We're sorry but this page doesn't work properly without JavaScript enabled. Please enable it to continue.
Feedback

Sea water turbidity analysis from Sentinel-2 images: atmospheric correction and bands correlation

00:00

Formal Metadata

Title
Sea water turbidity analysis from Sentinel-2 images: atmospheric correction and bands correlation
Title of Series
Number of Parts
351
Author
Contributors
License
CC Attribution 3.0 Unported:
You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal purpose as long as the work is attributed to the author in the manner specified by the author or licensor.
Identifiers
Publisher
Release Date
Language
Production Year2022

Content Metadata

Subject Area
Genre
Abstract
Sea water turbidity analysis from Sentinel-2 images: atmospheric correction and bands correlation Sea water turbidity is a measure of the amount of light scattered by particles in water. It is due to the presence of suspended particles, which it is operationally defined as the fraction in water with less than 2 µm in diameter. Plankton can also generate turbidity, but high turbidity events are dominated by high concentrations of inanimate inorganic particles. High levels of suspended sediments in coastal regions can occur as consequence of high sediment load from rivers, from bottom sediment resuspension due to wave actions or due to anthropogenic activities, such as dredging operations or bottom resuspension from ship propellants. The increase of turbidity can determine negative environmental effects both on the biotic and abiotic marine ecosystem. In highly anthropized coastal marine systems, like harbours, sediments represent a sink for contaminants and resuspension can contribute to propagate pollution to unpolluted areas (Lisi et al., 2019). Many marine water quality monitoring programmes measure turbidity. Traditional methods (e.g., in situ monitoring) offer high accuracy but provide sparse information in space and time. Earth Observation (EO) techniques, on the other hand, have a potential to provide a comprehensive, fast and inexpensive monitoring system to observe the biophysical and biochemical conditions of water bodies (Caballero et al., 2018; Saberioon et al., 2020; Sagan et al., 2020). Hence, some of the authors are developing a semi-empirical model for predicting water turbidity by combining Sentinel-2A data and machine learning methods using samples collected along the North Tyrrhenian Sea (Italy). Field data collected at the study site from April 2015 to December 2020 were made available by ARPAL, even though most of these data refer to low turbidity events. In the framework of this research activity, Sentinel-2A multispectral optical images, freely available within the EU Copernicus programme, are elaborated. It’s well known that such products are provided at Level-1C (L1C) Top of Atmosphere (TOA) and at Level-2A (L2A) Bottom-Of-Atmosphere (BOA). L2A BOA reflectance products are preferred as they are already corrected for effects of the atmosphere. However, the official L2A data are available for wider Europe from March 2018 onwards. The necessity to use the complete on-site dataset to calibrate the predicting model, and not only data after March 2018, required the identification of the most appropriate algorithm for atmospheric correction of L1C images relative to study area between 2015 and 2018. Hence, a comparison between the available L2A BOA product ant the corresponding L1C image corrected in different open source environment was performed. In particular, the free and open source QGIS and GRASS GIS, and the Sentinel Application Platform (SNAP), provided by ESA/ESRIN free of charge to the Earth Observation Community, published under the GPL license and with its sources code available on GitHub, were used. Both image-based method, i.e. the Dark Object Subtraction (DOS) method in QGIS, and physically-based methods, i.e. the Second Simulation of Satellite Signal in the Solar Spectrum (6S) method in i.atcorr module of GRASS GIS and the Sen2Cor algorithm inside SNAP, were applied (Lantzanakis et al., 2017). The great advantage of the DOS method is that it focuses only on the spectral and radiometric characteristics of the processed image, hence it doesn’t require remote or in-situ atmospheric measurements. But the performed correction doesn’t seem so accurate. Instead, the physically-based approach requires atmospheric measurements and parameters, that are difficult to be identified so to be coherent in space and time with the processed image. The most complex physical parameter to set is Aerosol Optical Depth (AOD), which is a dimensionless parameter related to the amount of aerosol in the vertical column of the atmosphere over the target station. It usually range from 0 to 1, with values less than 0,1 that corresponds to a clean atmosphere with high visibility, and values higher than 0,4 that corresponds to hazy atmosphere with very less visibility. AOD is spatially and temporally very variable. It can be estimated from AERONET (AErosol RObotic NETwork), a federation of ground-based remote sensing aerosol networks with more than 25 years of data. A station which measured the Aerosol Optical Depth at 500 nm at Level 2 (quality-assured) at the same time as the scene was taken, is not always available nearby the site under study. Hence the evaluation of AOD variability in time and space was analysed for the area and the events of interest, so to identify the proper values. Expecially i.atcorr seems very sensitive to the set values of AOD. Once the proper method for atmospheric correction was identify, it was applied to the L1C images relative to the collected field data from April 2015 to March 2018. Then, the correlation between the in-site dataset and the individual bands known to be most sensitive to water turbidity, i.e. blue (B2), green (B3), red (B4) and near infrared (B8 and B8A) bands, was analysed, finding good results for the visible bands, and a weak correlation with NIR bands. In addition, indexes defined by the ratio between the three visible bands were checked to see which combination could best highlight the turbidity of the water from the Sentinel-2 images. Preliminary results seem to confirm that the identified EO technique could provide a fast and inexpensive monitoring system to observe sea water turbidity along the Northern Tyrrhenian Sea (Italy).
