Satellite Retrievals of Karenia brevis Harmful Algal Blooms in the West Florida Shelf Using Neural Networks and Comparisons with Other Techniques

Abstract

We describe the application of a Neural Network (NN) previously developed by us, to the detection and tracking, of Karenia brevis Harmful Algal Blooms (KB HABs) that plague the coasts of the West Florida Shelf (WFS) using Visible Infrared Imaging Radiometer Suite (VIIRS) satellite observations. Previous approaches for the detection of KB HABs in the WFS primarily used observations from the Moderate Resolution Imaging Spectroradiometer Aqua (MODIS-A) satellite. They depended on the remote sensing reflectance signal at the 678 nm chlorophyll fluorescence band (Rrs678) needed for both the normalized fluorescence height (nFLH) and Red Band Difference algorithms (RBD) currently used. VIIRS which has replaced MODIS-A, unfortunately does not have a 678 nm fluorescence channel so we customized the NN approach to retrieve phytoplankton absorption at 443 nm (a_ph443) using only Rrs measurements from existing VIIRS channels at 486, 551 and 671 nm. The a_ph443 values in these retrieved VIIRS images, can in turn be correlated to chlorophyll-a concentrations [Chla] and KB cell counts. To retrieve KB values, the VIIRS NN retrieved a_ph443 images are filtered by applying limiting constraints, defined by (i) low backscatter at Rrs 551 nm and (ii) a minimum a_ph443 value known to be associated with KB HABs in the WFS. The resulting filtered residual images, are then used to delineate and quantify the existing KB HABs. Comparisons with KB HABs satellite retrievals obtained using other techniques, including nFLH, as well as with in situ measurements reported over a four year period, confirm the viability of the NN technique, when combined with the filtering constraints devised, for effective detection of KB HABs.

1. Introduction

A new approach is needed for the detection and tracking of KB HABs that frequently plague the coasts and beaches of the West Florida Shelf (WFS) using 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 [1]. Effective KB HABs detection and tracking approaches are needed for use with VIIRS so that NOAA can extend its HABs monitoring capabilities, which previously relied on MODIS-A imagery [2,3,4,5,6,7,8,9,10,11]. Unfortunately, VIIRS, unlike its predecessor MODIS-A, does not have a 678 nm channel to detect chlorophyll fluorescence, which is used in the normalized fluorescence height (nFLH) algorithm [7,8], or in the Red Band Difference (RBD) algorithm, developed by our CCNY group [5,6]. Both of these techniques have demonstrated that the Rrs 678 nm from the MODIS-A fluorescence band helps in effectively detecting and tracking KB HABs in the WFS [3,4,5,6,7,8]. To overcome the lack of a fluorescence channel on VIIRS, the alternative approach described here, 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 neural network (NN) inversion algorithm, previously developed and reported by us [12,13,14,15], that takes VIIRS Rrs measurements at the 486, 551 and 671 nm bands (or 488, 555, and 667 nm for MODIS-A) as inputs, and produces as output the related Inherent Optical Properties (IOPs), namely: absorption coefficients of phytoplankton (a_ph443) dissolved organic matter (a_g) and non-algal particulates (a_dm) as well as the particulate backscatter coefficient, (bb_p) all at 443 nm. In this work, however, it is only with the NN output of a_ph443 that we are concerned with and which can be converted to equivalent [Chla] using empirical relationships for specific absorption in the WFS which have been determined from in situ measurements [16]. For the next step, to obtain KB values from the VIIRS NN retrieved a_ph443 image, we apply additional constraints, defined by (i) low backscatter manifested as a maximum Rrs551 value and (ii) a minimum [Chla] threshold [4,17,18] (and hence an equivalent minimum a_ph443min value) both known to be associated with KB HABs in the WFS. These two constraining filter processes are applied sequentially to the VIIRS NN retrieved a_ph443 image. First an image is made of retrieved VIIRS Rrs551. A mask is then made of all pixels with Rrs551 ≥ Rrs551max, the maximum value known to be compatible with the existence KB HABs. This mask is then applied, in a filter process hereinafter dubbed F1, to the VIIRS NN retrieved a_ph443 image to exclude pixels with Rrs551 ≥ Rrs551max. The residual image, after application of filter process F1, will only show a_ph443 values that also comply with Rrs551 ≤ Rrs551max. Then, in a second filter process, hereinafter dubbed F2, all values of a_ph443 ≤ a_ph443min are eliminated. The residual image will now only show a_ph443 values that are compatible with both criteria for KB HABs, and are therefore representative of KB HABs.

