11 May 2017 Satellite retrievals of Karenia brevis harmful algal blooms in the West Florida shelf using neural networks and impacts of temporal variabilities
Author Affiliations +
J. of Applied Remote Sensing, 11(3), 032408 (2017). doi:10.1117/1.JRS.11.032408
We apply a neural network (NN) technique to detect/track Karenia brevis harmful algal blooms (KB HABs) plaguing West Florida shelf (WFS) coasts from Visible-Infrared Imaging Radiometer Suite (VIIRS) satellite observations. Previously KB HABs detection primarily relied on the Moderate Resolution Imaging Spectroradiometer Aqua (MODIS-A) satellite, depending on its remote sensing reflectance signal at the 678-nm chlorophyll fluorescence band (Rrs678) needed for normalized fluorescence height and related red band difference retrieval algorithms. VIIRS, MODIS-A’s successor, does not have a 678-nm channel. Instead, our NN uses Rrs at 486-, 551-, and 671-nm VIIRS channels to retrieve phytoplankton absorption at 443 nm ( a ph 443 ). The retrieved a ph 443 images are next filtered by applying limits, defined by (i) low Rrs551-nm backscatter and (ii) a minimum a ph 443 value associated with KB HABs. The filtered residual images are then converted to show chlorophyll-a concentrations [Chla] and KB cell counts. VIIRS retrievals using our NN and five other retrieval algorithms were compared and evaluated against numerous in situ measurements made over the four-year 2012 to 2016 period, for which VIIRS data are available. These comparisons confirm the viability and higher retrieval accuracies of the NN technique, when combined with the filtering constraints, for effective detection of KB HABs. Analysis of these results as well as sequential satellite observations and recent field measurements underline the importance of short-term temporal variabilities on retrieval accuracies.
El-Habashi, Duran, Lovko, Tomlinson, Stumpf, and Ahmed: Satellite retrievals of Karenia brevis harmful algal blooms in the West Florida shelf using neural networks and impacts of temporal variabilities



We have previously described preliminary results with a neural network (NN) approach for the detection1 and tracking of Karenia brevis harmful algal blooms (KB HABs) that frequently plague the coasts and beaches of the West Florida shelf (WFS) using visible-infrared imaging radiometer suite (VIIRS) satellite data. Such a monitoring capability for KB HABs is important because of their negative impacts on ecology and health. More specifically, high KB HABs levels pose a threat to fisheries and human health and directly affect tourism and local economies.2 Effective KB HABs detection and tracking approaches are needed for use with VIIRS so that NOAA can extend its HABs monitoring capabilities. These previously relied on MODIS-A imagery34. and specifically on the remote sensing reflectance signal at 678 nm, (Rrs678) at the chlorophyll fluorescence wavelength. This was used in MODIS-A with the normalized fluorescence height (nFLH) and related red band difference (RBD) techniques to effectively help in KB HABs retrievals. However, the current VIIRS satellite, unlike its predecessor MODIS-A, does not have a 678-nm channel to detect chlorophyll fluorescence. To overcome the lack of a fluorescence channel on VIIRS, the NN approach bypasses the need for measurements of chlorophyll fluorescence, allowing us to extend KB HABs satellite monitoring capabilities in the WFS to VIIRS.

The essence of the approach is the application of a standard multiband NN inversion algorithm, previously developed and reported by us.1314.15.16 This approach takes VIIRS Rrs measurements at the 486-, 551-, and 671-nm bands (or 488, 555, and 667 nm for MODIS-A) as NN inputs and produces the related inherent optical properties (IOPs) at 443 nm as outputs: namely, the absorption coefficients of phytoplankton (aph443), dissolved organic matter (ag), and nonalgal particles (adm) as well as the particulate backscatter coefficient, (bbp). In this work, it is only the NN output of aph443 with which we are concerned. This is used to generate an aph443 image, which is then converted into an equivalent [Chla] image, using empirical relationships for specific chlorophyll absorption values in the WFS, which have been determined from in situ measurements.17 Next, to obtain KB values from the VIIRS NN retrieved aph443 image, we apply two filter processes, based on constraints known to be associated with KB HABS in the WFS. These constraints are: (i) low backscatter at 551 nm, manifested as a maximum permissible value of Rrs551Rrs551max and (ii) a minimum permissible [Chla]min threshold value9,18,19 and hence an equivalent minimum permissible value: aph443aph443min. Following application of these two filter processes, the residual image will now show only aph443 values that are compatible with both criteria for KB HABs and are, therefore, representative of KB HABs.

VIIRS retrievals of KB in the WFS, using our NN and five other retrieval algorithms, are then compared and evaluated for their efficacy against datasets of in situ measurements. To have enough data for meaningful analysis, in situ data sets that cover the available data from the start of the VIIRS mission in January 2012 to October 2016 were created for these comparisons. The comparisons confirmed the viability and potential of the NN technique, when combined with the filtering constraints devised, for effective detection of KB HABs in the WFS. Analysis of results as well as sequential satellite observations and field measurements in the WFS also show the importance of short-term temporal variabilities and underline their impact on retrieval accuracies.


Materials and Methods Background


Neural Network Algorithm Background

For the development of our NN algorithm,1,13,1415.16 a synthetic data set of 20,000 IOPs was produced based on the NASA bio-optical marine algorithm data set.20 These IOPs whose range and variability is well represented in the literature2122. were then used as inputs to a four component bio-optical model,15,24,34 which, in conjunction with a HydroLight-based,35 parameterized forward model, described in Ref. 32, Lee produced 20,000 sets of Rrs values at 486, 551, and 671 nm (for VIIRS) and 488, 555, and 667 nm for MODIS-A. The NN was trained on 10,000 of these values and tested on a 10,000 synthetic subset as well as on field data to solve the inverse problem36,37 of retrieving physical variables, including aph443, from Rrs values at 486, 551, and 671 nm, and at 488, 555, and 667 nm. The algorithm is a standard multiband NN inversion algorithm that takes VIIRS Rrs measurements at the 486-, 551-, and 671-nm bands and MODIS-A measurements at 488-, 555-, and 667-nm bands in the WFS as inputs, and produces as outputs the related IOPs, namely: aph443, ag, and adm as well as bbp, all at 443 nm. (As mentioned in the Sec. 1, it is only the aph443 output of NN that we are concerned with here). Detailed descriptions of the NN are given in Refs. 1314.15.16, and Ref. 1 shows the necessary parameters for implementation of the NN with a MATLAB tool for obtaining satellite retrievals.


Development of Criteria for Filters to be Used in Processing VIIRS (and MODIS-A) Retrievals of Rrs551 and [Chla] in the WFS to Obtain Residual Images Depicting KB HABs Concentrations

A brief description is given below of the approach used for devising filtering criteria to apply to VIIRS (and MODIS-A) retrieved images of Rrs551 and aph443 (or [Chla]) in the WFS, to evolve residual images depicting KB HABs concentrations. This approach, which is described in detail in Ref. 1, is summarized and updated here. Satellite data are obtained from VIIRS and MODIS-A-sensors as level two data products from NASA’s processing stream, available from Ref. 38. Pixels are excluded from processing if they have been flagged for land, cloud, failure in atmospheric correction, stray light, bad navigation quality, both high and moderate glint, negative Rayleigh corrected radiance, and solar zenith larger than 70.


VIIRS retrievals of Rrs551 and aph443 and determination of limiting values of Rrs551max and aph443 min compatible with KB HABs in the WFS

This section updates and revises an earlier analysis carried out in Ref. 1 to determine limiting values of Rrs551max and aph443min compatible with KB HABs in the WFS. By searching more thoroughly through available data bases, we are able to constrain comparison of VIIRS satellite retrievals on 10/9/12 to in situ measurements carried out all on the same day. This constraint provides a much more convincing comparison of satellite retrievals to in situ values from which to select our limiting values than was done in Ref. 1. In the latter, available retrievals from VIIRS on one clear day (9/2/14) were compared with in situ measurements, not carried out on the same day, but rather over an overlapping 22-day period (8/27/2014 to 9/17/2014) during which conditions could clearly have changed much more substantially.

