Evaluation of MODIS Products in the Great Australian Bight and Adjacent Coastal Waters

Satellite remote sensing data can produce global environmental data and is easily accessible and widely used by the scientific and non-scientific community. However, to use satellite data, it is important to know its limitations and how it validates against in situ measurements for the different regions. Here, field measurements of chlorophyll-a concentration and euphotic depth within the Great Australian Bight, Gulf St Vincent and Spencer Gulf were used to validate ocean colour products derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the Aqua satellite. The field data include in situ and in vivo chlorophyll-a concentration, which were compared against MODIS chlorophyll-a products derived from three algorithms (OC3M, Carder, and Garver-Siegel-Maritorena (GSM)), as well as euphotic depth measurements derived from photosynthetically active radiation (PAR) profiles, which were compared against two MODIS euphotic depth products (derived semi-analytically and from surface chlorophyll-a). The OC3M product performed well in open waters, with errors below the 35% NASA accepted limit, but it overestimated chlorophyll-a values in shallow (<50 m) waters. The GSM product produced the lowest errors, but also showed a smaller dynamic range, while the Carder product produced higher errors than GSM and it also showed small dynamic range. The relationships between the MODIS and in situ euphotic depth were robust, with errors lower than 20%. MODIS products showed weaker or no significant relationships to in situ measurements in the Eastern Great Australian Bight. This is thought to be due to the summertime subsurface upwelling pool that is characteristic of the area. Based on these results, the OC3M product provides the most reliable estimates of chlorophyll-a, and is recommended for further applications of MODIS imagery, if the limitations in shallow waters are taken into account. Alternatively, the GSM product could be a better option if the algorithm were locally adjusted. Changes in the sampling methodology to imHow to cite this paper: Rodriguez, A.R., King, E., Doubell, M. and van Ruth, P. (2018) Evaluation of MODIS Products in the Great Australian Bight and Adjacent Coastal Waters. Open Journal of Marine Science, 8, 431-449. https://doi.org/10.4236/ojms.2018.84024 Received: September 25, 2018 Accepted: October 23, 2018 Published: October 26, 2018 Copyright © 2018 by authors and Scientific Research Publishing Inc. This work is licensed under the Creative Commons Attribution International License (CC BY 4.0). http://creativecommons.org/licenses/by/4.0/ Open Access A. R. Rodriguez et al. DOI: 10.4236/ojms.2018.84024 432 Open Journal of Marine Science prove the algorithms are discussed. Derived euphotic depth products can be used with confidence in applying MODIS products for monitoring water clarity, ecosystem health or primary productivity in the region.

in situ measurements in the Eastern Great Australian Bight.This is thought to be due to the summertime subsurface upwelling pool that is characteristic of the area.Based on these results, the OC3M product provides the most reliable estimates of chlorophyll-a, and is recommended for further applications of MODIS imagery, if the limitations in shallow waters are taken into account.Alternatively, the GSM product could be a better option if the algorithm were locally adjusted.Changes in the sampling methodology to im-