Keywords
AreaMathematical analysisCross-correlationPresentation of a groupComputer animation
SatelliteMathematical analysisObservational studyMusical ensembleCross-correlationComputer fontMathematical analysisSatelliteObservational studyMappingInformationCASE <Informatik>Computer animation
MathematicsWater vaporNegative numberMeasurementParameter (computer programming)Sound effectParticle systemIntegrated development environmentComputer animation
InformationMatter waveAngleExecution unitScatteringDegree (graph theory)SpacetimeMeasurementSurfaceElectronic signatureSpectrum (functional analysis)Function (mathematics)VotingSatelliteInformation retrievalOpen sourceSoftware frameworkStrategy gameMaß <Mathematik>Observational studyAreaMultiplicationCompilation albumProduct (business)Musical ensembleVaporAngular resolutionMusical ensembleExecution unitSet (mathematics)InformationView (database)AreaSatelliteSurfacePosition operatorCondition numberMedical imagingDifferenz <Mathematik>Water vaporEvent horizonComputer programProduct (business)HistogramObservational studyIntegrated development environmentRemote procedure callMultiplication signMeasurementImage resolutionObject (grammar)Dependent and independent variablesElectronic signatureParameter (computer programming)TesselationComplex systemMetreAngleMatter wavePoint (geometry)Information retrievalReflection (mathematics)Component-based software engineeringVotingSpacetimeFood energyDegree (graph theory)Greatest elementParticle systemMusical ensembleState observerType theoryOpen sourceMereologyIncidence algebraGreen's functionSpectrum (functional analysis)ScatteringComputer animation
Process (computing)ScatteringProduct (business)Level (video gaming)GeometryGreatest elementObservational studyOpen sourceGreatest elementParticle systemElectronic signatureEnergy levelSound effectReflection (mathematics)Computer animation
SoftwareGrass (card game)Density of statesInformationChainProcess (computing)DataflowProduct (business)ImplementationMusical ensembleSystem identificationPrice indexPoint cloudGroup theoryDifferenz <Mathematik>Computer-generated imageryLogical constantPixelSatelliteSpectrum (functional analysis)SimulationRayleigh-ZahlSound effectObject (grammar)Characteristic polynomialLevel (video gaming)Set (mathematics)Numbering schemePresentation of a groupInformationMedical imagingGrass (card game)Musical ensembleUniform resource locatorGreatest elementReflection (mathematics)PixelSatelliteObject (grammar)Particle systemSoftwareResultantPoint cloudProduct (business)Water vaporLibrary (computing)Energy levelSound effectDifferenz <Mathematik>WordMappingMultiplication signOpen sourceComputer animation
Standard deviationService (economics)Population densityMetadataDiscrete element methodProduct (business)Digital signaloutputParameter (computer programming)Parameter (computer programming)Arithmetic meanSoftwarePixelComputer fileOpticsProduct (business)Multiplication signMedical imagingPosition operatorGeometryService (economics)Level (video gaming)InformationGrass (card game)MetadataComputer animation
Workstation <Musikinstrument>WebsiteDistanceParameter (computer programming)Characteristic polynomialComputer networkMeasurementRoboticsOpen sourceWebsiteRoboticsCharacteristic polynomialGrass (card game)SoftwareMeasurementParameter (computer programming)AreaCASE <Informatik>Workstation <Musikinstrument>Observational studyComputer animation
Cellular automatonSound effectParameter (computer programming)Workstation <Musikinstrument>Statistical dispersionComputer-generated imageryLogical constantMusical ensembleProcess (computing)PixelMeasurementPairwise comparisonCross-correlationMusical ensembleSingle-precision floating-point formatMatter waveMedical imagingCross-correlationParameter (computer programming)Greatest elementWorkstation <Musikinstrument>CASE <Informatik>Sound effectSet (mathematics)Square numberGrass (card game)Reflection (mathematics)SoftwareError messageSubject indexingInformationCoefficient of determinationArithmetic meanEnergy levelFehlerschrankeMusical ensembleMusical ensembleValidity (statistics)Right angleRootVariable (mathematics)BendingRange (statistics)Resultant1 (number)Computer animationDiagram