Figure 1 shows in situ measured KB cell counts associated with a prominent KB HAB bloom in the WFS with its peak occurring on 10/09/2012, from the NOAA HABSOS.39 This shows the bloom area with indicators (circles of different sizes) for measured KB HABs cell counts for that date. We next compare NN retrievals from VIIRS observations to these cell counts. Using Rrs486-, 551-, and 671-nm measurements from VIIRS as inputs to the NN, we retrieve aph443 for the same date and area shown in Fig. 1. These retrievals are shown in Fig. 2, which also shows on the right-hand scale the equivalent [Chla] converted from aph443 using the empirical relationship of Ref. 17


aph443=0.051×[Chla]0.74,  m1.

Fig. 1

NOAA HABSOS data39 with in situ KB concentrations for October 9, 2012.


Fig. 2

NN retrieved aph443 (LHS) equivalent [Chla] (RHS).


The challenge then remains to retrieve from Fig. 2 only the aph443 or equivalent [Chla] that represents KB HABs by filtering out pixel values incompatible with KB HABs. To do this, it is necessary to first define and apply the filters, which, as discussed in Sec. 1 and in more detail in Ref. 1, serve to eliminate pixels that have Rrs551 backscatter and aph443 (or [Chla]) values that are incompatible with KB HABs. The two filter processes are summarized next.

Considerations of backscatter values coexisting with KB Refs. 12, 18, and 19 show maximum permissible values of backscatter, compatible with KB HABs, as bbp5500.0045  m2  mg1 at 550 nm. Then, if Rrs551 is taken to serve as a proxy for backscatter, the equivalent max permissible value is Rrs5517.0×103  sr1. However, by inspection of numerous VIIRS retrievals of Rrs551 against more than 100 simultaneous or near simultaneous in situ measurements of KB HABs occurrences in the WFS over the 2012 to 2016 period (discussed in Sec. 3), we concluded that a value of approximately Rrs5516.0×103  sr1 appears more appropriate as the highest permissible Rrs551 limiting value compatible with the existence of KB HABs. So that would define the limits of the first processing filter, hereinafter denoted as F1, which when applied to VIIRS retrievals of Rrs551, would be used to eliminate pixels with higher values, as being incompatible with the existence of KB HAB blooms. The application of this filter process F1 is described in the next section.


Backscatter limiting values compatible with KB HABS and filter process F1

Figure 3(a) shows VIIRS retrievals of Rrs551 for the same KB HABs peak date 10/9/2012 and location. When filter process F1 is applied to Fig. 3(a), all pixels with Rrs5516.0×103sr1 are screened out, leaving the residual pixels in Fig. 3(b) as a mask retaining only pixels compatible with KB HABs as far as backscatter is concerned.

Fig. 3

(a) VIIRS Rrs551 image; (b) residual Rrs551 image after F1 mask (light gray) is applied showing residual (Rrs5510.006  sr1). Note that these images are overlaid with NOAA-HABSOS KB cell counts, from Fig. 1. (c) Residual image after filter F1 is applied to Fig. 2 retrievals. White areas represent cloud cover or invalid data.


We next apply this pixel mask, from Fig. 3(b), to the VIIRS NN retrieved aph443 for the same October 9, 2012, date and location, and shown in Fig. 2. Eliminating pixels outside the mask area, results in eliminating all aph443 values in Fig. 2 that do not satisfy the Rrs5516.0×103  sr1 requirements for compatibility with KB HABs. Residual aph443 values, resulting from this F1 filter process, and, shown in Fig. 3(c) now satisfy the Rrs551 limit requirements for compatibility with KB HABs.


Filter process F2 to apply limits to NN retrieved aph443 (or equivalent [Chla]) images to obtain residual images with values compatible with and depicting KB HABs concentrations

We next consider the appropriate F2 filter values for retaining only aph443 values that were in the F1 filtered image, Fig. 3(c), that are compatible with KB HABs. Refs. 18 and 19 show the minimum [Chla] concentration compatible with KB HABs in the WFS as [Chla]1.5  μg·L1. This is converted into equivalent aph443 using the empirical relationship17which gives aph4430.0688  m3 as the minimum permissible value. However, by inspection of VIIRS NN retrieved aph443 values against more than 100 in situ measurements of KB HABs cell values in the WFS over the 2012 to 2016 period (see discussion in Sec. 3) we find that a value of approximately aph4430.061  m3 or equivalent [Chla] is more appropriate as the limiting values of aph443 and [Chla] for compatibility with KB HABs.

Applying this value as filter F2 to Fig. 3(c), we eliminate pixels that are not compatible with KB and end up with the residual Fig. 4(a). This shows only residual values with aph4430.061  m3 which at the same time satisfy the backscatter F1 filter requirement Rrs551<6.0×103  sr1. These are, therefore, residual values satisfying both F1 and F2 filter requirements and are, therefore, compatible with and representative of KB HABs. Comparison of the residual aph443 (left-hand scale) and equivalent [Chla] (right-hand scale) in Fig. 4(a), which represents the retrieved KB HABs values, with the overlaid cell count information from Fig. 1, shows good qualitative agreement. Figure 4(b) shows the equivalent NN aph443 from MODIS-A for the same date and location, and also filtered for the same Rrs551 backscatter and aph443 compatibility requirements (but using the equivalent 488-, 555-, and 667-nm bands of MODIS-A as inputs to the NN instead of the VIIRS bands). Again, we see good qualitative agreement with the in situ values of Fig. 1.

Fig. 4

(a) VIIRS retrieved NN aph443 (left-hand scale) and equiv. [Chla] (right-hand scale) after filter process masks F1 and F2 are applied. These residual values are, therefore, compatible with and show the extent of the KB blooms; (b) MODIS-A NN retrieved aph443 (left-hand scale) and equiv. [Chla] (right-hand scale) after filter process masks F1 and F2 are applied. Residual values are, therefore, compatible with and indicate KB blooms. Light gray represents F1 and F2 masks and white represents cloud cover or invalid data.



Rrs spectra associated with KB HAB blooms

We next examine the relationship between VIIRS Rrs multiwavelength spectra and the different concentrations of in situ KB HABs reflected in the wide variety of [Chla] values of Fig. 4(a) for the bloom of 10/09/2012. Figure 5(a) shows Rrs reflectance spectra observed for low bloom concentration locations. It is color-coded to reflect the measured in situ concentrations at these locations. Figure 5(b) shows the Rrs spectra for medium and high bloom locations. As can be seen, both the magnitude and shape of these reflectance spectra change with varying underlying in situ bloom concentrations. Thus, with increasing cell concentration, the increased absorption is seen to dominate over the low backscatter associated with KB cells, so that magnitudes of Rrs decrease as bloom concentrations increase. Furthermore, for higher bloom concentration, this combination of absorption and low backscatter features results in a decrease in the blue part of the spectrum with respect to the green. The features, shown below, retrieved from satellite observations, have also been previously observed with in situ measurements of KB HABS.18

Fig. 5

Matchup of VIIRS Rrs reflectance spectra on 10/09/2012 for: (a) low in situ cell count locations. (b) Medium cell to high counts locations; (c) very low bloom Rrs (orange), high bloom Rrs (green) and high bloom Rrs normalized to low bloom Rrs value at 410 nm (dashed green). (d) Differences between low and normalized high bloom.


The change in spectral shape is brought out in Fig. 5(c), in which the Rrs spectrum for a high bloom concentration (green) is normalized to the reflectance spectra (orange) at 410 nm, of very low KB (i.e., low [Chla] shown dotted in red). The difference between the two spectra (high and low bloom or [Chla]) is shown in Fig. 5(d), whose shape exhibiting with the decrease in green reflectance relative to the blue and red helps provide confirmation that we are dealing with KB HABs of increasing intensity.