Introduction
Satellite systems are able to obtain global environmental data with great temporal coverage, and have provided important insights into marine ecosystem dynamics [1] [2] [3] [4].Nowadays, satellite data is easily accessible from a number of institutions and is widely used not only by the scientific community but also by the general public.In some areas, ocean color-based chlorophyll-a concentration (chl) and sea surface temperature data are used operationally to indicate potential fishing and conservations zones [5] [6].In addition, measurements of ocean color-based euphotic depth (Z eu ) are essential for modeling primary productivity within the ocean [7].However, satellite images must be calibrated with in situ measurements to ensure the quality of the data for different biomes [8] [9] [10] [11].
The processing of satellite ocean color data requires an atmospheric correction and the development of in-water bio-optical algorithms.Phytoplankton is the main optically active component of open marine waters, where changes in the optical properties are directly related to the chl.These waters are known as Case 1 waters [12] [13].In coastal waters, colored dissolved organic matter (CDOM) and suspended inorganic matter, together with bottom reflectance, can also contribute significantly to changes in optical properties.In these cases, the different optically active constituents need to be characterized to accurately estimate the chl [14] [15].These waters are known as Case 2 waters [16] [17].
Most operational satellite chl products are based on empirical relationships between chl and the reflectance ratio in several spectral bands.These products perform well in open waters [17] [18] [19] but may fail in coastal waters due to the presence of additional optically active substances [20] [21] [22].Semi-analytical ocean color algorithms use a model to obtain characteristics of seawater optical components [23] [24] [25] but are more sensitive to errors in the atmospheric correction than empirically derived algorithms [10].Similarly, the Z eu can be estimated empirically (derived from surface chl measurements) or by determining the vertical distribution of light from the optical properties [23] [26], and therefore is also influenced by water type and atmospheric corrections.
The  [10].Owing to the remoteness of the Great Australia Bight (GAB) region in southern Australia (Figure 1), there has been a paucity of in situ data available for the assessment of satellite data.Previous validation studies in southern Australia have focused on the Spencer Gulf region; in particular, the southern area of the Gulf surrounding Port Lincoln, which supports the lucrative Southern Bluefin Tuna aquaculture industry [27] [28].
The purpose of this study is to evaluate the performance of satellite-based measurements of chl and Z eu using field-based measurements in the GAB, Gulf St Vincent and Spencer Gulf in order to determine whether the imagery is a reliable data source and to identify the most accurate products available for on-going monitoring and various research applications in the region.Satellite products evaluated here include: chl derived from the OC3M [18] [19], Carder [29] [30] and Garver-Siegel-Maritorena (GSM; [24]) algorithms, and Z eu derived from surface chl [31] and semi-analytically [23] from MODIS-Aqua.

Study Area
The area of interest includes the GAB, Gulf St Vincent and Spencer Gulf, extending from 130˚S to 139˚Sand 31˚SE to 37˚E (Figure 1  Coincident with the water sampling, in vivo chl measurements were taken from Conductivity-Temperature-Depth profiler (CTD) casts using either a Chelsea Aqua Tracka III fluorometer (Chelsea Technologies Group, Surrey, UK) or an ECO FL fluorometer (WetLabs, Rhode Island, US) attached to the CTD.
From the fluorescence profiles we extracted the in vivo surface chl values at 5 m depth.In cases where two surface chlvalues (i.e.upcast and downcast) were available, we used an average of the two readings and excluded those values where the logarithmic differences were greater than 0.5.

Euphotic Depth
During the DEPM surveys and SAIMOS project, profiles of photosynthetically active radiation (PAR) in the water column were recorded from CTD casts, us-ing either a Satlantic PAR-LOG-600 m PAR sensor (Satlantic LP, Halifax, Canada) or a Biospherical QSP-200L Log Quantum Scalar Irradiance Sensor (Biospherical Instruments Inc.San Diego, CA) attached to the CTD.From the PAR profiles we calculated the Z eu , defined as the depth where PAR is reduced to 1% of the initial value at the surface.Firstly we calculated the PAR attenuation coefficient as the slope of the regression line of the log-transformed PAR with respect to depth, and then used the standard method to calculate the euphotic depth following the Beer-Lambert equation: Since the PAR profiles could be influenced by the boat shadow, we calculated the Z eu including only PAR values corresponding to depths greater than 5 m.In cases where two PAR profiles (i.e.upcast and downcast) were available, we used an average of the values calculated from the two casts and excluded those were the differences were greater than 5 m.

MODIS Imagery
MODIS Aqua level-2 daily imagery was obtained from the IMOS archive for Australia.These data were derived by processing raw 5-minute granules to level-1b and then level-2 using the NASA SeaDAS [42] processing package (v7.0.1).The level-2 processing step includes a correction for contribution of the atmosphere and results in data products with the native nadir resolution of approximately 1 km 2 .MODIS data were only extracted from the archive for days where the difference between the time of the data collection and the satellite overpass did not exceed 5 hours.We selected the following level-2 products: the chl derived from the OC3M, Carder and GSM algorithms, their corresponding level 2 flags, and the Z eu values.The OC3M product is based on an empirical algorithm: a fourth degree polynomial regression between the pigment concentration and the spectral ratios of ocean reflectance [18] [19].The GSM is a semi-analytical algorithm based on the bio-optical model inversion algorithm of Garver and Siegel [43], and optimized using a simulated annealing technique [24].The Carder algorithm is also semi-empirical, but uses a more complex approach than the GSM, calculating the spectralab sorption properties of the sea water and splitting them into those associated with phytoplankton pigments and those associated with biological degradation products (e.g.CDOM).The absorption coefficient of phytoplankton chl is then adjusted in relation to the chl and the availability of light and nutrients [29] [30].
The Z eu is based on the semi-analytical algorithm of Lee et al. [44], computed from the absorption and backscattering coefficients.In addition, we used the surface chl values (OC3M) to compute the empirical Z eu (Z eu _chl) following the approach of Morel et al. [31] Equation (1).