It will be shown that when both these filter condition are applied to VIIRS NN a_ph443 retrievals, they can be used to effectively delineate and quantify KB HABs in the WFS. The KB HABs retrieved in this manner also show good correlations with in situ KB HABs measurements as well as with nFLH retrievals and other techniques to which the same filtering criteria have been applied, confirming the viability of the approach.

2. Materials and Methods

Neural Network Algorithm Background

In the work presented here, we use our synthetically trained NN algorithm [12,13,14,15] to solve the inverse problem [19,20] of retrieving physical variables, including a_ph443, from VIIRS observations of Rrs at 486, 551 and 671 nm in the WFS. A brief background summary is given here. Additional details are available in the cited references and in the Appendix. The latter also gives the necessary parameters for using the NN with a MATLAB tool for satellite retrievals by any interested reader. As mentioned in the introduction, we use a standard multiband NN inversion algorithm that takes VIIRS Rrs measurements at the 486, 551 and 671 nm bands as inputs, and produces as outputs the related IOPs, namely: a_ph443, a_g and a_dm as well as bb_p, all at 443 nm. (As mentioned earlier, it is only with the a_ph443 output of NN that we are concerned with here).

Synthetic Dataset

For development of the algorithm, available measurement data from the NASA Bio-Optical Marine Algorithm Data set [21], (NOMAD) was used as a basis for establishing a range of possible IOP parameters and examining potential relationships between them. The full range of parameters, applicable to both oceanic and coastal waters, and the relationships established between them were used as inputs to a bio-optical model [14,22,23]. The range and variability of IOPs used in our bio-optical modeling is well represented in the literature [22,23,24,25,26,27,28,29,30,31,32,33,34,35]. Our model, as illustrated in Figure 1a, has four components: water, phytoplankton, colored dissolved organic matter (CDOM) and non-algal particulates (NAP) as inputs to create a synthetic data base consisting of 20,000 IOP combinations covering the entire observable range of possible IOP parameters. These IOPs were then used with a HydroLight based [36], parameterized forward model, described in Lee 2002 [23] to generate corresponding sets of Rrs values at 486, 551 and 671 nm for VIIRS, and at 488, 555 and 667 nm for MODIS-A, associated with each of the 20,000 IOPs (and the related different combinations of the four components) [12,13,14,15].

Algorithm Training

Once the comprehensive set of Rrs spectral channel data together with the associated IOP’s are assembled, a robust NN can be developed to connect the Rrs directly to the physical parameter to be retrieved. As is standard, a subset of the data is used to train the underlying NN and the rest used for validations. In our case, 50% (10,000) of these sets of Rrs and related IOPs were used [12,13,14,15] to train the NN model to obtain the relationship between input Rrs values at the 486, 551 and 671 nm bands for VIIRS and the output a_ph, a_g, a_dm, bb_p coefficients at 443 nm, which is at the peak of phytoplankton absorption spectrum, and thus exhibits the most sensitive variation in natural waters (and where a_g and a_d are mutually constrained through an empirically derived relationship based on a study of NOMAD [21] values). Figure 1b shows the VIIRS NN Architecture used. In our implementation, it was found that a one-hidden-layer multilayer perceptron (MLP-NN) with six neurons at the hidden layer was able to model our problems well. Since this architecture appeared appropriate for our needs, we did not further explore how this relationship might be improved by increasing the number of neurons or layers.

The VIIRS NN algorithm was tested and evaluated using multiple experiments. It was (a) tested initially on half the simulated data set that was not used in the training (b) field data from the, (NOMAD) and (c) data from our own field campaign in the Chesapeake Bay during 2013, which represents well the typical range of water optical properties and chlorophyll-a concentrations in coastal regions. These tests confirmed good retrievals of a_ph443 by the NN [13,14] and provided strong motivation for its use in the current work. The results of the present work have further confirmed its viability for effective retrievals of a_ph443 in the coastal waters of the WFS from VIIRS measurements at the 486, 551 and 671 nm bands. In this context, we would also add that though we are estimating a_ph443 at 443 nm, the Rrs values used as inputs to the NN to retrieve it are at longer wavelengths, 486, 551 and 671 nm, and do not require measurements at the shorter wavelength channels, M1 (415 nm) and M2 (443 nm) which are typically most impacted by atmospheric correction difficulties.