Evaluation of VIIRS KB HABs Retrievals, Obtained Using NN and Other Algorithms, against In Situ Measurement for Matchups Occurring over 2012 to 2016 Period

We next examine VIIRS NN retrievals of KB HABs in the WFS, using the limiting F1 and F2 filter approach described in Sec. 2, against coincident, or near coincident, in situ measurements. We also compare these VIIRS NN retrievals and in situ matchups against retrievals and matchups obtained using other direct retrieval algorithms. In doing so, it is not the intention to present an exhaustive comparison of the many approaches that have been studied for the detection of KB HABs in the WFS. Some of these are listed in Refs. For more details, the reader is referred to recent excellent surveys and comparisons in Refs. Rather, we focus on techniques that have the potential to retrieve images from VIIRS observations that will show concentration and distribution of KB HABs at the time of observation. Below, we briefly summarize the main retrieval techniques that might be used for KB HABs retrievals.


Existing and Potential KB HABs Algorithms to be Considered for VIIRS Retrievals

Existing or potential satellite algorithms that are relevant to the detection of KB HABs in the WFS rely on retrieved chlorophyll-a concentrations [Chla] as part of their processing. They fall primarily into three categories: (i) using the remote sensing reflectance fluorescence signal at Rrs678 measured by the MODIS-A satellite; (ii) using blue-green ratio algorithms, obtained from the NASA OC3/OCI20,4041.42 and include the chlorophyll-a anomaly technique and related approaches;4344. and (iii) [Chla] retrievals that are obtained from NASA products, including generalized inherent optical property (GIOP) and quasi-analytical algorithm (QAA) (see below for definitions), and a more recent WFS region specific empirical algorithm, the red green chlorophyll-a index (RGCI) which retrieves [Chla] using two visible VIIRS or MODIS-A bands.54

Those methods using the 678-nm fluorescence signal include: (a) the nFLH, where FLH is a measure of the solar stimulated chlorophyll-a fluorescence, obtained from water leaving radiance Lw (λ). Normalized nFLH34.,5556.57 is obtained from water leaving radiances normalized to the down-welling light at the sea surface; nFLH is in turn computed as the difference between the observed normalized water leaving radiance [nLw(678)] and a linearly interpolated nLw(678) from the two surrounding bands; (b) the RBD techniques; and (c) the KB bloom index (KBBI).4,5 While these techniques, or a combination of them, have generally exhibited good retrieval statistics,9,12 they are, unfortunately, not applicable to VIIRS, the successor satellite, which unlike its predecessor, MODIS-A does not have a 678-nm fluorescence channel. Since, in this paper, we are concerned only with retrieval techniques applicable to VIIRS, fluorescence techniques are not included in any further discussions.

In the following sections, we focus on comparing VIIRS [Chla] retrievals against coincident or near coincident in situ measurements of KB cell counts and their equivalent [Chla]. In prior work,1 we analyzed satellite retrievals against in situ measurements, in which NN VIIRS retrievals were compared against retrievals using OCI/OC3 (Refs. 42 and 43) and RGCI.54 In this paper, we have now added comparisons of NN retrievals with those using the GIOP5859.60 and QAA61,62 widely used algorithms. This remedies an important omission in our previous publication1 and extends NN retrieval comparisons to include comparisons with all available VIIRS direct retrieval techniques. Thus, algorithms compared with NN now include GIOP and QAA, as well as OCI/OC3, and RGCI for new overlap time windows (see Secs. 3.2 and 3.3). The salient features of these different NASA retrieval algorithms are summarized below.

The OCI/OC3 indexes, both NASA products,43 use [Chla] retrieval algorithms that make use of the ratio of blue/green bands in the MODIS-A and VIIRS satellites20,4041.42.43 and were found to yield exactly the same retrievals for our WFS conditions.

GIOP model. While there are numerous semianalytical algorithms (SAAs) existing to estimate IOPs, GIOP allows construction, evaluation, and selection of specific modeling assumptions from different SAAs at runtime, in order to generate a unified IOP model. This NASA algorithm5859.60 returns spectral marine absorption and backscattering coefficients for water column constituents [e.g., colored dissolved organic material and algal and nonalgal particles] in m1, calculated using the default global configuration of the GIOP model. We use it here to retrieve [Chla] from VIIRS observations for matchups with in situ measurements.

Quasi-analytical algorithm version 5 (QAA_5)61,62 developed by Lee et al. to derive the absorption and backscattering coefficients by analytically inverting the spectral remote-sensing reflectance [Rrs(λ)]. It starts with an empirical estimate of the total absorption coefficient at a reference wavelength (550, 555, or 560) and then analytically calculates the backscattering coefficient at the same wavelength. The amplitude of these coefficients at other wavelengths is obtained using an empirical estimate of the particulate backscattering spectral shape and the measured remote-sensing reflectance. After the total absorption coefficient is known, it can be further decomposed into the algal and nonalgal components. We use it here to retrieve [Chla] from VIIRS observations for matchups with in situ measurements.

RGCI is an empirical WFS region specific retrieval algorithm54 that retrieves [Chla] from observations of MODIS-A and VIIRS bands. For VIIRS, the [Chla] retrievals are given by


[Chla]RGCI=0.1×exp(11.8×Rrs671Rrs551),  μgL1.

The above algorithms, all applicable with VIIRS, will be used for retrieval comparisons against in situ measurements and comparisons against VIIRS NN.


Available Matchups for VIIRS Observations of the WFS Over the 2012 to 2016 Period

The ultimate test for the viability of KB HABs satellite retrieval techniques is their ability to match retrieved values with concurrent in situ measurements. However, it is difficult on any one day to find sufficient matchups between satellite observations and concurrent, or near concurrent, in situ measurements to obtain statistically meaningful results. We, therefore, extended the study period to look at all available WFS matchups, Fig. 6, between VIIRS measurements and in situ data at concurrent dates and over the 2012 to 2016 period for which there was available VIIRS data.

Fig. 6

Showing in situ locations of Karenia brevis cell counts cover the range 0.01 to 9.2  × 106  cells·L1 obtained from Florida Fish and Wildlife Conservation Commission (FWC). Zoomed area illustrates the extent of the underline details of KB values available in VIIRS retrievals for the period of 2012 to 2016.


We then looked for matchups where the overlap time windows between satellite observations and in situ measurements were 15 and 100 min.

It was found that there were 36 matchups of available in situ measurements that satisfy the matchup conditions for satellite observations within a 15-min overlap time window, including those pixels that show KB cell concentrations at values below those considered to be officially labeled as blooms. The additional conditions stipulated for matchup were that pixel centers were: 0.3 miles or less from the in situ measurement location. This represents an empirical approach to ensure that pixel values could be reasonably assumed to reflect the related in situ measurements and hence reduce potential impact of patchiness63 within the pixel (0.7  km2 for VIIRS and 1  km2 for MODIS). Pixels were also excluded from the matchup comparison if they had been flagged for land, cloud, failure in atmospheric correction, stray light, bad navigation quality, both high and moderate glint, negative Rayleigh corrected radiance, and solar zenith larger than 70 deg, and any pixels that had water leaving radiance spectra with negative values in any one of its wavelengths. Cell count sample measurements also had to be at less than 1 m depth and at concentrations 104  cells·L1. It should be noted47 that [Chla] of 1  μg·L1 is taken as 105  cells·L1. The in situ cell count data were obtained from the Florida Fish and Wildlife Conservation Commission’s Fish and Wildlife Research Institute (FWC-FWRI). The search for matchup between VIIRS satellite and in situ observations on the same day and within a 100-min window of the overpass time showed 93 cases that satisfied the matchup conditions as specified above. The seasonal distribution of these 93 matchups between satellite observations and in situ measurements used in the comparison of retrieval accuracies is shown in Table 1.

Table 1

Seasonal distribution of matchups used in the comparison of retrieval accuracies.

January 2012 to March 2016WinterSpringSummerFallTotal
Number of matchups18616893


Comparisons of VIIRS Retrievals Using NN and Other Algorithms Against KB HABs In Situ Measurements in the WFS Over the 2012 to 2016 Period for 100- and 15-min Windows between the Overpass Time and In Situ Observations

We next revise the comparisons of NN retrieval accuracies from comparisons with OCI/OC3 and RGCI in Ref. 1 to (i) reflect different matchup overlap time windows with in situ measurements and (ii) also extend them to now include comparisons with GIOP and QAA. The change in time windows is also a significant change from Ref. 1. Thus, the matchup overlap time windows between retrievals and in situ measurements for the statistical comparison of NN retrievals with other techniques were changed to 100 and 15 min in this paper, from the previous time windows of: daily, 60, and 30 min in Ref. 1. The daily matchup was eliminated as being too inaccurate.