Comparison between Field-Based and MODIS Data
We compared the measurements obtained via the field-based sampling (both in situ and in vivo chl) with those derived from MODIS for each of the three chl products and the two Z eu values following a modified general match-up exclusion protocol [45].We edited the dataset using the Level 2 processing flags and masked out those pixels where any of the NASA operational Level 3 Ocean Color Processing flags (Table 1) were set.We identified and averaged the 9 pixels closest to the field-based measurement (within an array of 9 by 9 pixels centered on each of the locations).Only match-ups where at least 5 out of the 9 pixels contained valid data were included.
We calculated the relationship between the field-based value and the MODIS value via linear regression, and determined the goodness of fit through the coefficient of determination (R 2 ) and the Root Mean Square Error (RMSE).We tested the data for normality and log-transformed the values prior to the calculations when necessary.

Chlorophyll-a
The relationship between the MODIS OC3M product and the DEPM in situ chl was weak (Table 2; Figure 2(a)), with an R 2 value of 0.32 and a root mean     2).
The Carder and GSM products (Figure 2(b) and Figure 2(c)), designed for Case 2 waters, were not affected by the exclusion of shallow waters from the analysis (Table 2) and their RMSE were within the 35% accepted NASA limit.
Both products still showed squared correlation coefficients to be lower than 0.5.
As per the OC3M, both products showed a better relationship with in situ data in the west GAB (R 2 ~ 0.35) and also around Kangaroo Island (R 2 > 0.5) and no significant relationships in the east GAB (p-value = 0.16).
The method used to calculate the in situ chl from the water samples was not able to resolve small variations in concentrations, leading to an apparent quantisation, or binning, of the field-based values (Figure 2).The limited precision in the methodology may result in apparent errors in MODIS when the datasets are compared.To account for these apparent errors, we repeated the analysis with a second field-based dataset based on in vivo chl values measured using a CTD fitted with a fluorometer (CTD-F) during DEPM surveys [40].Chl data derived from fluorometers should be used with caution, since changes in the phytoplankton population, environmental state, physiological state, quenching, instruments and calibrations may all lead to erroneous or biased measurements [48]- [53].In order to minimize the bias within and between instruments we considered each instrument calibrated within a period of a year as a separate dataset.
The relationships between the MODIS and in vivo chl data are generally stronger than with the in situ data, with R 2 values > 0.5 for all regressions performed, except for the Carder product (Table 3, Figure3).However, most of the MODIS data overestimated the in vivo chl values.Such bias is likely due to the methodology used to measure in vivo chl rather than an overestimation from MODIS, since fluorometers tend to underestimate the chl present in sea water [53].For 2004, the OC3M achieved the best relationship in terms of R 2 values (R 2 = 0.83), however, it also had the greatest RMSE (36%).In contrast, for 2005,  OC3M performed the best on both criteria, with an R 2 and RMSE of 0.86 and 21%, respectively.Differences between the performances of the algorithms may be related to the location of the sampling stations within the years (Figure 3 & Figure 4).During 2004, a number of stations were located in shallow waters within the gulfs, while in 2005 all the stations were located in open waters, with more stations in the west GAB and KI areas.
We repeated the regressions for 2004 excluding locations with depth < 50 m (which excludes the gulfs).This improved the OC3M relationship (RMSE = 29%), while the results for the other two algorithms remained the same.As before, the RMSEs for all the algorithms were lower for the west GAB region.
The relationship between the MODIS and the SAIMOS in situ chl (Table 4) was moderate, with an R 2 value > 0.5 for all products.The number of data points used for the regressions however is very small, and the results are not significant at the 99% level (n < 10, p = 0.02), hence the results should be used with caution.
Regressions between MODIS and the in vivo data from SAIMOS were non-significant and involve a low number of data points (n < 10, p-value > 0.05; not shown).

