Operational continental scale monitoring and baseline determination of water quality are fundamental for reporting on marine and inland water Sustainable Development Goals (SDG). Surface water is an important global resource and plays an essential role in biochemical cycling, maintenance of biodiversity, human wellbeing, and prosperity [1
]. In the Australian context, existing water quality information is often limited, and inland water, in particular, is not well represented in global datasets or even nationally. Incompatibilities in the content and scale of state and territory environment reports confound attempts to integrate water resource data into a consistent national reporting framework [3
]. In addition, existing monitoring practices do not encompass the breadth of temporal and spatial scales required for standardized baseline establishment (SDG indicator 14.1.1) or reporting on continental scale indicators (SDG indicator 6.3.2) germane to human health, environmental sustainability, and economic prosperity.
Earth observation (EO) methods have the potential to provide a consistent reporting mechanism, providing synoptic coverage of spatial and temporal variations in water quality [2
]. Water quality parameters that can be derived from EO data include chlorophyll and phycocyanin concentrations [6
], coloured dissolved organic matter (CDOM) [9
], the concentration of total suspended solids (TSS) [10
], and descriptors of the light environment of the water column, such as Secchi disk depth (SDz) [7
], vertical attenuation (Kd), water clarity [12
], and water colour [14
], among others. However, their application at a continental scale is restricted by a lack of regional data for parameterization and validation of water quality information [4
Quantitative use of EO data for the synoptic assessment of aquatic water quality has increased in recent times. Retrieval of water quality information from EO data can provide an improved understanding of the spatial and temporal variability within water bodies for the resource managers by filling temporal gaps between periodic in situ observations. Freely available, medium resolution EO data, such as from the Landsat suite of sensors and Sentinel-2 provide a useful platform in which to investigate temporal changes in water quality across wide geographic scales, e.g., [11
Methods implemented to retrieve water quality parameters are strongly determined by the scale under consideration (local, regional, or global). At the local scale, a locally tuned empirical algorithm can be successfully implemented to retrieve water quality data. Due to spatial variability in optically active water column constituents, empirical tuning using geographically limited, in situ data fail at the regional or global scale in areas with unique optical properties [22
] or where the optical complexity of extreme events such as floods, blackwater events, or algal blooms are not captured in the empirical model [4
]. For example, the SeaWiFS global chlorophyll algorithm (OC4) failed to accurately retrieve chlorophyll from satellite data in Antarctic waters [23
] and sediment plumes off the Santa Barbara Channel [24
The application of physics-based (analytical) methods to retrieve environmental variables from EO data offers a solution across a range of scales in optically complex aquatic regions. This approach is strongly driven by an understanding of the relationship between inherent optical properties (IOPs) and the water-leaving radiative signal, which is ultimately detected by the satellite borne sensor. Several quasi-analytical algorithms have been tested and specifically tuned to local water bodies, e.g., Tokyo Bay [25
], Venice Lagoon [26
], and Lake Constance [27
]. Physics-based algorithms have also been implemented on a more regional basis in the English Channel and French Guiana [28
], coastal Australia [29
], the Adriatic Sea [31
], and Florida and the Arabian Sea [32
Optical water types (OWT) is one method implemented to facilitate the retrieval of water column constituent data. OWTs are defined as different water masses that are represented by a collection of similar optical characteristics of the water components, resulting in similar reflectances [33
]. Classifications based on OWTs can provide information about the concentration of optically significant constituents in each class, offering the possibility to observe the trends and status of important biogeochemical and biological variables. OWTs can be implemented to empirically retrieve water column constituents, e.g., [31
], train neural networks, e.g., [35
] or parameterize bio optical models, as suggested by [34
]. This knowledge can then be implemented into monitoring strategies that support the assessment metrics of the UN SDGs related to aquatic ecosystems, such as water quality retrieval, algal bloom detection, estimation of primary production, refining the distributions of biogeographic provinces, and developing indices of marine biodiversity.