The 100-min time window conforms to the approximate time between consecutive overpasses of the VIIRS satellite and consecutively retrieved images, which can also provide evidence of the temporal variations observed with KB HAB blooms. These are the subject of Sec. 3.5. The shorter 15-min time window was selected after an investigation for this paper found 15 min to be the shortest time window, which at the same time provided enough matchup points for meaningful statistical analysis. This approach is borne out by the results that show much better matchups for the 15-min window. This has important implications regarding the validity of satellite observations of KB HABs.

The relationship between retrieved [Chla] using the NN, OCI/OC3, and RGCI algorithms and the in situ KB cell count measurements for the 100- and 15-min matchup time windows are shown in Fig. 7. For NN, OCI/OC3, and RGCI retrievals, it was found that there were 93 valid observations within the 100-min window between overpass time and the in situ measurements. In these results, R2 is the coefficient of determination. To determine R2, the orthogonal linear regression approach (OR) was used where errors are assumed to exist for both variables. The error (ϵ) is calculated as the sum of orthogonal distances. OR estimates of X on Y will minimize the orthogonal distance from the observed data points to the regression line64


where β0   and β1 are intercept and slope and the OR estimate of the slope is


where SXX and SYY are the covariance for X and Y, respectively, and SXY is the correlation for X and Y.

Fig. 7

In situ comparison for the available matchups points for 100- and 15-min overlap windows with in situ observations obtained using the different algorithms (NN, OCI/OC3, and RGCI). Color coding of the dots denotes distance to shore, with red being the closest.


Figure 7(b) shows that the results for 23 matchups observed when the observations window between overpass time and in situ measurements are restricted to 15 min, in which the impact of temporal variations is reduced. As can be seen, correlations and errors greatly improve for the 15-min window, compared to the 100-min window observations. It is also observed that NN retrievals exhibit the best performance for both time windows, in terms of both correlations and errors.

When VIIRS retrieval comparisons were extended to using GIOP and QAA algorithms, it was found that both algorithms exhibited negative values or no retrievals in many instances. When these negative values were excluded, there remained 68 valid matchups for the 100-min overpass window. The retrieved [Chla] for these 68 matchups is shown against in situ cell counts in Fig. 8, for all retrieval techniques.

Fig. 8

Results for 68 in situ observations within 100 min of VIIRS overpass for the six algorithms showing retrieved [Chla] against KB HABs cell counts for NN, GIOP, OC3, OCI, RGCI, and QAA retrievals. Color coding of the dots denotes distance to shore, with red being the closest.


When the overlap observation window is reduced to 15 min, we have 18 valid matchups remaining, after excluding negative values associated with GIOP and QAA. Results for these 18 matchups are shown in Fig. 9. Again, as can be seen, correlations and errors greatly improve for the 15-min window over those for the 100-min window, Fig. 8.

Fig. 9

Results for 18 in situ observations within 100 min of VIIRS overpass for the six algorithms showing retrieved [Chla] against KB HABs cell counts for NN, GIOP, OC3, OCI, RGCI, and QAA retrievals. Color coding of the dots denotes distance to shore, with red being the closest.


The results in Figs. 8 and 9, which illustrate the impact of observation time windows on retrieval accuracies for all algorithms, are summarized in Tables 2 and 3. These results support the conclusion that, at least for these preliminary and somewhat limited data sets, the NN retrievals exhibit the best performances against the in situ measurements, for both the longer (100 min) and, more importantly, the shorter (15 min) overlap time windows. This was observed both in terms of higher correlations and lower errors against the in situ measurements.

Table 2

Statistics of comparison for Fig. 7.

y-axis [Chla] (μg L−1)x-axis KB cell counts (cells L−1)R2ϵSlope and interceptN
NNVIIRS 100-min window0.386.79y=0.46x−1.7193
NNVIIRS 15-min window0.760.59y=0.48x− 1.85 23

Table 3

Statistics of comparison for Figs. 8 and 9.

y-axis [Chla] (μg L−1)x-axis KB cell counts (cells L−1)R2ϵSlope and interceptN
NNVIIRS 100-min window0.425.01y=0.52x−2.0468
NNVIIRS 15-min window0.790.42y=0.47x−1.8318


Assessing Validity of Filter Limits on Rrs551 and aph443 ≥ 0.061 m−3 (or [Chla]) in Context of Comparisons Against In Situ Measurements

We also use the above matchups to assess the validity of the F1 and F2 filter limits (discussed in Sec. 3.1) for Rrs5516.0×103  sr1 and aph4430.061  m3 or [Chla]. Figure 10 shows the NN retrieved aph443 matchups falling within the 15-min window, which were retrieved (without any filter constraints being applied for Rrs551 and [Chla]). These show that there are three false negatives (marked with dark-gray arrows) and two false positives (marked with light-gray arrows) erroneously included within the Rrs551 and [Chla] limiting values, out of 36 matchup points. It should be noted that the spatial variability within a pixel on the same day can be quite large and can include both extremely low and high KB cell concentrations.1 This means that subpixel variability in the bloom concentration may be a key factor in creating apparent “false” positives. Planned future statistical analysis of false positives and negatives in retrieval results against in situ measurements is expected to result in further refinements of these limiting values and improvements in retrieval accuracies.

Fig. 10

The matchups falling within the 15-min window, with both filter thresholds.



Consecutive Satellite Images to Examine Temporal Changes

From the results in Sec. 3.3, it was seen that reducing the time window between satellite and in situ observations can generally significantly increase the accuracy between VIIRS retrieved [Chla] and in situ measured KB cell counts. These changes can be quite rapid.65,66 To explore the potential for detecting HAB bloom changes over relatively short periods from overlapping consecutive satellite overpasses, we have also examined changes in three consecutive overlapping satellite images in Fig. 11. Two of these are from VIIRS, 96 min apart and an intermediate one is from MODIS-A, 70 min after the first VIIRS image.

Fig. 11

The MODIS-A granule and two VIIRS granule within 100 min on November 3, 2014.


We next examine NASA OC3 retrievals of [Chla] for these three consecutive granules. Figs. 12(a)12(c) show the retrieved NASA OC3 [Chla] products from these consecutive VIIRS-MODIS-VIIRS images for the WFS near Sarasota, Florida, on 11/3/2014.

Fig. 12

(a)–(c) Changes in bloom for consecutive satellite images of region one.


Since environmental factors, including wind direction and currents, are known to affect temporal changes, including down- and upwelling and transportation of KB blooms in the WFS,47,65,6768.69 we include wind information available for that date. Figure 13 shows data obtained from the National Data Buoy Center website for the C-MAN stations at Venice, Florida (Station VENF1). (27°4’21” N 82°27’10” W) showing the variability of the wind on 11/3/14 at the time of the VIIRS overpasses.

Fig. 13

Temporal changes in wind directions during November 3, 2014, obtained from VENF1 station (27°4’21” N 82°27’10” W). Zoomed image shows wind directions for the time period of the consecutive satellite images. The wind vectors were plotted in the direction toward which the wind is flowing.


It should be noted that in shallow waters, the angle between the wind-induced surface water movement and the wind direction can be as low as 15 deg (or even less depending on the wind velocity and water depth) rather than up to the 45-deg predicted by idealized spiral Ekman models for deeper waters.70 We can probably assume that wind-induced current direction will be about 15-deg with the wind direction in the relatively shallow waters under consideration. This does not include tidal or other sources of current for which we have no information on that date. For qualitative comparison purposes, we show approximate wind directions and possible wind induced current directions (at 15 deg), overlaid in Fig. 12.