Compilation albumEnergy levelMusical ensembleLevel (video gaming)Set (mathematics)Medical imagingSound effectMappingRight angleMaxima and minimaWater vaporCartesian coordinate systemShape (magazine)Computer animation
Product (business)Latent heatComputer-generated imageryOpticsInformationGreatest elementLimit (category theory)Sound effectObservational studyCross-correlationMaxima and minimaImage resolutionSpacetimeAreaVirtual machinePredictionImage resolutionPoint cloudCondition numberSoftware developerInformationGreatest elementSpacetimeMeasurementSet (mathematics)State observerGoodness of fitProduct (business)Observational studySound effectCore dumpComplex analysisOpticsPredictabilityVirtual machineSystem callCross-correlationReflection (mathematics)CASE <Informatik>Musical ensembleComputer animation
Transcript: English(auto-generated)
Thank you Bianca for your kind presentation and thank you everyone for attending. So yes, today I will present our work that was done to obtain, to analyze sea water turbidity from Sentinel-2 images.
So first of all I will say something about turbidity, what turbidity is and why we study for water quality purposes. I will recall briefly some satellite remote sensing basics information that make it useful for water quality analysis and then I will go deeper into our case study.
I will present all the steps for the study that goes from atmospheric correction to band analysis up to the realization of turbidity maps. I will show some results and get some conclusions. So what is turbidity? I think we're all pretty familiar about turbidity but what is it exactly?
Turbidity is a measure of the amount of light scattered by the particles that are in water. For sea water, changes in turbidity are due to, for example, algal blooms but the most relevant contribution to changes in turbidity is due to sediment.
For example, sediment coming from river flows. Especially when related to anthropogenic activities like beach nourishment or dredging activities, the increase of turbidity can be associated to negative environmental effects. That's why turbidity is one of the parameters that is considered for European legislation
for water quality monitoring. In situ, turbidity monitoring is done by a probe. As said, turbidity is not the only measure that is taken so usually it's measured with a multi-parametric probe that also measures other parameters.
The measured turbidity depends of course by the particles that are present in water but also on the instrument, on the wavelength of the light and the angle at which the detector is positioned. For example, nephilometric turbidity unit, NTU, measures the scattered light at 90 degrees from the incident light beam measured with a white light.
In situ monitoring offers high accuracy data but it is of course time consuming and offers information that is sparse both in time and space. On the other hand, remote sensing and Earth observation can provide fast and economic
when we consider free available data and most of all a synoptic view of the situation. Remote sensing and in particular optical remote sensing measures the solar energy that is reflected by the Earth's surface.
The different objects and features of the Earth's surface have a unique spectral response called spectral signature. And this is the fact that we can use to study seawater surface. In fact, the seawater surface has a spectral signature that depends on the optical significant
constituents. There are phytoplankton pigments, color-disturbed organic matter and suspended matter. We can see in the right image the spectral signature of the different components and we can notice that water absorbs energy of most of the wavelength while the presence
of chlorophyll causes a small peak of reflectance in the green area. But sediments is the particle that causes the higher values of reflectance not only
in the visible spectrum but also in the near infrared region. So the aim of the study was to retrieve turbidity information in a study area. The study area is here reported. It's in the North Iranian Sea that includes Liguria region and Tuscany region.