Various methodologies were developed in aquatic remote sensing studies to classify OWTs. Most often, some form of cluster analysis is implemented in the process. Cluster analysis groups water into sets based upon differences in the magnitude and shape of reflectance using different degrees of implicit or explicit knowledge [1
]. A selection of recent studies, classifying reflectance data of water into generalized OWTs on scales ranging from local to global is presented in Table 1
Coastal and inland waters are generally considered optically complex ‘Case 2′ waters [6
]. Therefore, the observed reflectance spectra from different Case 2 waters share common features, as their optics are a function of a high variability in the concentrations of different optically active constituents [38
]. Due to the range of bio-optical properties in these zones, classifying these waters is a challenge [33
]. In addition, atmospheric correction algorithm failures [40
] and the high spatio-temporal variability in hydrodynamics and biological processes [38
] makes the classification of these waters more arduous. Due to this complexity, much effort has been focused on classification of coastal and inland waters.
Selected literature on optical water type classification.
Selected literature on optical water type classification.
|Spectral Data||Classification ||Dataset||Reference|
|SeaWiFS data ||ED a|
|Normalized SeaWiFS data ||ISODATA b||Global|||
|In situ Rrs c ||hierarchical||English Channel|||
|In situ Rrs||FCM d||Global|||
|In situ Rrs||FCM d||Chinese lakes|||
|In situ Rrs||FCM d||Global lakes|||
|In situ Rrs||thresholding||Yellow Sea|||
|Normalized In situ Rrs||hierarchical||Eastern English Channel|
|Normalized In situ Rrs ||k-means||Global|||
|Simulated rrs e||WFCM f ||Estonia|
Several studies have successfully implemented optical classification techniques to formulate OWTs at a global scale (e.g., [1
]). However, apart from incorporating data from three Southern African reservoirs [35
], most of these global studies used data collected predominantly in the Northern hemisphere. Most of the regional studies also report on areas in the Northern Hemisphere (Table 1
). This study aims to improve our knowledge of the optical complexity of Australian coastal and continental waters. Although several regional datasets have been reported on, e.g., [49
], this will be the first attempt at analysing the data at a continental scale in Australia. To this end, a database of observations, capturing a range of optical data based on seasonality and geography, was used to:
Develop a method to define distinct OWTs.
Create a set of synthetic generalized inherent optical properties (GIOPs), based on the key features of each unique OWT.
Present a case study as an example of a potential application of implementing the GIOPs water quality monitoring at a drainage basin scale.
Concurrent in situ AOP measurements for 73 of the 110 inland data points shows a relatively good comparison between model derived and measured Rrs
). Linear percentage error [28
] indicates wavelength-dependent differences, introduced by the confounding effects of environmental factors, such as atmospheric variability, the water surface state (with swell-, wave-, and wavelet-induced reflections), and refraction of diffuse and direct sky and sunlight [55
]. Errors are larger where the signal is smallest, e.g., below 450 nm and above 700 nm. The agreement between modelled and measured Rrs
gave us confidence to use the modelled spectra to analyse spectral variability on Australian waters.
Spectral clustering resulted in 32 and 17 unique clusters for the coastal and inland waters datasets, respectively. The results of the analysis of variance test applied to the SIOP data for each cluster are listed in Table 3
. The test showed that the spectral slope constant of NAP absorption coefficient (γ aNAP) contributed significantly to spectral variability within both the coastal and inland waters dataset while chlorophyll concentration contributed significantly to spectral variability within the coastal dataset and NAP concentration contributed significantly to spectral variability within the inland dataset. The absorption of CDOM at 440 nm (aCDOM
(440 nm)) had a strongly significant effect on spectral variability within the coastal dataset and a lesser effect on spectral variability within the inland dataset.
Following clustering and merging, most of the modelled spectra and modelled SIOP sets were clustered in five groups within the coastal dataset (240 of the 286 observations) and six groups within the inland dataset (102 of the 110 observations). The remaining 54 spectra represents small clusters (n < 4) or individual in situ observations that were too distinct, both spectrally and bio-optically, to be grouped with any of the bigger clusters or each other. Figure 4
shows the variability in optically active water column constituents (a–c) and SIOPs (d–h) of the 11 clusters that emerged from the data analysis. Figure 5
and Figure 6
shows the variability in the concentration, absorption and scattering characteristics of the optically active water column constituents of the 11 clusters. A summary of the main features of the 11 clusters is presented in Table 4
captures the variability in spectral shape of each of the observations that describe the 11 OWTs in the dataset. There is a clear distinction between the clusters that made up the coastal dataset (01–05) and those of the inland dataset (06–11). The coastal spectra tend to be dominated by response curves that peaks in the blue region of the spectrum while the inland spectral responses peak predominantly within the green and red region of the spectrum.
shows the spatial distribution of 885 locations with relatively permanent water bodies of sufficient size to not be affected by adjacency effects. The colour scale indicates the number of Sentinel-2 MSI observations retrieved from each point, while the three major drainage regions in the AOI are delineated in shades of grey. There were generally fewer observations on the western and northern sides of the MDB and in the far north of the NEC.
