Comparison between simulated Landsat-7,
Landsat-8, Sentinel-2 and Sentinel-3 satellite data
for detecting inland water quality variables
Date: 25/06/2015 Marit van Oostende Student ID: 10187448 University of Amsterdam Supervisor: Emiel van Loon
The water quality of fresh water bodies is important for human health, biodiversity and aquatic ecosystem health. New earth observation satellites currently and to be deployed in the near future have the potential to improve remote sensing for inland waters and will enable continued time series on the OACs; Chlorophyll-α, Coloured dissolved organic matter and Non-algal particulate matter. Higher spectral resolution and careful placement of spectral bands has been shown to improve water quality retrieval whether the sensors used were designed for terrestrial or ocean applications. However, a sensor specifically designed for the monitoring of inland water quality may not be cost effective. By evaluating the water quality retrieval accuracy that can be achieved from reflectance spectra obtained from common satellite sensors, this study aims to identify a cost-effective compromise by
identifying the most suitable sensor for this purpose. The output of this study is a comparison between simulated terrestrial sensors Landsat-7, Landsat-8 and Sentinel-2 and the coastal – ocean sensor Sentinel-3 for retrieving OAC data. We applied this comparison to five different lakes along a temperate to tropical gradient. The spectral inversion method
(algorithm) used was the adaptive linear matrix inversions of forward simulations of spectra performed in EcoLight, a radiative transfer numerical model imbedded in IDL.
Incorporating signal to noise factors of each sensor Sentinel-3 is the best suited to retrieve OAC data. When lakes are considered too small for the 3 sensor pixel size, Sentinel-2, with its smaller pixel size, may be the most useful for this purpose.
Table of contents
1. Introduction...4 2. Theoretical background...7 3. Methods...14 4. Results...16 5. Discussion...22 6. Conclusions...23 Acknowledgements...23 Citations...24List of definitions
AbbreviationsaLMI Adaptive linear matrix inversion
CHL Chlorophyll-α
CDOM Coloured dissolved organic matter
CPC Cyano-phycocyanin
CPE Cyano-phycoerythrin
IOP Inherent optical property LDCM Landsat data continuity mission NAP Non-algal particulates
OAC Optically active constituents SIOP Specific inherent optical propertie TSM Total suspended matter
Parameters
a Total absorption coefficient m-1
a*NAP (440) Specific absorption of NAP at the 440 nm m2g-1
a*phy CHL specific absorption spectrum m2mg-1
aCDOM/a(CDOM) Absorption coefficient of CDOM m-1
aNAP/a(NAP) Absorption coefficient of NAP m-1
aphy/a(CHL) Absorption coefficient of CHL m-1
a(w) Absorption coefficient of pure water m-1
b Total backscattering coefficient m-1
bbNAP Backscattering coefficient of NAP m-1
bbphy Backscattering coefficient of CHL m-1
bb*NAP(550) Specific backscattering of NAP at 550nm m2g-1
bb*phy(550) Specific backscattering of CHL at 550nm m2mg-1
bbphy/b(CHL) Backscattering coefficient of CHL m-1
bbNAP/b(NAP) Backscattering coefficient of NAP m-1
bbw/b(w) Backscattering coefficient of pure water m-1
c Attenuation coefficient m-1
CCHL Concentration of CHL
-CCDOM Concentration of CDOM
-CNAP Concentration of NAP
-Ed Downwelling irradiance W m-1nm-1
Lsky Downwelling radiance from the sky W m-1sr-1nm-1
Lu Total upwelling radiance W m-1sr-1nm-1
rrs Remote sensing reflectance sr-1
s Direction
-SCDOM Spectral slope constant for CDOM absorption coefficient nm-1
SNAP Spectral slope constant for NAP absorption coefficient nm-1
YNAP Power law exponent for NAP backscattering coefficient
-1. Introduction
The quality of inland water bodies is important for consumption, agriculture, fishing, recreation and ecosystems. It is affected by a number of factors including urbanisation, population growth, land use change, deforestation, farming, overexploitation and contamination from industries. Therefore water quality monitoring is essential to observe the condition and discover trends in the water body
constituents. Inland waters are defined in this thesis as inland surface freshwaters. Water quality refers to the physical, chemical and biological content of the water and may vary. It does not describe an absolute but rather a condition relative to the use or purpose of the water. The most important optical water quality variables of inland water are the optically active constituents (OACs); chlorophyll-α (CHL), coloured dissolved organic matter (CDOM), non-algal particulates (NAP) and
cyano-phycocyanin (CPC)[ CITATION Gue15 \l 1043 ]. Inland waters are also referred to as Case 2 waters. Case 1 waters are optically relatively simple waters, where algae and its breakdown products are the OACs, often summarised in the chlorophyll concentration. Case 1 waters are usually oceans. Case 2 waters are more complex, more OACs are relevant than just chlorophyll and they influence each other[ CITATION Dek03 \l 1043 ].