We now examine in more detail region 1, in Fig. 12, and shown in the zoomed images in Fig. 14. The bloom, as delineated by the [Chla] color contour in the images, appears, qualitatively, to increase in concentration and expand in the southwest direction over the 96-min interval between the consecutive overlapping VIIRS-MODIS-VIIRS images. These changes are reflected in the associated zoomed pixel images on the right-hand side of Fig. 14 and appear to provide qualitative visual indications of expansion of the bloom and its increasing [Chla] concentration in the southwest direction. They also appear broadly consistent with the directions of wind and likely currents, though there is no specific or quantitative evidence of linkage. Furthermore, no movements of [Chla] distribution patterns are discerned that would indicate transport. Nor is there evidence for identifying any specific causes for the changes observed, e.g., whether these are due to upwelling/downwelling effects or otherwise.

Fig. 14

Changes in bloom for consecutive satellite images of region 1 using retrieved OC3/OCI [Chla].


There is, also however, a caveat that might call into questions some of the above observations. We are aware of artifacts in the VIIRS retrieval imagery that tend to appear near the western edge of the image granule making it appear that [Chla] concentrations are higher than the same values detected near the eastern edge. Under those circumstances, these artifacts could be playing a part in the apparent increase of the bloom in the consecutive VIIRS images in Figs. 12 and 14, which are from near the eastern and western edges of the consecutive VIIRS granules (Fig. 11). This possibility may be negated, however, by an examination of consecutive VIIRS retrievals in adjacent bloom free waters. These retrievals, which are found to be identical, are shown blue on the left-hand side of Fig. 14, for bloom free waters very closely adjacent to the blooms in region 1, and therefore, also near the eastern and western ends of the VIIRS image granules for the consecutive images. Thus, Fig. 14 (left-hand side) shows that, at least for these bloom free waters, VIIRS retrievals from both near granule edges are identical in the consecutive VIIRS images and are not affected by edge artifacts. Their validity is further confirmed by an identical MODIS-A retrieval, also in Fig. 14, whose image is not from a near granule edge. The fact that the two consecutive VIIRS retrievals are the same in bloom free waters and are unaffected by edge artifacts in turn supports the interpretation that the changes observed at the bloom/nonbloom boundaries (dotted rectangle) in consecutive images in Fig. 14(a) and 14(c) are not due to edge artifacts, but do in fact actually show expansion of the bloom into bloom free waters. This interpretation probably remains valid, even though near edge artifacts might still significantly impact the accuracies of changes in the higher bloom concentrations away from the bloom/nonbloom boundaries, such as those indicated by the color-coded concentrations in the zoomed images from the solid rectangle in Fig. 14 (and shown in right-hand side). These color-coded concentrations may, therefore, not be accurate because of near edge artifacts. In further support that it is changes that are being observed, it should be noted that these artifacts that may apply to retrievals because of western near edge effects in the VIIRS images would not apply to, or impact, the MODIS-A image, which is not at an edge, and which also shows similar expansion of the increasing [Chla] bloom Fig. 14, region 1.

For comparison with the VIIRS OC3 retrievals, in Figs. 12 and 14, which were discussed above, we show, in Fig. 15, retrievals using our NN technique for the same consecutive VIIRS observations. Results are qualitatively very similar to those from OC3, Fig. 14. Again, they also show bloom expansion to the southwest. NN retrieved [Chla] values can be expected to differ in details from those of OC3, because of the different (higher) retrieval accuracies demonstrated by the NN technique (see Sec. 3.3). The zoomed features in Figs 15(a), 15(d), 15(e), and 15(h) show similarity with the equivalent OC3 retrievals in Fig. 14. Interestingly, we also observe essentially identical retrievals in the zoomed pixels of Figs. 15(b) and 15(f), which are from apparently unchanged high [Chla] areas in Figs. 15(c) and 15(g), respectively. This might imply that the Rrs values at 486, 551, and 671 nm used as inputs into the NN are identical. Future matchups with in situ measurements may shed more light on these matters.

Fig. 15

Changes in bloom for consecutive VIIRS images of region 1 using NN retrieved [Chla].


In the context of the above discussion, it is also worth reiterating that, as was noted previously in Sec. 2.2, all quality flags were applied for VIIRS and MODIS-A OC3/OCI ocean color level 2 information, as well as the Rrs values at 486, 551, and 671 nm for the NN retrievals, downloaded from NASA, which already has the appropriate corrections for both atmosphere and observation angle applied.71 Water-leaving reflectances for our selected consecutive scenes are assumed to be independent of viewing geometries after BRDF/Albedo and atmospheric corrections are made. We would also note that the viewing angles for the consecutive VIIRS images are close, <5  deg (θviirs_1 to θviirs_2. Furthermore, even if no correction was made for the 60  deg observation angles (less than that, for the critical western near edge involved in the above images), potential errors would be relatively small,7273.74 (1%) and would have no discernable impact on our qualitative interpretations that [Chla] changes are observed in the retrieved images, while at the same time, not saying anything with regard to absolute magnitudes or accuracies.

In conclusion for this section on consecutive satellite images, it is recognized that additional studies that include comparisons from consecutive VIIRS retrievals of Rrs values, as well as comparisons of [Chla] retrievals against simultaneous in situ measurements, would be needed to clarify the nature and magnitude of the changes being observed. However, given the difficulty of carrying out comprehensive calibration measurements of this type, we believe that it is reasonable to conclude that while consecutive overlapping satellite images can provide some evidence of temporal changes in KB HABs concentration in the WFS, they are unlikely to provide accurate or reliably useful information on the absolute magnitudes involved.

It should also be noted that while consecutive overlapping images appear to show temporal changes, there is insufficient evidence from them to attribute the relative contributions of drift, patchiness, upwelling/downwelling, or a combination of any of these to the causes of the changes. Section 3.6 presents the results of recent field measurements of KB HABs in the WFS, which much more solidly confirm KB HABs temporal variabilities as well as patchiness. They also support conclusions that the significantly improved retrieval accuracies that are obtained with shorter overlap time windows between satellite retrievals and in situ measurements (Sec. 3.3) reflect the impact of temporal variabilities.


Field Measurements

The evidence for temporal changes, intrapixel variations, and patchiness associated with blooms in the WFS is further supported more definitively by recent field measurements made in conjunction with Mote Marine Laboratories on 1/19/2017 off Lido Key, near Sarasota, Florida. Figure 16 shows the transect of measurements made. Several of these measurements were made at stations subpixel distances apart (generally 300 m) on an outward leg and were then repeated for the same stations, matched as closely as possible with a GPS, on a return leg. Samples were taken at a 0.5-m depth and cell concentrations obtained by analysis at Mote Marine Laboratories.

Fig. 16

Transect of outward and return legs of field measurements.


The values shown in Fig. 17 give the KB cell counts at the different stations and the times of measurement. These are color coded, so that the same row color indicates the results for the same station on both the outward and return leg. It should be noted that, as might be expected with these high cell counts, the latter showed an excellent match with simultaneous collocated high performance liquid chromatography (HPLC), [Chla] measurements. From Fig. 17, it can be seen that the changes in values at the same station generally increase with the time interval (between outward and return legs). Thus, the greatest change is for Station CV1701/CV1713 from 7.52×106 to 1.552×106  cellsL1 over the 120-min time interval between the two measurements. The second greatest change is that for the next station, CV1702/CV1712, where the change is from 1.776×106 to 1.326×106  cellsL1 for the 87-min time interval.

Fig. 17

Results of field measurements on 1/19/2017.


For the shortest time between measurements, CV1706/CV1708 the change is 0.952 to 0.690×106  cellsL1 over the 21-min interval between measurements. [It should also be noted that there is also a slight discrepancy in station positions recorded due to drift from measurement start (with GPS initial colocation) to completion and recording.]

These results illustrate both the intrapixel variations that can typically occur (as well as inter pixel variations) and also confirm the temporal variations that can be expected. The relative contributions of drift or upwelling/downwelling to the results are not known.

In general, the consecutive satellite images and the field measurement observations lend support to our underlying thesis that the significantly increased bloom retrieval accuracy that occurs by shortening of the overlap time window between observation and in situ measurement matchups from 100 to 15 min is due to temporal changes in the observed bloom. They also serve to underline that derived magnitudes from satellite observations may be valid only for brief periods. To deal with these uncertainties, we are currently examining temporal and spatial averaging possibilities.