To determine to what extent the nSSM could match Sentinel-2 MSI surface reflectance data to the 11 OWTs, each convolved Rrs
spectrum was compared to the cluster medians and ranges and assigned to the cluster that yielded the lowest nSSM metric. Ideally, the input spectra would be matched back to the original groupings that they were clustered to. However, with the reduced number of spectral bands implemented for the Sentinel-2 MSI spectra (compared to the 90 spectral bands used for the cluster analysis), some of the finer spectral differences between the clusters will be degraded. Table 5
shows an overall classification accuracy of 69% with a Kappa of 62%. Some clusters are more uniquely separable than others (e.g., 11). A few of the clusters with smaller sample sizes are more often mis-classified (e.g., 05, 07, 08, and 10). It should be noted that 08, 07, and 10 are comprised of smaller clusters that were bio-optically similar, thus compromising the spectral uniqueness of the groupings. Within the clusters with a higher number of observations, generally data belonging to the coastal dataset is most often misclassified into another coastal OWT. OWT 06, representing deep, clear inland lakes is often confused with coastal clusters, representing clear coastal waters. Cluster 09 appears to not be spectrally unique with misclassifications into several other classes.
Of the 69% of the convolved Rrs spectra that were correctly matched to the original OWT clusters, 20% were matched with an nSSM value of 1.0 or larger. 40% of the convolved Rrs spectra that were incorrectly matched to the original OWT clusters had an nSSM value exceeding 1.0. This value was selected as the maximum threshold where matches with a larger nSSM value were considered unclassified.
The surface reflectance of the Sentinel-2 MSI pixel observations was compared to each convolved modelled spectrum in each of the 11 clusters and assigned to the cluster that yielded the lowest nSSM metric. An nSSM threshold was defined by determining the lowest nSSM where most of the modelled OWT spectra were assigned to their original cluster. Any match greater than 1.00 was considered spectrally distinct from all the existing OWTs and labelled “unclassified”.
and Figure 10
show the limnological, landscape, and temporal distribution of OWTs matched to Sentinel-2 MSI observations over each waterbody within the three drainage regions. The observations, which were not matched with the existing OWTs, are labelled a red colour. There appear to be more unclassified observations in the inland lakes and rivers class than in the coastal waters (Figure 9
The inland lake and river classes show a clear limnological split east and west of the Great Dividing Range (Figure 9
). Waterbodies in the west (MDB) have a larger proportion of observations with dominant spectral responses in the green and red part of the spectrum and waterbodies in the east (NEC and SEC) have a higher proportion of observations with dominant spectral responses in the blue part of the spectrum.
The MDB is a predominantly inland drainage basin with a wide range of climatic and geohydrological conditions (Table A3
in Appendix A
). There is a strong bimodal clustering of OWTs within the lakes and rivers class in this region (Figure 9
) The observations in the lakes in the lower elevations (<200m) falls predominantly in OWT classes where absorption is CDOM and NAP dominated with high amounts of suspended sediment (11) or chlorophyll (08). Lakes in higher elevations generally have clearer waters with a spectral response dominated by blue reflectance (02, 03, 06). Similarly, rivers in the lower elevations are more dominated by OWT classes with higher NAP and chlorophyll concentrations, while the higher elevations are predominantly clearer waters that are dominated by CDOM absorption (Figure 9
). There appears to be a seasonal shift in water quality across the MDB drainage region with the frequency of NAP dominated OWTs decreasing during the winter months (June, July, August), and with an increase in the observation frequency of clearer OWTs during this period (Figure 10
). The lake class in the higher elevations has the most unclassified observations in this drainage region.