As mentioned above, a more detailed description of the NN, and its background are given in the Appendix. The parameters needed for its implementation and the retrieval a_ph443 from inputs of Rrs 486, 551 and 671 nm. The related calculations that can be readily performed by interested readers using a MATLAB script are also described in the Appendix. We would be interested to hear the results of other retrievals, particularly for other areas where KB HABs are known to exist.

3. Results and Discussion

3.1. VIIRS Retrievals of Rrs551 and aph₄₄₃ and Determination of Limiting Values of Rrs551max and aph₄₄₃min in a KB Bloom Environment

Figure 2 shows KB cell counts associated with a prominent KB HAB bloom in the WFS with its peak occurring around 2 September 2014. This bloom has been extensively studied [7,8,15]. Figure 2, from the NOAA HABSOS site [37] shows the bloom area with indicators (circles of different sizes) for KB HABs cell counts for concurrent or near concurrent dates over the period (27 August–17 September 2014). Figure 3a shows VIIRS retrievals of Rrs551 for 2 September 2014 and for the same location. It shows, overlaid, the cell count information from Figure 2. Considering backscatter data as a KB indicator, Refs. [3,17,18] give maximum permissible values of backscatter, compatible with KB HABs, as bb_p550 ≤ 0.0045 m²·mg⁻¹ at 550 nm. In addition, for Rrs 551 which is taken to serve as a proxy for backscatter, the equivalent max permissible value is Rrs551 ≤ 7.0 × 10⁻³ sr⁻¹. However, by inspection of Figure 3a of retrieved [38] and in situ [37] values we concluded that a value of approximately Rrs551 ≤ 6.0 × 10⁻³ sr⁻¹ appears more appropriate as the highest permissible limiting value for compatibility with KB HABs. If that value is applied to Figure 3a, as filter process F1, all pixels with Rrs551 ≥ 6.0 × 10⁻³ sr⁻¹ are screened out, leaving the residual pixels in Figure 3b as compatible with KB HABs as far as backscatter is concerned.

We next apply the pixel mask shown in Figure 3b, resulting from application of Filter 1 conditions to VIIRS Rrs551 retrievals, to Figure 4a which shows a_ph443 retrieved by the NN for the same 2 September 2014 date and location. This process eliminates all a_ph443 values in Figure 4a that do not satisfy Rrs551 ≤ 6.0 × 10⁻³ sr⁻¹ requirements for compatibility with KB HABs. The residual values, shown in Figure 3b satisfy the Rrs551 limit requirements. Cell counts from Figure 2 are again shown overlaid in Figure 4b.

We next examine the appropriate F2 filter values for minimum a_ph443 consistent with KB HABs. Now Refs. [17,18] give the minimum [Chla] concentration compatible with KB HABs in the WFS as [Chla] ≥ 1.5 µg·L⁻¹. This is converted into equivalent a_ph443 using the empirical relationship [16]:

$$a_{ph443}=0.051\times {\left[Chla\right]}^{0.74},m^{-1}$$

(1)

Which gives a_ph443 ≥ 0.0688 m⁻³ as the minimum permissible value. Again, by inspection of Figure 4b against overlaid in situ values, we conclude that a value of approximately a_ph443 ≥ 0.061 m⁻³ is more appropriate as the limiting value for compatibility with KB HABs.

Applying this value as Filter F2 to Figure 4b we eliminate pixels that are not compatible with KB and end up with Figure 5a. This shows only residual values with a_ph443 ≥ 0.061 m⁻³ which also satisfy Rrs551 < 6.0 × 10⁻³ sr⁻¹. These are therefore residual values compatible with and representative of KB HABs. Comparison of the residual a_ph443 (left hand scale) and equivalent [Chla] (right hand scale) in Figure 5a, which represent the retrieved KB HABs values, show good qualitative agreement with the KB HABs cell counts in Figure 2.