Summary and Conclusions

In the work reported here, NN algorithms using Rrs values from the 486-, 551-, and 671-nm VIIRS bands are used to retrieve an image of aph443 values in the WFS. Then, these additional limiting constraints are applied, in two filter processes, F1 and F2 to eliminate from that image all aph443 pixels that are not compatible with the existence of KB HABs. The residual image then shows only retrieved aph443 values and their equivalent [Chla] values that are consistent with the existence of KB HABs. This procedure was then used to retrieve KB HABs in the WFS. The efficacy of these NN retrievals was evaluated by comparison of retrieval accuracies obtained against simultaneous colocated in situ measurements.

For meaningful quantitative comparisons, it is important to have many data points. Accordingly, we sought all available matchups between VIIRS NN aph443 (and equivalent [Chla] retrievals) and in situ KB cell count measurements for the period 2012 to 2016 for which there was available VIIRS data. These comparisons showed that for VIIRS observations, the NN technique appeared to offer good potential for effective retrievals of KB HABs cell counts in the WFS. More specifically, these comparisons showed that when the overlap time window between in situ observations and satellite overpass measurements was reduced from 100 to 15 min, retrieval accuracies greatly improved and showed increased correlations and reduced errors. The comparisons against in situ matchups were also carried out for VIIRS retrievals using other algorithms: OCI/OC3, GIOP, QAA, and RGCI.

It was seen that for the available and somewhat limited data sets, the NN retrievals exhibited the best retrieval accuracies, among the techniques tested, against the in situ measurement data, for both the longer (100 min) and, more importantly, the shorter (15 min) overlap time windows. This was observed both in terms of higher correlations and lower errors against the in situ measurements. The results confirm the potential efficacy for detecting and quantifying KB HABs in the WFS using the NN technique.

Finally, consecutive satellite images qualitatively illustrated the temporal changes that can be associated with KB HABs in the WFS. The evidence for temporal changes was even more strongly complemented by recent field measurements off Sarasota, Florida, that quantitatively confirm the temporal changes and patchiness observed with KB HABs in the WFS.


This work was partially supported by grants from National Oceanographic and Atmospheric Administration (NOAA) through NOAA CREST, and NOAA JPSS, as well as the NASA Public Health and Air Quality Program (NRA: NN13ZDA001N). We would like to thank the NASA Ocean Color Processing Group for satellite imagery. We also thank NOAA HABSOS and Florida Fish and Wildlife Conservation Commission’s Fish and Wildlife Research Institute (FWRI) groups for making in situ KB data available to us. The authors declare no conflict of interest.



A. El-Habashi et al., “Satellite retrievals of Karenia brevis harmful algal blooms in the West Florida shelf using neural networks and comparisons with other techniques,” Remote Sens. 8(5), 377 (2016).RSEND3http://dx.doi.org/10.3390/rs8050377Google Scholar


US Department of Commerce, National Oceanic and Atmospheric Administration, “Harmful algal blooms,” NOAA’s National Ocean Service, 16 November 2009, http://oceanservice.noaa.gov/hazards/hab (22 November 2016).Google Scholar


C. Hu et al., “Red tide detection and tracing using MODIS fluorescence data: a regional example in SW Florida coastal waters,” Remote Sens. Environ. 97(3), 311–321 (2005).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2005.05.013Google Scholar


R. Amin et al., “MODIS and MERIS detection of dinoflagellates blooms using the RBD technique,” Proc. SPIE 7473, 747304 (2009).PSISDG0277-786Xhttp://dx.doi.org/10.1117/12.830631Google Scholar


R. Amin et al., “Novel optical techniques for detecting and classifying toxic dinoflagellate Karenia brevis blooms using satellite imagery,” Opt. Express 17(11), 9126–9144 (2009).OPEXFF1094-4087http://dx.doi.org/10.1364/OE.17.009126Google Scholar


M. Tomlinson, T. Wynne and R. Stumpf, “An evaluation of remote sensing techniques for enhanced detection of the toxic dinoflagellate, Karenia brevis,” Remote Sens. Environ. 113(3), 598–609 (2009).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2008.11.003Google Scholar


G. A. Carvalho et al., “Satellite remote sensing of harmful algal blooms: a new multi-algorithm method for detecting the Florida red tide (Karenia brevis),” Harmful Algae 9(5), 440–448 (2010).HALNE7http://dx.doi.org/10.1016/j.hal.2010.02.002Google Scholar


G. A. Carvalho et al., “Long-term evaluation of three satellite ocean color algorithms for identifying harmful algal blooms (Karenia brevis) along the west coast of Florida: a matchup assessment,” Remote Sens. Environ. 115(1), 1–18 (2011).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2010.07.007Google Scholar


I. M. Soto Ramos, “Harmful algal blooms of the West Florida shelf and Campeche bank: visualization and quantification using remote sensing methods,” PhD Dissertation, University of South Florida College of Marine Science (2013).Google Scholar


D. Blondeau-Patissier et al., “A review of ocean color remote sensing methods and statistical techniques for the detection, mapping and analysis of phytoplankton blooms in coastal and open oceans,” Prog. Oceanogr. 123, 123–144 (2014).POCNA80079-6611http://dx.doi.org/10.1016/j.pocean.2013.12.008Google Scholar


C. Hu et al., “A harmful algal bloom of Karenia brevis in the Northeastern Gulf of Mexico as revealed by MODIS and VIIRS: a comparison,” Sensors 15(2), 2873–2887 (2015).SNSRES0746-9462http://dx.doi.org/10.3390/s150202873Google Scholar


I. M. Soto et al., “Evaluation and optimization of remote sensing techniques for detection of Karenia brevis blooms on the West Florida shelf,” Remote Sens. Environ. 170, 239–254 (2015).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2015.09.026Google Scholar


I. Ioannou et al., “Neural network approach to retrieve the inherent optical properties of the ocean from observations of MODIS,” Appl. Opt. 50(19), 3168–3186 (2011).APOPAI0003-6935http://dx.doi.org/10.1364/AO.50.003168Google Scholar


I. Ioannou et al., “Deriving ocean color products using neural networks,” Remote Sens. Environ. 134, 78–91 (2013).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2013.02.015Google Scholar


I. Ioannou et al., “Remote estimation of in water constituents in coastal waters using neural networks,” Proc. SPIE 9240, 92400T (2014).PSISDG0277-786Xhttp://dx.doi.org/10.1117/12.2067772Google Scholar


A. El-habashi and S. Ahmed, “Neural network algorithms for retrieval of harmful algal blooms in the West Florida shelf from VIIRS satellite observations and comparisons with other techniques, without the need for a fluorescence channel,” Proc. SPIE 9638, 96380B (2015).PSISDG0277-786Xhttp://dx.doi.org/10.1117/12.2195339Google Scholar


C. Le and C. Hu, “A hybrid approach to estimate chromophoric dissolved organic matter in turbid estuaries from satellite measurements: a case study for Tampa Bay,” Opt. Express 21(16), 18849–18871 (2013).OPEXFF1094-4087http://dx.doi.org/10.1364/OE.21.018849Google Scholar


J. P. Cannizzaro et al., “A novel technique for detection of the toxic dinoflagellate, Karenia brevis, in the Gulf of Mexico from remotely sensed ocean color data,” Cont. Shelf Res. 28(1), 137–158 (2008).CSHRDZ0278-4343http://dx.doi.org/10.1016/j.csr.2004.04.007Google Scholar


J. P. Cannizzaro et al., “Detection of Karenia brevis blooms on the West Florida shelf using in situ backscattering and fluorescence data,” Harmful Algae 8(6), 898–909 (2009).HALNE7http://dx.doi.org/10.1016/j.hal.2009.05.001Google Scholar


P. J. Werdell and S. W. Bailey, “An improved in-situ bio-optical data set for ocean color algorithm development and satellite data product validation,” Remote Sens. Environ. 98(1), 122–140 (2005).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2005.07.001Google Scholar


IOCCG, “International ocean-colour coordinating group,” June 2003, http://www.ioccg.org/groups/lee_data.pdf (9 March 2015).Google Scholar