The SEC has a temperate climate with predominantly moderate summer rainfall (Table A3
), resulting in an increase in the observation frequency of more sediment dominated water types in the summer months (December, January, February, Figure 10
). It has many coastal waterways, characterized by estuarine waters (Figure 9
). The lakes in this region are predominantly clear while many of the rivers in the lower elevations have water that is dominated by suspended sediments (Figure 8
). This region has relatively similar proportions of unclassified observations within the coastal, lake, and river classes. The rivers class in the lower elevations has the most unclassified observations within this class in this drainage region.
The monsoonal climate of the NEC (Table A3
) results in a lower observation frequency in the summer months when conditions are cloudy (Figure 9
). This region comprises of numerous drainage basins with predominantly natural waterways in the north. The southern drainage basins have more man-made water storages. There is also a climatological gradient with wet tropical conditions in the north and drier savannas towards the south and west. The coastal waterways in this region is characterized by estuarine waters (Figure 9
). The lakes and rivers in this region are predominantly clear during the observation period while some rivers in the lower elevations have water that is dominated by suspended sediments (Figure 9
). This region has relatively similar proportions of unclassified observations within the lake and river classes, while the coastal waterways have a relatively low proportion of unclassified observations.
The objective of this study was to improve our knowledge of the optical complexity of Australian waters. A workflow was developed to cluster the modelled spectral response of a range of in situ bio-optical observations collected in Australian coastal and continental waters.
Coastal and inland Case 2 waters share common features and represents a continuum from fresh waters in catchments to coastal marine environments [1
]. Despite this continuum, there is a clear distinction between the clusters that made up the coastal dataset and those of the inland dataset in this study. Similar to a global study by [1
], the coastal spectra tend to be dominated by response curves that peaks in the blue region of the spectrum, while the inland spectral responses peaks predominantly within the green and red region of the spectrum.
The coastal clusters represent Australian coastal waters ranging from clear blue to more turbid estuarine waters. The first three coastal OWTs (01, 02, 03) have a strong blue reflectance signal, relative to the rest of the spectrum with absorption budgets dominated by CDOM absorption. The 01 OWT most closely resembles open ocean waters, with NAP absorption affecting the absorption budget to a small degree and very low inputs from phytoplankton. The 02 OWT represents coastal waters with a strong estuarine influence. It has the lowest γaNAP, suggesting a higher proportion of large organic particles [79
]. The 02 OWT closely resembles the 01 OWT but has a higher phytoplankton component. The remaining two coastal OWTs (04, 05) represents moderately reflective tropical waters [49
]. The 04 OWT is dominated by phytoplankton absorption and has relatively high chlorophyll content, suggesting that this represents eutrophic lagoonal waters. The higher aCDOM
(440 nm) and smaller γaCDOM of the OWT indicates a coral reef matrix influence [49
]. The higher NAP concentration and large γaNAP of the 05 OWT implies that this OWT represents tropical waters with a stronger coastal influence than 04.
The inland clusters represent eastern Australian continental waters ranging from relatively clear alpine reservoirs to turbid shallow lakes. All these OWTs have a maximum reflectance peak around 570–585 nm. The first inland OWT (06) has relatively low reflectivity and low concentrations of water column constituents, denoting clear waters with high transparency. The two green inland OWTs (07, 08) are both highly reflective water types with high chlorophyll concentrations. They are spectrally very similar with spectral differences mainly limited to a difference in the 705 nm reflectance peak and a lower reflectance in the blue region by 08, characteristic of the stronger CDOM absorption in this OWT. The 07 OWT has a smaller median γ bb
NAP and a larger median γ aNAP than 08, suggesting a predominance of larger organic particles in the water column [79
]. The last three OWTs represents waters that are dominated by NAP and CDOM absorption with a relatively small phytoplankton component. The 09 OWT has a moderate reflectivity and represents fairly clear inland waters. The comparatively high NAP absorption and low b*b
NAP suggests a water column that is characterized by small suspended particles. Both the 10 and 11 OWTs are highly reflective sediment-laden water types with high NAP concentrations and high CDOM absorption. The large γaNAP of 10 suggests larger amounts of organic particulate material [79
] while the relatively low b*b
NAP of the 11 OWT suggests waters with smaller suspended particles [82
To test the applicability of the 11 derived OWTs for water quality monitoring at a drainage basin scale, temporal analysis was undertaken on Sentinel-2 MSI surface reflectance observations. A total of 885 sample locations across three drainage regions were used in the case study. The satellite-derived surface reflectance observations, sourced from DEA, captured a wide range of climatological and limnological conditions. The analysis demonstrated clear limnological and seasonal differences in water types within and between the drainage divisions, which corresponds with general limnological, topographical, and climatological factors. Due to the operational period of the Sentinel-2 satellite platforms coinciding with severe drought conditions [83
], there were generally fewer observations on the western and northern sides of the MDB, as surface water extents were significantly reduced. There were also fewer observations collected over this region before the Sentinel-2 platform was fully operationalized, which further confounded the analysis. Less observations in the far north of the NEC are predominantly due to cloud cover. As much of this region is subjected to monsoonal conditions in summer [84
], there are larger proportions of observations made in winter, which will affect any analysis of seasonal water quality trends from satellite data in this region.