There are three ways in which water quality are usually measured: laboratory analysis, in situ remote sensing1 and earth observation2. Earth observation satellites can provide water quality data on a daily
basis on a large scale, which is not possible with field-based approaches (laboratory analysis and in situ remote sensing) only. Earth observation provides an objective, wide viewing, high frequency and continuous measurement tool. Field- and earth observation measurements can be used to complement and validate each other[ CITATION Gue15 \l 1043 ].
Earth observation satellites measure spectra from space at different wavelengths (spectral bands). These spectra can be used for determining OACs. There are four approaches by which spectral reflectance measurements can be used to estimate concentrations of OACs. First there is the empirical method. Herein statistical relationships are sought between measured spectral values and measured water parameters. This is the least scientific method, as a causal relationship does not necessarily exist between the parameters used. Second is the semi-empirical method; the spectral characteristics of the compounds sought are more or less accurately known. This spectrometric knowledge can be included in the statistical analysis. Reasonable algorithms can be found by common sense and improved by experience. Algorithms that use single bands, band ratios, band arithmetic or multiple bands as independent variables in different regression analyses is a widely used example of the empirical approach. This method suffers from the fact that extrapolation beyond the range of constituents observed may produce erroneous results. Thirdly, there is the analytical method, a difficult method wherein reflectance spectra are simulated using radiative transfer theory and the results cannot be
methods. In this semi-analytical based computer model, OACs can be estimated with an initial input of concentration-specific IOP (SIOP) datasets, concentrations and remote sensing reflectance (Rrs), measured in this case with the TriOS Ramses fieldspectroradiometer. IOP data is measured in this study with the BB9 and ac-s instruments that measure backscattering at nine wavelengths and beam absorption and attenuation in hyperspectral wavelengths.
Current remote sensing of inland waters is limited by the fact that high spectral resolution imagery has a low spatial resolution and vice versa [ CITATION Jul13 \l 3081 ]. New earth observation satellites will be deployed in the near future or have recently been deployed, and have the potential to improve remote sensing for inland freshwaters and have a long lasting impact[ CITATION Dek12 \l 3081 ]. These satellites include the Landsat Data Continuity Mission (LDCM) or Landsat-8, and the Sentinel-2 and Sentinel-3 mission which can be of great value as they give free data access, and can be used for continuous inland water quality monitoring[ CITATION Pal14 \l 3081 ].
The launch of Landsat-8 in February 2013 ensures the continuous stream of satellite data which is essential for monitoring. It has been stated that Landsat-8 data will be comparable to other Landsat records in terms of spatial resolution, swath width, global geographic coverage and spectral coverage on the land cover [ CITATION Iro12 \l 3081 ]. This article by Irons et al., was written before the launch of Landsat-8 and therefore a comparison between Landsat-7 and Landsat-8 data is still required. There is only one study which compares Landsat-7 and Landsat-8 data. This study compares spectral bands with sample points and vegetation indices. However, they do emphasise that more comparison analysis between Landsat-8 and other sensors should be carried out [ CITATION LiP14 \l 3081 ]. It has been proven that data from Landsat 1 to 7 can be used interchangeably to measure and monitor the same landscape phenomena[ CITATION Vog01 \l 3081 ]. The Landsat satellite series is the longest running earth observation mission, operating since 1972.
The ESA’s Sentinel missions, like the LDCM, will provide high resolution optical imagery and continuity of earth observation data collection. As this mission is relatively new, no comparison studies have been conducted between these two or with the LDCM. Further details of the satellite sensors are discussed in the Theoretical background. Hence the question is: Is it possible, in relation to water quality, to compare Landsat-8 and Sentinel data performance and consequently to be able to relate them to older legacy or archival datasets (Landsat 5 & 7) allowing trend analysis across various sensors?
The satellite sensors operate on diverse spectral, spatial and temporal resolutions [ CITATION Roy14 \l 3081 ] [ CITATION Ber12 \l 3081 ]. This can be problematic when different band
placements and sensitivities for certain wavelengths may give different results on OAC concentrations for the same water body. The different spectral resolutions could have serious implications for relating new satellite data to older data of the Landsat series. Successful application of any multi-spectral satellite sensor ultimately depends on the ability of that sensor to adequately describe the shape of the reflectance spectrum and hence relate shape to water quality concentrations. Improving the placement and width of spectral bands leads to stronger correlations with OACs and increasing the number of such bands allows for a greater range of OACs to be retrieved (Dekker, 1993). The spectral resolution of different satellite sensors will therefore determine the ability to discriminate a range of OACs. Sensors with limited spectral resolution will be restricted in their ability to discriminate different OACs and will derive concentrations with less accuracy (Dekker, 1993).Higher spectral resolution and careful placement of spectral bands has been shown to improve water quality retrieval
reflectance spectra obtained from these common satellite sensors, this study aims to identify the most suitable sensor for determining water quality in inland waters, and also determine whether data from these satellites can be used interchangeably and therefore trends can be spotted.