Euphotic Depth
The relationships between the MODIS and the in situ Z eu from the DEPM dataset were generally very good (Table 5, Figure 4(a) and Figure 4(b)).The Z eu estimated from the surface chl values (R 2 = 0.73, RMSE = 11%) were better than those derived by the Lee (2007) semi-analytical algorithm (R 2 = 0.69, RMSE = 18%), however, the former tended to underestimate the Z eu of the clearest waters.
There were no substantial differences in the performance of both algorithms when disregarding the shallow waters (depth < 50 m).As with the chl values, the correlations between the satellite and in situ datavaried between the different areas of the GAB and the best matches occurred within the west GAB, where the RMSE from both algorithms fell below 10%.
In contrast, the relationships between MODIS and the SAIMOS in situ Z eu (Figure 4(c) and Figure 4(d)) were less robust and the best matches occurred for the chl derived Z eu within the east GAB.Again, there were a smaller number   of data points for this comparison, and the data collected by SAIMOS was not only from the upwelling season (February and March) but across seasons.

Discussion
To guarantee the quality and consistency of remote sensing data, they must be validated against field-based observations [8]  In general, the OC3M chl product performed better in open waters, while it Open Journal of Marine Science tended to overestimate chl in shallow (depth < 50 m) waters.The OC3M was designed primarily for case 1 waters and, hence, the overall poor performance in the gulfs and shallow waters (which are also generally coastal) is not surprising.
Removing data from shallow waters improved the accuracies, with RMSE errors falling below the accepted NASA 35% error limit [46] [47].In contrast, the semi-analytical products based on the model of Carder et al. [29] and GSM, which were designed to have an improved performance in optically complex waters, produced consistent results irrespective of whether shallow waters were included.The GSM semi-analytical algorithm produced the lowest errors, although it produced an average slope of ~0.6, overestimating low of chl and vice versa and it presented a smaller dynamic range.
Based on these results, the OC3M chl product provides the most reliable estimates of chl and is recommended for further applications of MODIS imagery in the region, if the limitations in shallow waters are taken into account.Alternatively, the GSM algorithm could be a better option if the algorithm were locally adjusted.
The correct evaluation of the different satellite products requires accurate field-based measurements; however the methodology used to infer in situ chl was subject to errors (due to limited precision and bias inherent to the methodology).The use of a high-performance liquid chromatography (HPLC) method (currently used by the SAIMOS project) will improve the accuracy of the in situ chl data, while developing strict calibration and quality control protocols for the in vivo data will provide a more consistent dataset [56].
Optical variability across the study environment affected the performance of the three chl products examined here, and further improvements would be necessary before their utilization for an operational monitoring of the GAB shallow waters.All products showed better relationships with the DEPM in situ data within the west GAB and failed to give significant results in the east GAB.Such differences may be due to the fact that the east GAB is characterized by a subsurface chl maximum during the upwelling season (Jan/Feb/Mar; [11] [33] [35] [36] [57]).Field measurements of chl are taken at a discrete location in the water column, while the satellite observations are a weighted mean over the light-penetrating depth of the water column.Therefore, changes in the vertical profile may not be properly recorded by discrete sampling while changes on the vertical profile for deeper (subsurface) areas will not be properly estimate by the satellite.
To generate a locally improved algorithm, sampling efforts should be directed to collecting a comprehensive bio-optical dataset, including in situ remote sensing reflectance and spectral absorption coefficients, that would help to determine whether the performance of the products is related to the effect of the spectral ratios considered or to the determination of the spectral contributions of different seawater constituents [8] [9] [58].
The relationships between the MODIS and in situ Z eu were robust, with RMSE lower than 20%.For the GAB, the Z eu values derived from the surface chl values were better than those derived semi-analytically in terms of RMSE, although they had lower slopes.As with the chl data, the relationships with the in situ data from the sardine DEPM surveys (collected during the upwelling season) in the east GAB were notably weaker.The results of the MODIS Z eu derived from surface chl profiles were robust even in shallow water.The Z eu does not depend on a specific absorption coefficient and, compared to chl, it is much easier and more accurate to determine in the field and less subject to errors due to limited precision and bias inherent to the methodology, thus Z eu produced very small errors compared with the chl data.Our results indicate that derived Z eu could be used with confidence in applying MODIS products for monitoring water clarity, ecosystem health or primary productivity in the region.