3.2. Comparisons of NN KB HABs with Other Retrieval Techniques

We next compare the NN retrievals of KB HABs, described in the last section, with retrievals using other techniques. 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 [2,3,4,5,6,7,8,9,10,11]. For more details, the reader is referred to recent excellent surveys and comparisons in [3,11]. We focus here, instead, on comparisons with three distinct categories of retrieval techniques.

The existing 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 retrieve it primarily in one of three ways: (i) using the remote sensing reflectance fluorescence signal at Rrs678 measured by the MODIS-A satellite; and (ii) using blue-green ratio algorithms, exemplified by the NASA OC3 and OCI products [21,39,40,41]; and (iii) a more recent WFS region specific empirical algorithm, the Red Green chlorophyll-a Index (RGCI) which retrieves [Chla] using 2 visible VIIRS or MODIS-A bands [42]. Those methods (i) using the 678 nm fluorescence signal, include the normalized fluorescence height (nFLH) [2,3,4,5,6,7,8,9,10,11] and Red Band Difference (RBD) techniques [5,6]. (nFLH is FLH normalized to solar radiance); Those methods (ii) using blue-green ratio algorithms, include the chlorophyll-a anomaly technique and related approaches [43,44,45,46,47,48,49,50,51,52]. The method using RGCI, has been modified very recently, and has been reported for [Chla] retrievals but has not yet been applied to KB retrievals [42].

Amongst the various available KB HABs retrieval techniques that have been comprehensively evaluated, nFLH and RBD exhibited the best KB HABs retrieval statistics [3] for the WFS. Accordingly, our paper focuses on comparisons of NN retrievals against retrievals using (i) nFLH/RBD fluorescence and (ii) OCI/OC3 approaches; and (iii) RGCI approaches. These are each described in Section 3.2.1, Section 3.2.2, Section 3.2.3 and Section 3.2.4 respectively, with detailed descriptions available in references [2,3,4,5,6,7,8,9,10,11,42,43,44,45,46,47,48,49,50,51,52,53].

3.2.1. Comparison of NN KB HABs Retrieval Techniques with Those Using the Normalized Fluorescence Height (nFLH) and Red Band Difference (RBD) Techniques

Both nFLH and RBD techniques use the magnitude of Rrs at the 678 nm fluorescence band relative to the values at adjacent bands on MODIS-A to detect the chlorophyll fluorescence [54] and [Chla] [55]. This approach, when coupled with the known low backscatter properties of KB, has been shown to help in effectively detecting KB in the WFS [3,17,18].

Since the nFLH technique requires the fluorescence signal at 678 nm, which is absent from VIIRS, comparisons of nFLH with the NN technique must be made using MODIS-A neural network retrievals. For this purpose, the same approach used with VIIRS NN (above) is used to obtain MODIS-A NN a_ph443 retrievals. For MODIS-A, the Rrs values at the 488, 555 and 667 nm bands are used as inputs to the NN. For application of the F1 filter process, Rrs 555 nm is used as the proxy for backscatter (instead of the Rrs 551 nm used with VIIRS). The same F1 and F2 filter processes, with the same limiting values of Rrs555 ≤ 6.0 × 10⁻³ sr⁻¹ for F1, and a_ph443 ≥ 0.061 m⁻³ for F2, were applied to MODIS-A NN retrievals for the same 2 September 2014 bloom date, to obtain KB compatible residual retrieval values. Figure 5b shows the MODIS-A NN retrieved a_ph443 (left hand scale) and the equivalent [Chla] (right hand scale), after application of F1 and F2 filters. As can be seen, Figure 5b shows good qualitative agreement with Figure 5a from VIIRS NN retrievals.

We next compare the nFLH NASA product available [54] with MODIS-A for retrievals of nFLH values for the same 2 September 2014 WFS bloom. nFLH uses the height of the Rrs 678 nm (over a baseline connecting adjacent Rrs values) as an indicator of chlorophyll fluorescence magnitude [54,55]. Retrieved values are shown in Figure 6a. The left hand scale shows the nFLH NASA product, while the right hand scale shows the equivalent [Chla] value, converted using the empirical relationship [7]:

$$[Chla]=1.255\times {(\mathrm{FLH}\times 10)}^{0.86},\mu g{L}^{-1}$$

(2)

Figure 6b shows the same nFLH retrievals, but with filters F1 and F2 applied, i.e., with the same pixels filtered out as in the MODIS-A NN retrievals in Figure 5b. It can be seen that Figure 6b shows good qualitative agreement with the images of Figure 5a,b.