Aim and output
The aim of the present study is to identify the most suitable satellite sensor for effective optically active constituent retrieval by investigating the accuracy of OAC measurements of five different inland freshwaters along a longitudinal gradient in Eastern Australia collected by existing satellites Landsat-7, Landsat-8, Sentinel-2 and the future Sentinel-3. In other words, how accurate is the simulated satellite OAC concentrations in relation to the in situ measured OAC concentrations? The output of this study is a comparison between simulated Landsat-7, Landsat-8, Sentinel-2 and Sentinel-3 satellite OAC data and laboratory measured OAC data, conducted with adaptive linear matrix inversions applied to generated spectra made in EcoLight
These satellites were selected because the data is free and relevant for measuring inland water quality. Also these satellites have a relatively small pixel size, a reasonable revisit cycle and are currently operating or will be operating in the near future [ CITATION Dek12 \l 3081 ]. Remote sensors like MODIS, MERIS, VIIRS, IKONOS, Quickbird, SPOT-5, GeoEYE, RapidEye and Worldview-2 will not be included as they do not meet all mentioned criteria (Table 1).
Satellite Spatial
resolution (m)
Revisit cycle Free of
charge (Y/N)
Operating now or in the near future (Y/N) MODIS 250-1000 Daily Y Y MERIS 300 2-3 days Y N VIIRS 750 750 Y Y IKONOS, Quickbird, SPOT-5, GeoEye 2-4 On-demand/2-60 days N Y RapidEye 6.5 Daily N Y Worldview-2 2 On-demand N Y Sentinel-2 20-60 5 Y Y Sentinel-3 300 2 Y Y Landsat-7 30 16 Y Y Landsat-8 30 16 Y Y
Table 1: Satellite sensors and some properties important for monitoring of OACs. Bold characters are the properties where satellites do not meet the criteria set by this study. Data modified from Dekker and Hestir (2012)
2. Theoretical background
Water constituents
Optical water quality variables which can be estimated with remote sensing are; 1. chlorophyll pigments (CHL),
2. phycocyanin (CPC, CPE), 3. total suspended matter (TSM),
4. coloured dissolved organic matter (CDOM), 5. vertical light attenuation (Kd), turbidity,
6. bathymetry and
7. emergent and submerged aquatic vegetation.
The temperature of the water surface skin layer can be estimated with thermal infrared remote sensing but is not further discussed here as this requires earth observation sensors with thermal bands which usually have lower spatial resolution and are not present on each of these sensors.. The first four variables (1 to 4) are the most important water quality variables. CHL is an indicator of phytoplankton biomass and nutrient status and is important for assessing the quality of drinking water and the light environment; CDOM is the optically measurable component of dissolved organic matter in the water and important for the light environment; TSM is important for assessing the concentration of particulates suspended in the water column and the light environment in water and CPC and CPE are indicators of cyanobacterial biomass, common in harmful and toxic algal blooms. These constituents all affect the water reflectance spectrum in different ways; an example is shown in Figure 1. Because the CPC and its reflectance absorption minimum falls outside of the spectral bands of Landsat and Sentinel-2 this shall not be retrieved in this study.
In the visible and near infrared region (~400-900 nm) the influence of OACs interacts to modify the shape and amount of the spectrally reflected signal. In wavelengths longer than 900 nm water itself is such a strong absorber that very little radiation is reflected from the water bodies. It is dependent on the type of satellite sensor system and its spectral response band placement what the quality of the measurement is per constituent[ CITATION Dek12 \l 3081 \m Gue15].
Figure 1: A typical reflectance spectrum from a eutrophic inland water body and the regions in which the different OACs influence the shape of that spectrum [ CITATION Gue15 \l 1043 ].
Remote sensing
Earth observation
The satellites used in this study and some of their properties are shown in Table 2.
Satellite Landsat-7 Landsat-8 Sentinel-2 Sentinel-3
Satellite sensor systems ETM+ OLI/TIRS MSI OLCI
Spatial resolution (m) 30 30 10, 20, 60 300
No of Spectral Bands 8 11 12 21
Revisit cycle (days) 16 16 5 2
Swath width (km) 185 185 290 1270
Launch date April 1999 February 2013 June 2015 Late 2015 Years in orbit/Minimum
design life (yr) 15/5 2/5 0/7 0/7
Table 2: Properties of the satellite sensor systems: ETM+, OLI/TIRS, MSI, OLCI
The above information was compiled from the official USGS, ESA and NASA web pages (https://earth.esa.int/web/guest/missions/esa-future-missions/sentinel-3,https://earth.esa.int/web/guest/missions/esa-future-missions/sentinel-2 http://landsat.usgs.gov/landsat8.php, http://geo.arc.nasa.gov/sge/landsat/l7.html).
The spatial resolution is the pixel size, or the smallest surface area on de earth surface that can be measured. Sentinel-2 has the smallest overall spatial resolution at 10 m or 100 m2 and sentinel-3 the
largest at 300m or 90000 m2. The satellite sensors record the amount of reflected light in each spectral
band for each pixel.