Conclusion
We have evaluated the performance of the satellite-based measurements of chl and Z eu against field-based measurements within the GAB and Gulf St Vincent and Spencer Gulf.The performance of the OC3M chl product, although reliable within open waters, is poor within the gulfs, while the GSM produces the least errors but has less dynamic range.The chl algorithms could be re-evaluated and improved to be used with confidence in the GAB area if a comprehensive bio-optical dataset was to be collected.The Z eu showed good performance and it is considered a reliable dataset to be used for future applications within the GAB.

Figure 1 .
Figure 1.Study area and locations of the field stations visited during the sampling campaigns of the sardine daily egg production method surveys (grey circles) and the South Australian Integrated Marine Observing System (black circles).

Level- 2 flags
Atmospheric correction failure Pixel is over land High sun glint Observed radiance very high or saturated High sensor view zenith angle Stray-light contamination is likely Probable cloud or ice contamination Coccolithophores detected High solar zenith Very low water-leaving radiance (cloud shadow) Derived product algorithm failure Navigation quality is reduced Aerosol iterations exceeded max Derived product quality is reduced Atmospheric correction is suspect Bad navigation Open Journal of Marine Science

Figure 2 .
Figure 2. Relationships between MODIS (a) OC3M, (b) Carder and (c) GSM algorithms with in situ chlorophyll-a from the DEPM dataset.The red line corresponds to the fitted line and the black line represents a perfect fit (1:1).The symbols correspond to (d) the different sampling stations from the DEPM surveys that matched MODIS data.Colors represent different bathymetric depths (green > 100 m, blue = 100 -50 m, orange = 50 -30 m, red ≤ 30 m).

Figure 3 .
Figure 3. Relationships between MODIS (a) OC3M, (b) CARDER and (c) GSM algorithms with the -a dataset during 2004 and between MODIS (d) OC3M, (e) CARDER and (f) GSM algorithms with the in vivo chlorophyll-a dataset during 2005.The red line corresponds to the fitted line and the black line represents a perfect fit (1:1).The symbols correspond to the different sampling stations from the DEPM surveys that matched MODIS data during (g) 2004 and (h) 2005.Colors represent different bathymetric depths (green > 100 m, blue = 100 -50 m, orange = 50 -30 m, red ≤ 30 m).

Figure 4 .
Figure 4. Relationships between MODIS euphotic depth based on (a) Lee (2007) algorithm and (b) surface chlorophyll-a concentration with the field-based euphotic depth data from the DEPM project and between MODIS euphotic depth based on (c) Lee (2007) algorithm and (d) surface chlorophyll-a concentrations with the field-based euphotic depth data from the SAIMOS project.The red line corresponds to the fitted line and the black line represents a perfect fit (1:1).The symbols correspond to the different sampling stations from the (e) DEPM and (f) SAIMOS surveys that matched MODIS data.Colors represent different bathymetric depths (green > 100 m, blue = 100 -50 m, orange = 50 -30 m, red ≤ 30 m).

[ 9 ]
[10] [11].Validation exercises have been performed around the world, resulting in improved regional algorithms[4] [9] [54][55].We analyzed the performance of various MODIS chl and Z eu products, using historical field-based datasets collected between 2003 and 2014 in the GAB and gulf regions of southern Australia.To our knowledge, this is the first study presenting a validation of remote sensing Z eu and chl imagery in the GAB area using in situ and in vivo measurements.Previous validation exercises have been carried out in Spencer Gulf in the vicinity of the Port Lincoln aquaculture Tuna Farming Zone[27] [28] for chl and SST, and their findings are consistent with the results shown here for gulf waters.
total of 118 stations, but only 10 matched with the MODIS data.

Table 1 .
Operational Ocean Color Processing masks.

Table 2 .
Regression statistics for MODIS and in situ chlorophyll-a data for the DEPM dataset.
(Figure 2(d)), intermediate in the central GAB, and non-significant in the eastern GAB (Table

Table 3 .
Regression statistics for MODIS and in vivo chlorophyll-a data for the DEPM dataset.

Table 4 .
Regression statistics for MODIS and in situ chlorophyll-a data for the SAIMOS dataset.

Table 5 .
Regression statistics for MODIS and in situ euphotic depth data.