For more quantitative comparisons, Figure 7a shows the correlation between equivalent [Chla] derived from nFLH in Figure 6a, against the retrieved MODIS-A NN derived equivalent [Chla]. To avoid complications from more complex waters in-shore areas, the comparison is made for the off-shore area A, shown demarcated in Figure 6a,b above. Without the KB compatibility filters applied, a high correlation, R² = 0.87 is obtained. With the F1 and F2 filters applied (eliminating non KB compatible pixels) the correlation becomes R² = 0.71, Figure 7b. The color coding of the dots in Figure 7a,b indicates distance from shore, with red being the closest. As can be seen the spread gets larger closer to the shore, where waters are more complex and the uncertainty increases.

When the comparison between nFLH and VIIRS NN retrieved [Chla] is expanded beyond area A to include all areas, including the more complex in-shore waters, the correlation between them drops drastically, to R² = 0.19, shown in Figure 8. This may be due to the poor retrievals associated with nFLH for in-shore areas, where increased backscatter can cause an apparent increase in the red-near infrared spectral peak, increasing the apparent nFLH chlorophyll fluorescence signal, but which is in fact only partially associated with the fluorescence peak [3,53,55,56]. Additional comparisons of nFLH and VIIRS NN in-shore retrievals are needed before more definitive conclusions can be arrived at.

3.2.2. Comparisons of MODIS-A NN Retrievals with Red Band Difference (RBD) Index Retrievals

We also examined retrievals from MODIS-A using the RBD index which is an empirical index previously developed by our CCNY group [5,6] often used in conjunction with the nFLH technique to confirm presence of KB blooms. The RBD index is defined by:

$$RBD=nLw(678)-nLw(667),W·m^{-2}·\mathsf{\mu }m·s{r}^{-1}$$

Figure 9 shows the RBD index, as defined above, retrieved from MODIS-A observations for the 2 September 2014 bloom, with filters F1 and F2 applied. As can be seen, this shows good qualitative agreement with both nFLH, MODIS-A NN [Chla], and HABSOS in situ images, Figure 2, Figure 5b and Figure 6b. As mentioned above, RBD, along with nFLH, has exhibited the best statistics for KB HABs retrievals [3].

3.2.3. Comparisons of VIIRS NN Retrievals with VIIRS Ocean Color Chlorophyll-a (OCI/OC3) Retrievals

We also compared the NN retrievals from VIIRS with retrievals using the OCI/OC3 indexes, both NASA products [53] for the same 2 September 2014 bloom. OCI and the related OC3 NASA products are [Chla] retrieval algorithms that make use of the ratio of blue/green bands in the MODIS-A and VIIRS satellites [21,39,40,41], and were found to yield exactly the same retrievals for our WFS conditions.

Correlations between VIIRS OCI/OC3 [Chla] and VIIRS NN equivalent [Chla] retrievals are shown in Figure 10a for the 2 September 2014 bloom for the off-shore area A. Without filter applications, R² = 0.9. When filters F1 and F2 are applied, the residual values consistent with KB HABs show, Figure 10b, a correlation of R² =0.78. Figure 10c shows a comparison for area A of nFLH equivalent [Chla] retrievals against OCI/OC3 [Chla] retrievals with an R² =0.75 without filters. Figure 10d shows R² = 0.50 with filters F1 and F2 applied, considerably lower than the correlation R² =0.71, Figure 7b, between nFLH equivalent [Chla] versus NN equivalent [Chla].

Again, when the comparison between OCI/OC3 and MODIS-A NN retrieved [Chla] is expanded beyond area A to include all areas, including the more complex in-shore waters, the correlation between them drops drastically, to R² = 0.07 (not shown) from the R² = 0.75 in Figure 10c for the off-shore area A. This may be due to the poorer retrievals associated with both nFLH and OCI/OC3 for more complex in-shore waters with increased backscatter [3,21,39,40,41,53,56].