Spectral resolution is determined by the number, the width and placing of the spectral bands. All earth observation sensor systems have spectral bands which are receptive to certain electromagnetic wavelengths. For these sensors they range between 400-12,000 nm. Only the spectral bands which range to 900 nm are useful for detecting OACs as they penetrate the water column and therefore only those are displayed in Table 3 and Figure 2 [ CITATION Gue15 \l 1043 ]. The panchromatic bands of the Landsats are also left out, as they have no added value to this study while those bands cannot detect OACs. Figure 2 shows how different the placements of the spectral bands are, comparing the different satellites.
Satellite Landsat-7 Landsat-8 Sentinel-2 Sentinel-3 Band Satellite sensor systems ETM+ Wavelength (nm) OLI/TIRS Wavelength (nm) MSI Wavelength (nm) / Spectral Resolution (m) OLCI Wavelength (nm) 1 483 443 443 / 60 400 2 565 483 490 /10 413 3 660 563 560 /10 443 4 852 655 665 /10 490 5 865 705 /20 510 6 740 /20 560 7 783 /20 620 8 842 /10 a:865 /20 665 9 674 10 681 11 709 12 754 13 761 14 764 15 768 16 779 17 865 18 885 19 900
Table 3: centre wavelengths (nm) of the different spectral bands. The spatial resolution of Sentinel-2 is also shown, as this varies per band.
The above information was compiled from the official USGS and ESA web pages
(http://landsat.usgs.gov/band_designations_landsat_satellites.php, https://sentinel.esa.int/web/sentinel/user-guides/sentinel-3-olci/resolutions/radiometric, https://sentinel.esa.int/web/sentinel/sentinel-2-msi-wiki/-/wiki/Sentinel%20Two/Performance).
Figure 2: Schematic overview of the spectral band placement of Sentinel-2, Sentinel-3, Landsat-8 and Landsat-7. Near infrared bands are displayed in grey.
In situ remote sensing
TriOS Ramses
The TriOS Ramses is a widely used in situ hyper spectral radiometer (350nm to 950 nm). It derives the down welling radiance from the sky Lsky; total upwelling radiance Lu and down welling irradiance
Ed . It has three radiometers; one held above water (Lsky) and two underwater sensors, which measure
radiance and irradiance (Lu, Ed) in the water [ CITATION Hom12 \l 1043 ]. With these three
parameters measured, remote sensing reflectance can be derived through[ CITATION Dek03 \l 1043 ]:
rrs (s)=
L
u(
s)
E
d (eq.1)
where s is the direction of the reflection.
The most important properties of the TriOS Ramses instrument are shown in Table 4.
BB9 and ac-s
The BB9 measures the optical backscattering b in the water. The ac-s measures optical absorption a and beam attenuation coefficient. With The BB9 and ac-s together, all inherent optical properties can be determined per site.
Instrument Deployment Measures Method Wavelength
(nm) Accuracy
TriOS Ramses
In water Lu, Ed, Lsky Three
Radiometers
320-950 0.3 nm
BB9 In water b Backscattering
Sensor
412-715
-ac-s In water c, a Spectrophotomete
r
400-730 0.01 m-1
Table 4: Some properties of the instruments used in the field. Data was retrieved from the official instrument sites; www.trios.de and wetlabs.com.
Radiometry
The colour of water is a complex optical feature, influenced by scattering and absorption processes as well as emission by the water column and of the reflectance by the substrate (Figure 3) [ CITATION Dek03 \l 1043 ].
Figure 3: Schematic diagram of the various processes that contribute to the signal as measured by a remote sensor [ CITATION Dek03 \l 1043 ].
Light incident upon water with all its components may be transmitted, scattered or absorbed. When the spectral absorption and scattering properties are determined of a water column, it is possible to calculate specific per unit spectral absorption and scattering, thus estimating the concentration of all constituents.
The IOPs only depend upon the medium. There are two main inherent optical processes, absorption (a) and scattering (b), with the sum described as the beam attenuation coefficient. An example of an IOP is the beam attenuation coefficient (c) and it represents the total loss of the light due to absorption and scattering combined [ CITATION AGD93 \l 3081 ]:
c=a+b (eq. 2)
The absorption coefficient of the medium as a whole, at a given wavelength, is equal to the sum of the individual absorption coefficients of the components present. Therefore:
a (total)=a ( w)+a ( NAP )+a (CHL)+a(CDOM )
(eq. 3) Where a(w) is the absorption of pure water, a(NAP) by non-algal particulate, a(CHL) by chlorophyll and a(CDOM) by coloured dissolved organic matter.Backscattering is the process by which photons change direction through interactions with matter and causes radiant energy to leave the water. It is caused by water b(w), by chlorophyll b(CHL) and non-algal particulate b(NAP) [ CITATION Kir83 \l 1043 ]:
Once the spectral absorption and scattering properties of the water samples have been determined it is possible to calculate concentrations and estimate reflectance and vice versa. This can be done in the program EcoLight accordingly [ CITATION Dek03 \l 1043 ]:
rrs=f (
b
a+b
)
(eq. 5)
aLMI analysis
The parameterisation of the IOP spectral shapes requires a detailed description of the often complex optical variability within freshwater systems. Adaptive models that use variable sets of specific IOP (SIOP) parameters can be used, which overcome problems of fixed SIOP sets. Brando et al. (2012) introduced an adaptive implementation of the LMI (aLMI), proposed by [ CITATION Hog96 \l 1043 ] , that incorporates a regional description of naturally occurring SIOPs derived from in situ measurements and samples collected simultaneously. This approach reduces computational
complexity and increases parameter retrieval accuracy by limiting the model retrievals to the natural variability expected within the study area [ CITATION Bra \l 1043 ].