A. Morel, “Optical properties of pure water and pure sea water,” Opt. Aspects Oceanogr. 1, 1–24 (1974).Google Scholar


H. R. Gordon et al., “A semianalytic radiance model of ocean color,” J. Geophys. Res.: Atmos. 93(D9), 10909–10924 (1988).http://dx.doi.org/10.1029/JD093iD09p10909Google Scholar


A. Morel, “Optical modeling of the upper ocean in relation to its biogenous matter content (Case I waters),” J. Geophys. Res. 93(10), 10749–10768 (1988).JGREA20148-0227Google Scholar


C. D. Mobley, Light and Water: Radiative Transfer in Natural Waters, Academic Press, San Diego, California (1994).Google Scholar


A. Bricaud et al., “Variability in the chlorophyll-specific absorption coefficients of natural phytoplankton: analysis and parameterization,” J. Geophys. Res.: Oceans 100(C7), 13321–13332 (1995).http://dx.doi.org/10.1029/95JC00463Google Scholar


R. P. Bukata et al., Optical Properties and Remote Sensing of Inland and Coastal Waters, CRC Press, Boca Raton, Florida (1995).Google Scholar


R. M. Pope and E. S. Fry, “Absorption spectrum (380-700 nm) of pure water. II. Integrating cavity measurements,” Appl. Opt. 36(33), 8710–8723 (1997).APOPAI0003-6935http://dx.doi.org/10.1364/AO.36.008710Google Scholar


A. Morel and S. Maritorena, “Bio-optical properties of oceanic waters-a reappraisal,” J. Geophys. Res. 106(C4), 7163–7180 (2001).JGREA20148-0227http://dx.doi.org/10.1029/2000JC000319Google Scholar


D. Stramski, A. Bricaud and A. Morel, “Modeling the inherent optical properties of the ocean based on the detailed composition of the planktonic community,” Appl. Opt. 40(18), 2929–2945 (2001).APOPAI0003-6935http://dx.doi.org/10.1364/AO.40.002929Google Scholar


A. M. Ciotti, M. R. Lewis and J. J. Cullen, “Assessment of the relationships between dominant cell size in natural phytoplankton communities and the spectral shape of the absorption coefficient,” Limnol. Oceanogr. 47(2), 404–417 (2002).LIOCAH0024-3590http://dx.doi.org/10.4319/lo.2002.47.2.0404Google Scholar


Z. Lee, K. L. Carder and R. A. Arnone, “Deriving inherent optical properties from water color: a multiband quasi-analytical algorithm for optically deep waters,” Appl. Opt. 41(27), 5755–5772 (2002).APOPAI0003-6935http://dx.doi.org/10.1364/AO.41.005755Google Scholar


M. Babin et al., “Variations in the light absorption coefficients of phytoplankton, nonalgal particles, and dissolved organic matter in coastal waters around Europe,” J. Geophys. Res.: Oceans 108(C7), 3211 (2003).http://dx.doi.org/10.1029/2001JC000882Google Scholar


Z. Lee, “Reports of the international ocean-colour coordinating group: fundamentals, tests of algorithms, and applications,” IOCCG Report 5, Dartmouth, Canada (2006).Google Scholar


C. D. Mobley and L. K. Sundman, Hydrolight 4.2 Technical Documentation, p. 84, Sequoia Scientific, Incorporated, Redmond, Washington (2001).Google Scholar


F. Aires et al., “A new neural network approach including first-guess for retrieval of atmospheric water vapor, cloud liquid water path, surface temperature and emissivities over land from satellite microwave observations,” J. Geophys. Res. Atmos. 106(D14), 14887–14907 (2001).Google Scholar


F. Aires, C. Prigent and W. B. Rossow, “Neural network uncertainty assessment using Bayesian statistics: a remote sensing application,” Neural Comput. 16(11), 2415–2458 (2004).NEUCEB0899-7667http://dx.doi.org/10.1162/0899766041941925Google Scholar


NASA’s OceanColor Web by the Ocean Biology Processing Group (OBPG) at NASA’s Goddard Space Flight Center, “Chlorophyll-a algorithm,” https://oceancolor.gsfc.nasa.gov/PRODUCTS (22 November 2016).Google Scholar


NOAA Harmful Algal Blooms Observing System, “HABSOS,” (15 July 2012), https://habsos.noaa.gov/ (22 November 2016).Google Scholar


J. O’Reilley et al., SeaWiFS Postlaunch Calibration and Validation Analyses, Part 3, , S. B. Hooker and E. R. Firestone, Eds., p. 49, NASA Goddard Space Flight Center, Hampton, Virginia (2000).Google Scholar


J. E. O’Reilly et al., “Ocean color chlorophyll algorithms for SeaWiFS,” J. Geophys. Res.: Oceans 103(C11), 24937–24953 (1998).http://dx.doi.org/10.1029/98JC02160Google Scholar


C. Hu, Z. Lee and B. Franz, “Chlorophyll a algorithms for oligotrophic oceans: a novel approach based on three-band reflectance difference,” J. Geophys. Res.: Oceans 117(C1), 73 (2012).http://dx.doi.org/10.1029/2011JC007395Google Scholar


NASA’s OceanColor Web by the Ocean Biology Processing Group (OBPG) at NASA’s Goddard Space Flight Center, “Chlorophyll-a algorithm,” https://oceancolor.gsfc.nasa.gov/atbd/chlor_a/ (22 November 2016).Google Scholar


K. Carder and R. Steward, “A remote-sensing reflectance model of a red-tide dinoflagellate off West Florida,” Limnol. Oceanogr. 30(2), 286–298 (1985).LIOCAH0024-3590http://dx.doi.org/10.4319/lo.1985.30.2.0286Google Scholar


R. P. Stumpf, “Applications of satellite ocean color sensors for monitoring and predicting harmful algal blooms,” Hum. Ecol. Risk Assess.: Int. J. 7(5), 1363–1368 (2001).http://dx.doi.org/10.1080/20018091095050Google Scholar


K. L. Mahoney, “Backscattering of light by Karenia brevis and implications for optical detection and monitoring,” PhD dissertation, Univ. of South Mississippi, Stennis Space Center, p. 116 (2003).Google Scholar


R. Stumpf et al., “Monitoring Karenia brevis blooms in the Gulf of Mexico using satellite ocean color imagery and other data,” Harmful Algae 2(2), 147–160 (2003).HALNE7http://dx.doi.org/10.1016/S1568-9883(02)00083-5Google Scholar


M. C. Tomlinson et al., “Evaluation of the use of SeaWiFS imagery for detecting Karenia brevis harmful algal blooms in the eastern Gulf of Mexico,” Remote Sens. Environ. 91(3), 293–303 (2004).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2004.02.014Google Scholar


R. Stumpf, M. Tomlinson, “Remote sensing of harmful algal blooms,” in Remote Sensing of Coastal Aquatic Environments: Technologies, Techniques and Applications, , R. Miller, C. Del Castillo and B. McKee, Eds., 277–292, Kluwer Academic Publishers, Springer, Dordrecht, The Netherlands (2005).Google Scholar


T. T. Wynne et al., “Detecting Karenia brevis blooms and algal resuspension in the western Gulf of Mexico with satellite ocean color imagery,” Harmful Algae 4(6), 992–1003 (2005).HALNE7http://dx.doi.org/10.1016/j.hal.2005.02.004Google Scholar


S. E. Craig et al., “Use of hyperspectral remote sensing reflectance for detection and assessment of the harmful alga, Karenia brevis,” Appl. Opt. 45(21), 5414–5425 (2006).APOPAI0003-6935http://dx.doi.org/10.1364/AO.45.005414Google Scholar


P. A. Tester et al., “Relationships among water column toxins, cell abundance and chlorophyll concentrations during Karenia brevis blooms,” Cont. Shelf Res. 28(1), 59–72 (2008).CSHRDZ0278-4343http://dx.doi.org/10.1016/j.csr.2007.04.007Google Scholar


R. P. Stumpf et al., “Skill assessment for an operational algal bloom forecast system,” J. Mar. Syst. 76(1), 151–161 (2009).JMASE50924-7963http://dx.doi.org/10.1016/j.jmarsys.2008.05.016Google Scholar