You can see here the study area and it's a very complex area from the turbidity point of view because it comprehends many different sources of turbidity. We have in the northern part a marine protected area. So it's supposed to be a very clean type of water.
But then we have a harbor. We have two river mouth, Magra river mouth and Arno river mouth. And again we have another harbor that is Livorno harbor.
So it's a very complex system. We use in situ turbidity measure collected from 2015 up to 2021 by the regional agency for the environmental protection of Liguria where I work and Tuscany. According to the official monitoring programs.
Turbidity is measured in a philometric turbidity units. You can see on the bottom left image the histogram of the turbidity data. This data set is not just the data that were collected in these six years but the data corresponding to the availability also of satellite image
because satellite image is not always available. We can immediately notice how the data set is very imbalanced towards low data and especially data lower than three NTU which means clear water. This is also due to the fact that the high turbidity events are associated
to intense meteorological events during which the sea condition does not allow in situ monitoring or the sky is cloudy so the satellite information is not available.
We use the Sentinel-2 multispectral instrument products because of its time and special resolution. Especially in the visible bands we have a special resolution of 10 meters. The products are available in areas of fixed size called tiles
and our study area is totally included in one tile. The revisit time in this area is about three days. Sentinel-2 provides data since 2015 but only since March 2018 the data are provided as so-called level 2 that is bottom of atmosphere reflectance.
We know in fact that the atmosphere interfere with the sunlight and changes the spectral signature. So the effects of the atmosphere are basically absorption. For example UV rays are blocked by ozone while carbon dioxide stops the thermal infrared
and by scattering due mostly by aerosol particles. So the so-called top of atmosphere reflectance which is the one that is measured by the sensor is composed by the bottom of atmosphere or surface reflectance plus the contribution of atmosphere.
So this is the scheme of our work. So first of all we collected of course the in-situ data and we collected the satellite data for the same day. As said we had to remove all the data with the presence of cloud
and then because not all the Sentinel data were available as a bottom of atmosphere information we had to do some atmospheric correction. And three open source softwares were tested and compared the QGIS, the GRASS GIS and the SNAP. And the results of the bottom of atmosphere obtained reflectance were compared
to the level 2 Sentinel-2 data because we wanted to have a homogeneous data set to use. So once the best method was identified and the whole data set was used to identify which bands were more correlated to turbidity and we only considered the bands that are known to be correlated to particles in water
that we saw before. So we found calibrated and validated a turbidity model that was used finally to obtain some turbidity maps. About atmospheric correction, just a few words. There are two main groups of atmospheric correction methods. The first one and is the easiest one are the image-based approaches.
They consist of the subtraction of a constant value from each pixel of the processed image. So this is a faster way because it doesn't require any information about the atmosphere. Basically it is based on the idea that the darkest pixel in the image
have a reflectance that is the lowest reflectance of the image that is due not because of the reflectance of the object but the effect of the atmosphere. So that is the value that is considered at the atmospheric effect and is removed by the other pixel. The physically based approach instead, they rely on the use of a radiative transfer model
to accurately compute the atmospheric effect. Examples are the 6S method which is used by GRASS GIS and the LibRATRAN method which is used by SNAP. Of course they need some atmospheric information at the time and location of the image acquisition.
So what parameters are required for the two methods? Of course some sensor and solar geometries information, the date and time of acquisition and the position of the image. But these data are easy to find because they are provided by the metadata file. We need the mean target elevation above sea level which is usually derived by
digital elevation model. And then we need the most difficult one are the data related to the atmosphere. So atmospheric model and aerosol model. A very important parameter is the aerosol optical depth or thickness at 550 nanometers. This is a dimensionless parameter that is related to the amount
of aerosol that is present in the vertical column of the atmosphere above the target. The higher the IOD, the less sunlight passes through the atmosphere. So for SNAP, this is a parameter that is automatically calculated by the
software from visibility using dark reference areas. If no dark pixels are present in the image, a forecast data obtained from Copernicus atmosphere monitoring service is used. While from GRASS GIS, we used an IOD parameter estimated from the aerosol robotic network.