3. The aLMI model first calculates the concentrations out of the input reflectance and one of the SIOP sets.
4. Then it calculates the modelled reflectance out of the measured concentrations with a matrix inversion. When there are more bands in the reflectance spectrum than concentrations (>3), the matrix is over determined and it uses singular value decomposition to conduct the inversion.
5. The last step of the loop is to see whether the used SIOP set is valid, so if the model optically closes. This is a vital component for testing the validity of the model. It compares the input reflectance to the modelled reflectance with the relative root mean square error.
6. The SIOP set with the best optical closure (the lowest ∆k) is used for determining
concentrations of OACs and the IOPs. Threshold conditions may be set to determine whether the spectral closure between the modeled and measured spectrum is sufficient to provide valid water quality information or not to avoid retrieving OACs that are too far from reality.
3. Methods
Site description
Six reservoirs have been sampled across Eastern Australia, spanning climatic regions from alpine to tropical latitudinal ranges. The reservoirs are: a deep alpine hydroelectric lake (Blowering Reservoir "BLO"); a temperate major headwater storage (Lake Hume, "HUM"); a semi-arid, small, shallow
vegetated weir subject to cyanobacterial blooms (Lake Cargelligo "CAR"); a temperate shallow artificial lake in the centre of Canberra (Lake Burley Griffin, "LBG"); and a reservoir with a large watershed that spans tropical rainforest and savannah in the north and dry tropical and semi arid savannah in the south (Burdekin Falls Dam, "BDF") [ CITATION Hes15 \l 1043 ]. These particular lakes have been chosen because of their diversity and recent databases (February/March 2013) with necessary water quality data are already available (provided by the Commonwealth Scientific and Industrial Research Organisation, CSIRO). Some properties of these lakes are shown in Table 5.
Site name
Reservoir name
Description Lat (S) Lon(E) Max
Depth (m) Surface area (km2) BDF Burdekin Falls
Dry tropics. Irrigation, urban supply.
20.627 147.045 40 220
BLO Blowering Cool temperate. Hydro-power, irrigation and recreation.
35.466 148.261 91 44
CAR Cargelligo Semi-arid grassland. Diversion weir/inflow wetland.
33.285 146.402 5 15 HUM Hume Warm temperate. Hydro-power,
irrigation, river regulation and recreation.
36.119 147.039 40 202
LBG Burley-Griffin
Dry-continental. Recreation and urban supply. 35.174 149.065 18 7 Site name Storage capacity ML x10^6
Primary inflow rivers Number
of sites CHL μg/La NAP a440/ma CDOM mg/La
BDF 1.86 Burdekin River, Belyando River 8 9.81 9.93 1.59
BLO 1.63 Tumut River 10 1.32 0.59 0.28
CAR 0.0004 Lachlan River 7 31.91 26.80 1.16
HUM 3.04 Murrey River 6 5.03 1.84 0.43
LBG 0.003 Molonglo River 5 12.48 7.75 2.01
Table 5: Description of sample sites, modified from [ CITATION Hes15 \l 1043 ]. a=Average of measured data
Materials
The materials required for this research are:
1. The TriOS Ramses, for in situ remote sensing reflectance data. 2. The ac-s and BB9, for measuring IOPs.
Field data collection and data analysis
Field data collection and filtration
On all five lakes five to ten sites were selected that are spatially distributed across the lake. All sites were visited within February and March (2013), the Austral summer. This is the temperate dry season and tropical wet season. Three samples for concentration determination were taken per site, the average of these samples is used as the measured concentration of the particular site. On every site the following activities were executed:
TriOS Ramses measurement BB9 and ac-s measurement GPS coordinates recording Water sampling
Taking photos of environmental conditions (sun glint, foam at surface, cloud cover etc.) Record water conditions/weather conditions/particulars
After every day in the field, all water samples need to be filtered to ensure that samples, holding organic matter, do not deteriorate. The protocol for filtration and analysis was handled according to [ CITATION Cle01 \l 1043 ]. The laboratory analysis is carried out by CSIRO Oceans & Atmopshere Laboratory in Hobart.
Data analysis and simulations
The following steps are taken in this research, the technical details are shown in Chapter 2: 1. The first step is to simulate satellite sensor reflectance. The remote sensing reflectance,
measured with the TriOS Ramses, is convolved with the program ENVI into the satellite bands (L7, L8, S2, S3) and into a generic 10nm band reflectance. The last reflectance spectrum has a spectral band every 10nm and will be used as the 'Measured Reflectance' in EcoLight as a reference.