L. Qi et al., “VIIRS observations of a Karenia brevis bloom in the northeastern Gulf of Mexico in the absence of a fluorescence band,” IEEE Geosci. Remote Sens. Lett. 12(11), 2213–2217 (2015).http://dx.doi.org/10.1109/LGRS.2015.2457773Google Scholar


R. M. Letelier and M. R. Abbott, “An analysis of chlorophyll fluorescence algorithms for the moderate resolution imaging spectrometer (MODIS),” Remote Sens. Environ. 58(2), 215–223 (1996).RSEEA70034-4257http://dx.doi.org/10.1016/S0034-4257(96)00073-9Google Scholar


A. Gilerson et al., “Fluorescence component in the reflectance spectra from coastal waters. Dependence on water composition,” Opt. Express 15(24), 15702–15721 (2007).OPEXFF1094-4087http://dx.doi.org/10.1364/OE.15.015702Google Scholar


M. J. Behrenfeld et al., “Satellite-detected fluorescence reveals global physiology of ocean phytoplankton,” Biogeosciences 6(5), 779–794 (2009).http://dx.doi.org/10.5194/bg-6-779-2009Google Scholar


J. Werdell, “Global bio-optical algorithms for ocean color satellite applications: inherent optical properties algorithm workshop at ocean optics XIX; Barga, Italy, 3-4 October 2008,” Eos, Trans. Am. Geophys. Union 90(1), 4 (2009).http://dx.doi.org/10.1029/2009EO010005Google Scholar


P. J. Werdell et al., “Generalized ocean color inversion model for retrieving marine inherent optical properties,” Appl. Opt. 52(10), 2019–2037 (2013).http://dx.doi.org/10.1364/ao.52.002019Google Scholar


P. J. Werdell et al., “Retrieving marine inherent optical properties from satellites using temperature and salinity-dependent backscattering by seawater,” Opt. Express 21(26), 32611–32622 (2013).OPEXFF1094-4087http://dx.doi.org/10.1364/OE.21.032611Google Scholar


Z. P. Lee, “QAA algorithm,” 2002, http://ioccg.org/groups/Software_OCA/QAA_v5.pdf (22 November 2016).Google Scholar


Z. Lee and K. L. Carder, “Absorption spectrum of phytoplankton pigments derived from hyperspectral remote-sensing reflectance,” Remote Sens. Environ. 89(3), 361–368 (2004).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2003.10.013Google Scholar


P. J. Franks, “Spatial patterns in dense algal blooms,” Limnol. Oceanogr. 42(5), 1297–1305 (1997).LIOCAH0024-3590http://dx.doi.org/10.4319/lo.1997.42.5_part_2.1297Google Scholar


L. Leng et al., “Ordinary least square regression, orthogonal regression, geometric mean regression and their applications in aerosol science,” J. Phys. Conf. Ser. 78, 012084 (2007).JPCSDZ1742-6588http://dx.doi.org/10.1088/1742-6596/78/1/012084Google Scholar


R. Arnone et al., “Diurnal changes in ocean color in coastal waters,” Proc. SPIE 9827, 982711 (2016).PSISDG0277-786Xhttp://dx.doi.org/10.1117/12.2241018Google Scholar


R. Arnone et al., “Ocean Weather—Interaction of physical and bio-optical processes across a river plume dominated shelf in the Gulf of Mexico,” in AGU Ocean Sciences, New Orleans (2016).Google Scholar


P. A. Tester and K. A. Steidinger, “Gymnodinium breve red tide blooms: Initiation, transport, and consequences circulation,” Liminol. Oceanogr. 42, 1039–1051 (1997).http://dx.doi.org/10.4319/lo.1997.42.5_part_2.1039Google Scholar


Y. Liu and R. H. Weisberg, “Patterns of ocean current variability on the West Florida shelf using the self-organizing map,” J. Geophys. Res. 110(C6) (2005).http://dx.doi.org/10.1029/2004JC002786Google Scholar


Y. Liu and R. H. Weisberg, “Seasonal variability on the West Florida shelf,” Prog. Oceanogr. 104, 80–98 (2012).POCNA80079-6611http://dx.doi.org/10.1016/j.pocean.2012.06.001Google Scholar


NASA ocean motion and surface currents, “Ocean in motion: ekman transport background,” http://oceanmotion.org/html/background/ocean-in-motion.htm (22 November 2016).Google Scholar


NASA’s OceanColor Web by the Ocean Biology Processing Group (OBPG) at NASA’s Goddard Space Flight Center, “Remote sensing reflectance (Rrs) algorithm,” https://oceancolor.gsfc.nasa.gov/atbd/rrs/ (22 November 2016).Google Scholar


H. Yang and H. R. Gordon, “Remote sensing of ocean color: assessment of water-leaving radiance bidirectional effects on atmospheric diffuse transmittance,” Appl. Opt. 36(30), 7887–7897 (1997).APOPAI0003-6935http://dx.doi.org/10.1364/AO.36.007887Google Scholar


H. R. Gordon and B. A. Franz, “Remote sensing of ocean color: assessment of the water-leaving radiance bidirectional effects on the atmospheric diffuse transmittance for SeaWiFS and MODIS intercomparisons,” Remote Sens. Environ. 112(5), 2677–2685 (2008).RSEEA70034-4257http://dx.doi.org/10.1016/j.rse.2007.12.010Google Scholar


C. D. Mobley et al., “Atmospheric correction for satellite ocean color radiometry,” A Tutorial and Documentation of the Algorithms Used by the NASA Ocean Biology Processing Group (2016).Google Scholar


Ahmed El-Habashi is a PhD student in the Department of Electrical Engineering at the City College of New York. He is a NOAA CREST fellow. His research in the Optical Remote Sensing Laboratory (ORSL) focuses on the remote sensing of ocean color and the detection of harmful algal blooms (HABs), as well as on the polarization properties of oceanic coastal waters and their relationship to fluorescence.

Claudia M. Duran is an undergraduate student in the Department of Earth System Science and Engineering at the City College. Her research in the ORSL has focused on the evaluation of satellite ocean color remote sensing techniques to retrieve HABs.

Vincent Lovko received his bachelor’s degree in environmental science and his PhD at the Virginia Institute of Marine Science, Gloucester, Virginia. He is a staff scientist and manages the Phytoplankton Ecology Research Program at the Mote Marine Laboratory in Sarasota, Florida. He leads Mote’s red tide program which conducts red tide monitoring and research under contract with the state of Florida, providing data for state red tide status updates and other collaborative data products.

Michelle C. Tomlinson received her BS degree in marine science biology from Southampton College of Long Island University and her MS degree in oceanography from Old Dominion University. She is an oceanographer with the National Centers for Coastal Ocean Science, NOAA. Her research focuses on the application of satellite-derived ocean color to detect, monitor, and forecast the occurrence of HABs, leading to the development of HAB forecast systems in the Gulf of Mexico and Lake Erie.

Richard P. Stumpf received his BS degree in environmental sciences from the University of Virginia, and his MS and PhD degrees in marine studies from the University of Delaware. He is an oceanographer with the National Centers for Coastal Ocean Science, NOAA. He leads research focusing on the application of satellite-derived ocean color to detect, monitor, and forecast the occurrence of HABs which has led to the development of an HAB forecast systems in the Gulf of Mexico and Lake Erie.

Sam Ahmed received his BA and MA degrees in electrical engineering from Cambridge University and his PhD from the University College, London. As director of the ORSL at the City College, New York, he has been involved in atmospheric LIDAR remote sensing of the atmosphere and in satellite remote sensing of ocean color.

Ahmed El-Habashi, Claudia M. Duran, Vincent Lovko, Michelle C. Tomlinson, Richard P. Stumpf, Sam Ahmed, "Satellite retrievals of Karenia brevis harmful algal blooms in the West Florida shelf using neural networks and impacts of temporal variabilities," Journal of Applied Remote Sensing 11(3), 032408 (11 May 2017). http://dx.doi.org/10.1117/1.JRS.11.032408
Submission: Received 30 November 2016; Accepted 11 April 2017


Neural networks

Image filtering


Earth observing sensors

Satellite imaging


Back to Top