It is possible to search for the nearest measurement site or the site that has the atmospheric characteristic closest to the study area. In our case, we identified five stations, Ispra and Modena because they are the closest to the study area,
but they're not coastal sites so they have high altitude. So we also analyzed three other station, Rome, Naples, and Palma de Mallorca in Spain, because they are at the coast, so they should have more similar characteristics. From the image on the left, we have an example of the AOD daily variability
for the different stations. What we can immediately notice is that the values are not constantly available. For example, for that date, the data from Naples stations were not available. On the right image, we can see the effect of the AOD parameter
on the atmospheric correction in GRASS GIS. Again, for just one day, an example, and one band. What we can notice is that changing the AOD parameter changes a lot the bottom of atmosphere reflectance obtained. And most of all, we can see that the AOD parameter of 0.15,
which is the one that we would use by looking at the stations, is not actually the one to give us the best bottom of atmosphere values compared to the level 2 Sentinel-2. So results obtained with the different softwares were compared to the level 2 data.
We can say that the QGIS is very fast, but is not very accurate, at least for this case. While the GRASS GIS software, which is a physically based approach, it can be very accurate, but in our case,
because we didn't have information about the AOD parameter, it's very difficult to use. On the other hand, SNAP, because it evaluates automatically, the AOD parameter was very efficient and actually accurate when compared to level 2 data. So the bottom of atmosphere reflectance obtained either directly from level 2
or after atmospheric correction was then used to identify which bands are more correlated to turbidity. Again, I said just the visible and near infrared bands were considered. And we can see here that the bands in the visible range
are the ones that are most correlated to turbidity. They were used to identify an index, were combined together to identify an index to relate it to turbidity. The correlation is reported here. To validate the correlation, a validation data set was used,
and you can see here the indexes R square, root mean square error, and mean absolute error were calculated to evaluate the accuracy. And you can see that the correlation that was found is linear with respect to the index,
but is actually non-linear with respect to the single bands. So the relationship was finally used to obtain turbidity maps. You can see in the left image the RGB composite or true color image for one specific day.
And on the right image, we have the turbidity map corresponding to the same day. Well, of course, we can notice that the land is evaluated as very high turbidity levels. Yes, of course. But we can clearly see that the shapes of the turbidity plumes are well represented.
The values are good values. I mean, they are physically good values. But even though they are very correlated to the calibration data set that we had, I mean, we cannot think to estimate values that are lower than the minimum value that we had,
or higher than the maximum value that we had for the calibration data set. This is another example of a turbidity map. Again, we can see how the turbidity plumes are well represented. I just wanted to recall that this is the Arno River, and we have here. Another river.
What we can see, though, in this image is that there are some effects due to the image, the Sentinel-2 image. We can also notice, not so well, but we can also notice them in the RGB composite. This is probably due to the fact that the Sentinel-2 image is actually not thought for application on water.
For example, they don't consider sun-glint correction, which is done instead with the Sentinel-3 ocean land color instrument. To conclude, some conclusion about the atmospheric correction. So as we said, the semi-empirical physically based model to obtain
water quality information from optical data required the use of bottom of atmosphere reflectance, which has to be obtained from atmospheric correction. From our experience, the physically based methods are performed better. However, the IOD, or in general,
the atmospheric information that are required can make it difficult to apply. Sentinel core from ESA was found to be the most suitable for our case study. To identify correlation, we think we perform good. We can actually say that earth observation can be an important tool
to support water quality monitoring. It provides especially wide information at significant resolution in space and time. The relationship shows a good agreement between measures and prediction. However, as said, it is highly dependent on the data set available.
Of course, it could benefit from new data, but they are very hard to obtain because they are limited to the sea condition and from cloud condition. To limit the effects of the atmosphere correction and improve the complexity of
the relationship, for example, including more bands, machine learning techniques are now under development. And before concluding and thanking you all, I would like to make an open call saying that if any of you works with water quality and turbidity, it's very welcome to cooperate and test our model with your data.
Thank you so much.