2. IOPs measured with the ac-s and BB9 are calculated into SIOP parameter sets. All the parameter sets are input for the aLMI analysis. The program will select the best fitting SIOP set with the lowest relative RMSE, and thus the best optical closure.
3. The last input for the aLMI analysis are the measured CHL, CDOM and NAP concentrations. These concentrations are varied by -10%, 0% and +10% to diminish measurement errors, and give 27 simulated spectra for each site.
4. When the input is completed the aLMI analysis can be carried out. The output of this analysis is an Excel file per satellite with the columns: Site station; Measured CHL; Measured CDOM; Measured NAP; Modelled CHL; Modelled CDOM; Modelled NAP; SIOP set nr; Optical Closure.
5. The last step is to upload all data and perform further analysis and visualisation in Matlab. It should be noted that the columns with the measured concentrations of OACs are the
concentrations used in the analysis and not the simulated concentrations of the 10nm spectrum. Results of this analysis are shown in Chapter 4.
4. Results
This chapter discusses the results with the assumption that concentrations of compounds are measured correctly in the field and that the aLMI analysis reflects the output of OAC concentrations correctly. First, some descriptive statistics are shown, then the results of the root mean square error between
measured and satellite are discussed and lastly the importance of the relation between CDOM and CHL is explained.
Descriptive statistics
Basic statistical analyses has been performed on the concentration data of CHL, CDOM and NAP of all lakes combined. The total number of samples is 972 per satellite sensor, after the variation of concentrations of OACs. Descriptive statistics are shown in the following Tables:
Table 6: Descriptive statistics of measured and simulated CHL concentrations.
Table 7: Descriptive statistics of measured and simulated CDOM concentrations.
CHL Measur ed L7 L8 S2 S3 Mean 11,71 14,45 5,34 14,27 13,64 Standard deviation 11,72 23,54 13,83 17,00 15,94 25th Percentile 2,27 2,38 0,56 4,53 3,47 75th Percentile 18,27 15,92 3,77 12,24 15,73 Range 49,70 177,03 107,53 66,18 61,14 Maximum 50,52 177,03 107,53 67,07 61,58 Minimum 0,82 0,00 0,00 0,89 0,44 Median 7,94 6,31 1,54 7,37 6,88 CDOM Measure d L7 L8 S2 S3 Mean 1,04 0,78 0,66 0,67 0,83 Standard deviation 0,67 0,67 0,43 0,52 0,57 25th Percentile 0,42 0,35 0,37 0,25 0,35 75th Percentile 1,57 1,05 0,81 1,06 1,28 Range 2,78 7,09 2,24 2,05 2,25 Maximum 2,96 7,09 2,26 2,06 2,39 Minimum 0,18 0,00 0,02 0,01 0,14 Median 1,17 0,58 0,57 0,50 0,59 NAP Measure d L7 L8 S2 S3 Mean 12,95 14,58 10,02 17,37 17,00 Standard deviation 14,54 18,18 9,29 18,83 18,38
Figure 5: The root mean square error between the measured OAC concentrations in the field and the modelled OACs per satellite.
Figure 5 shows clearly that, ascending from least suitable to most suitable satellite sensors and when considering spectral band placement only, is; Landsat-7, Landsat-8, Sentinel-2, Sentinel-3. This is expected as Landsat-7 only has three usable bands in the visible light spectrum and Landsat-8 only has four. Sentinel-3 performs better than Sentinel-2 as its band placement is especially made for detecting OACs, and also has more bands (respectively five and eleven). However, using more bands does not automatically mean that concentrations will be measured more accurately. This is clearly visible in the 10 nm spectrum. This spectrum has a 10nm wide band every 10 nm, but it measures OACs less accurately than Sentinel-3 (Table 9).
RMSE CHL (ug/L) RMSE CDOM
(a440/m) RMSE NAP (mg/L) 10nm 7.8748 0.5373 10.7595 L7 23.7391 0.6632 15.6603 L8 16.1744 0.6286 9.2909 S2 9.2823 0.5961 8.4340 S3 6.7564 0.4919 7.4816
Table 9: output of the root mean square error between all measured and modelled OAC concentration data.
This phenomenon is explained by [ CITATION Bra03 \t \l 1043 ]. They emphasize the importance of band placement when using matrix inversion methods, as it is very sensitive to noise. When using a spectral band every 10 nm, a lot of bands with a high level of noise are present. The spectral bands of Sentinel-3 are especially placed where signal levels are high for water quality variables (Figure 2). As a result, Sentinel-3 detects less noise than the 10nm spectral band, and thus has a better accuracy in measuring concentrations of OACs.
The next figures display the measured against modelled AOC concentrations and the 1:1 line. This again shows that Sentinel-3 has the highest and Landsat-7 the least accuracy. Additionally, these figures show that some lakes are measured more accurately than others. The next paragraph will expand on these results.
Figure 8:The measured concentrations of OACs plotted against the simulated Sentinel-2 concentrations of OACs.