3.2.4. Comparisons of VIIRS NN Retrievals with Red/Green Chlorophyll-a Index (RGCI) Retrievals

We next examine RGCI [Chla] retrievals for the same 2 September 2014 date and locations as in the above Sections. Equation (3a) below gives the relationship for RGCI MODIS-A [Chla] retrievals, and Equation (3b) that for RGCI VIIRS [Chla] retrievals [42].

$$[Chla{]}_{\mathrm{RGCI}}=0.1\times \text{exp }\left(11.8\times \frac{Rrs667}{Rrs547}\right),\mathsf{\mu }\mathrm{g}{\mathrm{L}}^{-1}$$

(3a)

$$[Chla{]}_{\mathrm{RGCI}}=0.1\times \text{exp }\left(11.8\times \frac{Rrs671}{Rrs551}\right),\mathsf{\mu }\mathrm{g}{\mathrm{L}}^{-1}$$

(3b)

We first compare VIIRS RGCI [Chla] retrievals against VIIRS NN [Chla] retrievals. Figure 11a shows a correlation of R² = 0.72 for open ocean area A. This drops to R² = 0.49 when filters F1 and F2 are applied, Figure 11b. This compares with R² = 0.9 and R² = 0.78 for the equivalent comparisons between OCI/OC3 and VIIRS NN [Chla].

We next compare RGCI [Chla] against nFLH MODIS-A [Chla] retrievals. We see that for the same off shore area A, Figure 11c, R² = 0.72. This is lower than the R² = 0.87 obtained for the equivalent comparison of nFLH against MODIS-A NN [Chla] retrievals, Figure 7a, above. When the comparison is not restricted to the off-shore areas, the correlation for RGCI-nFLH drops off completely to R² = 0.0 (not shown) compared to the previously given equivalent correlations for nFLH-MODIS-A NN of R² = 0.19, Figure 8 and R² = 0.07 for OCI/OC3 (not shown).

When the F1 and F2 filtering procedures are applied to the RGCI [Chla] retrievals, for area A, the correlation against nFLH MODIS-A [Chla] retrievals drops to R² = 0.50, Figure 11d. This compares to R² = 0.71 and R² = 0.50 for the equivalent correlations, shown previously, between nFLH and MODIS-A NN and MODIS-A OCI/OC3 respectively.

The salient points of Section 3.2.1, Section 3.2.2, Section 3.2.3 and Section 3.2.4 show that in open ocean area A, nFLH [Chla] retrievals from MODIS-A are well correlated with MODIS-A NN [Chla] retrievals R² = 0.87, less so with equivalent MODIS-A OCI/OC3 with R² = 0.75, and MODIS-A RGCI with R² = 0.72. All correlations drop when filters F1 and F2 are applied to obtain KB HABs retrievals to R² = 0.71, R² = 0.5 and R² = 0.5, respectively. For comparisons not restricted to the off-shore area A, all correlations with nFLH MODIS-A [Chla] retrievals drop off drastically to R² = 0.19, R² = 0.07 and R² = 0.0 for MODIS-A NN, MODIS-A OCI/OC3 and RGCI respectively. It should also be noted that all four retrieval techniques compared show qualitatively similar KB HABs retrieval images of the type shown in Figure 5a,b, Figure 6b, and Figure 9 (not shown for OCI/OC3 and RGCI).

3.3. Comparisons of VIIRS NN, OCI/OC3 and RGCI Retrievals with in Situ Cell Count Measurements

3.3.1. Evaluation of VIIRS KB HABs Retrievals against in Situ Measurement for Match-Ups Occurring over 2012–2015 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.

It is, however, difficult to find on any one day suffient matches between satellite observations between concurrent, or near concurrent in situ measurements to obtain statistically meaningful results, We therefore extended the study period to look at all available WFS match-ups, Figure 12, between VIIRS measurements and in situ data at concurrent dates and over the 2012–2015 period for which there was available VIIRS data. The conditions for match-up were: (1) pixel centers were 0.3 miles or less from the in situ measurement location; pixels are excluded from the match-up 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, viewing angle greater than 60, and solar zenith larger than 70, and any pixels which has water leaving radiance spectra with negative values in any one of its wavelengths. Cell counts sample measurements also had to be at less than 1 meter depth, and at concentrations ≥10⁴ cells·L⁻¹_. It should be noted that [Chla] of 1 µg·L⁻¹ is ~10⁵ cells·L⁻¹ [49].