The DEA Landsat archive, representing more than 30 years of ARD observations, is ideal to capture natural water quality variations caused by seasonal, limnological, and climatological factors as well as anthropogenic factors, such as land-use change [85
]. However, the atmospheric correction protocol implemented on DEA satellite data is based on standard terrestrial aerosol climatology model parameters. Atmospheric correction over water bodies is challenging because waterbodies are dark, and about 90% of the signal received by the sensor is not caused by the water itself [86
]. The large positive bias in the shape of the reflectance in the blue bands over water targets, which is introduced by this model [73
], renders the Landsat data unsuitable for the purpose of this study. With future reprocessing, incorporating a maritime aerosol type, in combination with the Landsat-8 top of atmosphere (TOA) reflectance-calibrated product and a glint-correction algorithm [73
], the spectral shape of the blue bands will be more suitable for aquatic applications on DEA datasets.
Although the Sentinel-2 MSI data on DEA are corrected with the same atmospheric correction parameterization as the Landsat data, the increased number of spectral bands and finer spatial resolution somewhat alleviates the existing issues with the spectral shape of the surface reflectance spectra. To further limit the confounding effects of the existing atmospheric correction protocol on the spectral shape of aquatic targets, the coastal blue band of the Sentinel-2 MSI sensor was excluded from the analysis.
The optical cluster analysis was carried out on hyperspectral modelled Rrs
spectra whilst the spectral matching of the case study was applied to five multispectral bands. The reduced spectral coverage of multispectral sensors limits the separability of water types because some water column constituents have characteristic absorption features in narrow sections of the spectrum [88
]. The reduced number of spectral bands implemented for the Sentinel-2 MSI spectral analysis degraded some of the finer spectral differences between the clusters. Due to the limited spectral resolution of the Sentinel-2 satellite data, the spectral signatures of some of the OWTs in this study became less uniquely separable than others. To mitigate the confounding effect of the reduced spectral resolution of the Sentinel-2 bands, a threshold metric was defined. An optimal threshold is a compromise between maximizing the probability of a correct match and minimizing the rate of false positive matches [89
]. The threshold defined for the nSSM metric balances the requirement to capture the maximum number of correct spectral matches with the need to identify areas where there are genuine gaps in the existing water quality database. To limit the confounding effect of environmental factors on the accuracy of results, care was taken to select only observations that were not affected by atmospheric variability, sun glint, and the effects of the brighter response of adjacent terrestrial targets [56
The OWT classification of waterbodies demonstrates clear limnological and seasonal differences in water types within and between the drainage divisions, which corresponds with general limnological, topographical, and climatological factors in those regions [84
]. Unclassified observations were clustered predominantly in lakes within the alpine regions and east of the Great Dividing Range. Observations within rivers (flowing water) are less frequently labelled unclassified. Targets with a higher albedo were more often matched to an OWT, suggesting that some of the unclassified observations in the deep, clear lakes in the alpine regions [84
] may be due to low target albedo. The water leaving signal at the satellite sensor is a small part of the total measured signal [86
]; this potentially constrains the ability of the Sentinel-2 sensor to record a sufficient amount of water leaving light through a set of atmospheric and air water interface conditions to allow a clearly distinguishable spectral signature that can be matched to an existing OWT.