Relation between CDOM and CHL
The lake characteristics are significant for interpretation of the data. In figures 7 to 9 it is shown that for some lakes better results were achieved than for others. This is mainly due to the relation between CHL and CDOM. These two compounds can be difficult to distinguish in a light spectrum. To determine what lakes are causing spectral and covariance ambiguities, Figure 10 was made. This figure is a visualisation of the Pearson correlation coefficient between CHL and CDOM.
Figure 10: The Pearson correlation between CHL and CDOM for the five different lakes. P-values are all below 0.005.
CDOM has two possible sources: it can be autochthonous (a breakdown product of CHL from algae); or it is allochthonous CDOM. This can originate from: vegetation that falls into the water and breaks down, agricultural run-off, soil leaching, etc. When CHL and CDOM are related, the main source for CDOM is CHL, otherwise CDOM is allochthonous. When a lot of allochthonous CDOM is present, then NRMSE of CDOM will be worse [ CITATION Hes15 \l 1043 ].
Figure 10 clearly shows a low correlation between CHL and CDOM in BDF, CAR and BLO. This makes sense as Burdekin Falls has two completely different main river inlets that mixes water up in
Figure 11: The normalised root mean square error per satellite for the OACs and lakes.
The normalized root mean square error is conducted on the satellite data in order to compare CHL, CDOM and NAP. NAP is most accurately measured in all lakes. Landsat-7 and Landsat-8 struggle with differentiating OACs because they have broad and only few bands (three and four respectively). Sentinel-3 performs best in differentiating all OACs because it has many carefully placed, narrow spectral bands.
5. Discussion
There are some side notes to be considered when looking at this research. This research does not take every factor of signal to noise in consideration, only spectral band placement and width. Signal to noise is a function of band width, band location, pixel size and sensor sensitivity[ CITATION Lei11 \l 1043 ].
SNR=f (Band widt h , Band location, Pixel
¿
Sensor sensitivity)
(eq. 5) When the pixel size is greater, the sensor has significantly more time per pixel to measure photons as more photons can reach the sensor, due to the bigger area When looking at a sensor with the same sensitivity and decreasing the pixel size, the amount of signal that this sensor receives is significantly lower [ CITATION Lei11 \l 1043 ]. Sentinel-3 has the biggest pixel size in this research:300mx300m. Sentinel-2 the smallest: 10mx10m. Satellites travel with a ground speed of 10km/sec above the earth surface, which translates to 0.03 seconds for 300 meters and 0.001 seconds for 10 meter. If you would decrease the pixel size of Sentinel-3 to the pixel size of Sentinel-2, the amount of signal collected would be 303 =27,000 times smaller. If the signal to noise ratio was, for instance, 700
for Sentinel-3 it would then be 4.3, which would be unacceptable. Sentinel-2 has, of course, a better signal to noise ratio due to a higher sensor sensitivity (Table 10).
Satellite Mean signal to noise ratio
Landsat-7 31
Landsat-8 113
Sentinel-2 142
Sentinel-3 496
Table 10: Mean signal to noise ratios for the different satellites over the spectral bands in the visible light spectrum. Data was collected from respectively www.nasa.com and www.esa.com.
However, with a bigger pixel size, it is harder (or impossible) to measure smaller inland lakes. The edges of lakes in a pixel are not useable for remote sensing. Therefore a smaller pixel size would be better for smaller lakes, but the signal to noise ratio will be lower.
6. Conclusions
This research compared the earth observation satellites Landsat-7, Landsat-8, 2 and Sentinel-3. There are, of course, other satellites that could have been compared, however these satellites all have drawbacks on spatial resolution, revisit cycle and costs. Out of these selected four, simulated Sentinel-3 data gave the best results for measuring CHL, CDOM and NAP in different types of lakes. When considering signal to noise ratio as well, Sentinel-3 is still the better option. However, its pixel size may be a drawback on smaller lakes. For smaller (or narrow) lakes, Sentinel-2 is the best option to use, with its smallest pixel size out of all four.
Concentration data of all satellites may be used interchangeably while they measure the same component. However, when doing this, it is important to know the differences between the satellites, especially in time series. One should take in consideration that Landsat-7 and Landsat-8 are not very accurate for this specific purpose. Landsat-8 performs significantly better than Landsat-7 because it has more and slightly differently placed spectral bands. Broad band systems like these measure a lot noise in comparison to narrow spectral band systems like the Sentinels. Although, the narrow bands must be carefully placed for its purpose, otherwise it will still measure a high level of noise, which can be seen in the 10nm spectrum.
In an optimal situation there would be a satellite with the appropriate band placement and width, pixel size and sensitivity, especially made for OAC detection. As long as this kind of satellite does not exist, Sentinel-3 (or Sentinel-2, depending on lake size) is the best sensor to use for research on OACs.
Acknowledgements
First of all, I would like to thank the CSIRO for giving me the opportunity to experience working and contributing on one of their researches and providing me with data. I also want to thank dr. Janet Anstee, dr. Hannelie Botha and dr. Arnold Dekker for their advice, comments and time.