The in situ cell count data was obtained from the Florida Fish and Wildlife Conservation Commission’s Fish and Wildlife Research Institute (FWC-FWRI). The search for match-up between VIIRS satellite and in situ observations on the same day showed 94 cases which satisfied the match-up conditions specified above. Figure 13a, above, shows the relationship between NN retrieved a_ph443 and the in situ KB cell measurements for these 94 cases. This shows a correlation with R² = 0.32. Converting the a_ph443 [16] to equivalent [Chla] for these same 94 cases, shows, in Figure 13b, the same correlation against KB cell counts, with R² = 0.32 for both these match-ups. Figure 13c,d, show the equivalent match-up for the same 94 points for retrieved OCI/OC3 and RGCI [Chla] against in situ KB cell counts, which shows lower correlations with R² = 0.28 and R² = 0.17 respectively.

Now, it is recognized that temporal variations due to bloom drift etc. between satellite observation and in situ measurement times can affect these results. Accordingly, the results were separated in accordance with observation time windows. Those occurring with less than 1 h, and 0.5 h between satellite and in situ observations are shown separately. Figure 14a (1 h), b (0.5 h) show the match-ups between VIIRS NN retrieved equivalent [Chla] against KB cell counts. The equivalent values for OCI/OC3 and RGCI [Chla] retrievals are shown in Figure 14c–f respectively. For the 1 h window, Figure 14a, the correlation for the NN retrievals increases from the daily R² = 0.32 to R² = 0.69. For OCI/OC3 retrievals, it increases from R² = 0.28 to R² = 0.33. For RGCI it increases from R² = 0.17 to R² = 0.62. For the 0.5 h window, the correlations increase to R² = 0.82, R² = 0.44 and R² = 0.7 respectively. With the NN being the highest, closely followed by RGCI, then OCI/OC3, though the average errors appear to be twice as high for RGCI as compared with NN.

Thus, it is seen that reducing the time window between satellite and in situ observations can generally significantly increase the correlation between VIIRS retrieved [Chla] and KB cell counts. More area specific studies need to be carried out before definitive conclusions can be drawn.

We also use the above match-ups to assess the validity of the F1 and F2 filter limits for Rrs551 ≤ 6.0 × 10⁻³ sr⁻¹ and a_ph443 ≥ 0.061 m⁻³ defined in Section 3.1 above. Figure 15 shows that for the NN retrieved a_ph443 match-ups falling within the ½ h window, which were retrieved without any filter constraints being applied on Rrs551 and a_ph443 all but two of them fall within these limits. Further refinements on these limiting filter values as a means of retrieving KB HABs will require planned future statistical analysis of false positives and negatives in retrieval results against in situ measurements.

Table 1 below shows summary table of retrieval comparisons, in situ results and KB thresholds criteria applied.

3.3.2. Evaluation of NN KB HABs Retrievals for Specific Bloom Events

We also compared the ability of NN KB HABs retrievals against in situ KB HABs measurements for specific algal blooms. We apply filter processes F1 and F2 above to VIIRS NN a_ph443 retrievals to arrive at KB compatible values, in Figure 16 and Figure 17a,b for three bloom dates: 28 August 2014, 16 November 2014 and 9 October 2012. The residual a_ph443 (left hand scale) and equivalent [Chla] values (right hand scale) are those compatible with KB HABs. Figure 16 shows the retrievals for the 8 August 2014, bloom now overlaid with actual KB cell counts for that date only (not for a range of overlapping dates as was used previously). There are a total of 20 match-ups on that day. There appears to be one false positive, indicated by the arrow. This presents an ambiguous situation, since it is near the edge of the bloom and changes may have occurred in the time window between in situ and satellite measurement, and secondly, while there is no in situ bloom measured at the surface at that date, there was one below the surface at nearly the same location, bringing into play the possibility of sub-pixel variations, indistinguishable when pixel averaging takes place. Figure 17a shows a bloom on 16 November 2014, near the coast, with a total of 6 match-ups. There are no false positives or negatives in this retrieval. This is particularly interesting since it is close to the coast, and therefore in more complex waters. Figure 17b shows a bloom on 9 October 2012, with 12 match-ups. It covers both near and offshore areas and has no false positives or negatives.

