rλ) was estimated. Their corresponding water-leaving reflectance (Rw = πR

Single band or band-combination index

On one hand, detection of floating algae presence is similar to detection of bright targets (e.g., oil/gas platforms (Liu et al., 2018) or ships (Heiselberg, 2016)) in the vast ocean, as they are all based on the concept of spatial anomalies. On the other hand, while the latter can be from single-band images (typically in the NIR) through the use of a roaming window and statistics-based threshold (Liu et al., 2018), the technique suffers from significant noise perturbations and lack of ability

Conclusions

Recurrent Ulva and Sargassum blooms in the Yellow Sea and East China Sea have stimulated research on their causes and consequences. While satellite remote sensing played a key role in monitoring and tracking the blooms, understanding their uncertainties requires accurate knowledge on their subpixel detection and discrimination limits. Using measurements from two popular satellite sensors, OLCI and MSI, and through simulated experiments, these limits are determined for most observing conditions.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (No. 41806208, No. 42076180) (Qi). We thank the European Space Agency, the U.S. NASA, and the U.S. Geological Survey for providing all satellite data. We thank the two anonymous reviewers for their thorough comments to help improve the presentation of this work.

References (43)

  • C.S. Thornber et al.

    Spatial and temporal variability in macroalgal blooms in a eutrophied coastal estuary

    Harmful Algae

    (2017)
  • Q. Vanhellemont et al.

    Advantages of high quality SWIR bands for ocean colour processing: examples from Landsat-8

    Remote Sens. Environ.

    (2015)
  • Y. Xiao et al.

    Remote sensing estimation of the biomass of floating Ulva prolifera and analysis of the main factors driving the interannual variability of the biomass in the Yellow Sea

    Mar. Poll. Bull.

    (2019)
  • J. Xiao et al.

    An anomalous bi-macroalgal bloom caused by Ulva and Sargassum seaweeds during spring to summer of 2017 in the western Yellow Sea, China

    Harmful Alage

    (2020)
  • Q. Xing et al.

    Monitoring seaweed aquaculture in the Yellow Sea with multiple sensors for managing the disaster of macroalgal blooms

    Remote Sens. Environ.

    (2019)
  • H. Zhang et al.

    Characterization of the development stages and roles of nutrients and other environmental factors in green tides in the Southern Yellow Sea, China

    Harmful Algae

    (2020)
  • D. Blondeau-Patissier et al.

    Phenology of Trichodesmium spp. blooms in the Great Barrier Reef lagoon, Australia, from the ESA-MERIS 10-year mission

    PLoS ONE

    (2018)
  • S.Y. Byeon et al.

    The origin and population genetic structure of the ‘golden tide’ seaweeds, Sargassum horneri, in Korean waters

    Sci. Rep.

    (2019)
  • Y. Chen et al.

    Spatial-temporal distribution of golden tide based on high-resolution satellite remote sensing in the South Yellow Sea

    J. Coast. Res.

    (2019)
  • H.R. Gordon

    Atmospheric correction of ocean color imagery in the earth observing system era

    J. Geophys. Res.

    (1997)
  • J. Gower et al.

    Ocean color satellites show extensive lines of floating Sargassum in the Gulf of Mexico

    IEEE Trans. Geosci. Remote Sens.

    (2006)
  • Cited by (39)

    • Mapping Ulva prolifera green tides from space: A revisit on algorithm design and data products

      2023, International Journal of Applied Earth Observation and Geoinformation
    View all citing articles on Scopus
    View full text