Citations
Aurin, D. A., & Dierssen, H. M. (2012). Advantages and limitations of ocean color remotesensing in CDOM-dominated, mineral-rich coastal and estuarine waters. Remote Sensing of Environment , 181-197.
Berger, M., Moreno, J., Johannessen, J. A., Pieternel, F. L., & Hanssen, R. F. (2012). ESA's sentinel missions in support of Earth system science. Remote Sensing of Environment , 84-90.
Brando, V. E., & Dekker, A. G. (2003). Satellite Hyperspectral Remote Sensing for Estimating Estuarine and Coastal Water Quality. Transactions on geoscience and remote sensing , 1378-1387. Brando, V. E., Dekker, A. G., Park, Y. J., & Schroeder, T. (2012). Adaptive semianalytical inversion of ocean color radiometry in optically complex waters. Applied Optics , 2808-2833.
Chappelle, E. W., Kim, M. S., & McMurtrey lll, J. E. (1992). Ratio Analysis of Reflectance Spectra (RARS): An Algorithm for the Remote Estimation of the Concentrations of Chlorophyll A,
Chlorophyll B and Carotenoids in Soybean Leaves. Remote Sensing of Environment , 239-247. Clementson, L. A., Parslow, J. S., Turnbull, A. R., McKenzie, D. C., & Rathbone, C. E. (2001). Optical properties of waters in the Australasian sector of the Southern Ocean. Journal of Gophysical Research , 31611-31625.
Dekker, A. G. (1993). Detection of optical water quality parameters for eutrophic waters by high reslution remote sensing. Amsterdam: Vrije Universiteit.
Dekker, A. G., & Hestir, E. L. (2012). Evaluating the Feasibility of Systematic Inland Water Quality Monitoring with Satellite Remote Sensing. CSIRO: Water for a Healthy Country National Research Flagship.
Dekker, A. G., Brando, V. E., Anstee, J. M., Pinnel, N., Kutser, T., Hoogenboom, E. J., et al. (2003). In Imaging Spectrometry of Water (p. Chapter 11).
Goward, S., Masek, J. G., Williams, D. L., Irons, J., & Thompson, R. J. (2001). The Landsat 7 mission: Terrestrial research and applications for the 21st century. Remote Sensing of Environment , 78 (1-2), 3-12.
Guerschman, J. P., Donohue, R. J., van Niel, T. G., Renzullo, L. J., Dekker, A. G., Malthus, T. J., et al. (2015). Earth observation for water resources management. The World Bank.
Irons, J. R., Dwyer, J. L., & Barsi, J. A. (2012). The next Landsat satellite: The Landsat Data Continuity Mission. Remote sensing of environment , 11-21.
Julian, J., Davies-Colley, R. J., Gallegos, C. L., & Tran, T. V. (2013). Optical Water Quality of Inland Waters: A Landscape Perspective. Annals of the Association of American Geographers , 309-318. Kerker, M. (1969). The Scattering of Light and Other Electromagnetic Radiation: Physical Chemistry: A Series of Monographs. Ney York: Academic Press inc.
Kirk, J. T. (1983). Light & Photosynthesis in Aquatic Ecosystems. Cambridge University Press. Leijtens, J., Perez-Calero, D., Dekker, A. G., & Peters, S. (2011). SWIMS (hyper Spectral Water Imaging & Monitoring System) - A hyperspectral constellation concept for aquatic ecosystems. Remote Sensing of Environment .
Li, P., Jiang, L., & Feng, Z. (2014). Cross-Comparison of Vegetation Indices Derived from Landsat-7 Enhanced Thematic Mapper Plus (ETM+) and Landsat-8 Operational Land Imager (OLI) Sensors . Remote Sensing , 310-329.
Matthews, M. W. (2011). A current review of empirical procedures of remote sensing in inland and near-coastal transitional waters. International Journal of Remote Sensing , 6855-6899.
Palmer, C. J., Kutser, T., & Hunter, P. D. (2015). Remote sensing of inland waters: Challenges, progress and future directions. Remote Sensing of Environment , 1-8.
Roy, D., Wulder, M., Loveland, T., Woodcock, C., Allen, R., Anderson, M., et al. (2014). Landsat-8: Science and product vision for terrestrial global. Remote Sensing of Environment , 154-172.
Serway, R., & Jewett, J. (2014). Physics for Scientists and Engineers. Boston: Brooks/Cole. Vogelmann, J. E., Helder, D., Morfitt, R., Choate, M. J., Merchant, J. W., & Bulley, H. (2001). Effects of Landsat 5 Thematic Mapper and Landsat 7 Enhanced Thematic Mapper Plus radiometric and geometric calibrations and corrections on landscape characterization . Remote Sensing of Environment , 55-70.
Wettle, M., & Brando, V. E. (2006). SAMBUCA Technical Report. A Semi-Analytical Model for Bathymetry, Un-mixing, and Concentration Assessment. CSIRO.