4. Summary and Conclusions

Detection of KB HABs in the WFS have been aided in the past by use of the 678 nm MODIS-A fluorescence band for retrievals using nFLH and RBD techniques. The lack of a fluorescence channel on the successor VIIRS satellite provided the impetus for the NN work reported here. In this approach, NN algorithms using Rrs values from the 486, 551 and 671 nm VIIRS bands are used to retrieve an image of a_ph443 values in the WFS. Then, as detailed above, the additional limiting constraints are applied, in two filter processes, F1 and F2 to eliminate from that image all a_ph443 pixels which are not compatible with the existence of KB HABs. These limiting criteria are defined by (i) low backscatter manifested as a maximum VIIRS Rrs551max value [3,17,18] and (ii) the minimum [Chla] and hence equivalent minimum (a_ph443) values that are known to be associated with KB HABs in the WFS. Thus, through the implementation of filter processes F1 and F2, all pixels in the VIIRS NN retrieved a_ph443 image that are incompatible with KB HABs are eliminated. The residual image then shows only retrieved a_ph443 values and their equivalent [Chla] values that are consistent with the existence of KB HABs. The specific limiting values for maximum Rrs551 and minimum a_ph443 used in the filter processes were based on published data, reinforced and refined by inspections of retrieval results against available in situ data. The retrieved values of a_ph443 and equivalent [Chla] retrieved using these filter limits, were then compared both qualitatively and quantitatively both against other satellite retrieval techniques and in situ measurements.

Comparisons of the NN technique were made first with the nFLH and RBD techniques, which, in reported comprehensive tests [3] exhibited the best statistics for KB HABs retrievals in the WFS from MODIS-A. Our comparisons showed good correlations between nFLH and the NN retrievals for open ocean waters. However, KB retrieval capabilities with the nFLH technique are severely limited for in-shore retrievals, where the more complex waters with increased non-algal particulate scattering interfere with the fluorescence signal making its interpretation more difficult [56]. Comparisons were also made between the NN technique and retrievals using OCI/OC3 and the RGCI for [Chla] retrievals. Again, these comparisons showed reasonably correlations in open ocean waters, but confirmed difficulties with retrievals in more complex inshore waters. Further work is also needed on our NN KB retrieval technique before we can properly evaluate its capabilities for in-shore retrieval and compare them more rigorously with those of MODIS-A nFLH, OCI/OC3 and RGCI.

Finally, the critical test for the viability of KB HABs satellite retrieval techniques is the comparison of their ability to match retrieved values with concurrent in situ measurements. We therefore sought to extend the range of available in situ comparisons. This was done by seeking all available match-ups between VIIRS NN a_ph443 (and equivalent [Chla] retrievals) and in situ KB cell count measurements for the period 2012–2015. These comparisons showed that when the window between in situ observations and satellite measurements was reduced from daily, to 1 h and 0.5 h, the correlations greatly improved for the VIIRS NN and VIIRS RGCI retrievals against in situ match-ups. These showed respectable R² = 0.82 and R² = 0.70 respectively for the 0.5 h window. The comparisons also supported the choice of values for the F1 and F2 filter processes. Comparisons were also made comparing VIIRS NN retrievals identifying KB HABs, and in situ KB cell counts measured on the same date for 3 specific bloom events. All three confirmed the viability of the VIIRS NN retrieval approach, with no apparent false negatives or positives. Comparisons with other NN retrieval techniques are not so straightforward. Several of the previously reported NN techniques have used in situ measurements of IOPs, including sea surface temperatures and other physical parameters as inputs to NNs for bloom prediction, apparently with some success [57,58,59]. The more directly comparable NN technique reported, is the NN algorithm product for [Chla] retrievals in Case 2 waters [60,61,62,63] from the MERIS satellite, which is no longer operational. This appears to give comparable results, in European waters, to OCI/OC3 retrievals in open waters, and possibly higher accuracies in coastal waters [61]. This MERIS algorithm does not appear to have been tested for KB retrievals in the WFS. Tests of historical data for such retrievals may yield interesting results.

In conclusion, results to date with the VIIRS NN technique appear promising. However, more match-ups, including more detailed considerations of observation times and their connections to false positives or negatives, sample depths, distance to observation pixel centers and sub-pixel variability and impact of complex in-shore waters need to be studied. More comprehensive statistics need to be obtained on the overall efficacy of the approach and its comparisons to other techniques.