Geographical Weighted Regression Model For Improved Near-shore Water Depth Estimation From Multispectral Imagery
This is a modal window.
The media could not be loaded, either because the server or network failed or because the format is not supported.
Formal Metadata
| Title |
| |
| Title of Series | ||
| Number of Parts | 183 | |
| Author | ||
| License | CC Attribution - NonCommercial - ShareAlike 3.0 Germany: You are free to use, adapt and copy, distribute and transmit the work or content in adapted or unchanged form for any legal and non-commercial purpose as long as the work is attributed to the author in the manner specified by the author or licensor and the work or content is shared also in adapted form only under the conditions of this | |
| Identifiers | 10.5446/32001 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language | ||
| Producer | ||
| Production Year | 2015 | |
| Production Place | Seoul, South Korea |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
|
FOSS4G Seoul 201534 / 183
7
8
47
53
54
65
73
74
79
82
84
92
102
103
105
124
126
127
130
141
142
143
156
161
162
170
176
178
181
183
00:00
AlgorithmMathematical modelRegression analysisLinear regressionMaterialization (paranormal)MereologySatelliteComputer fontWater vaporWeightConnectivity (graph theory)EstimatorMathematical analysisMedical imagingUniverse (mathematics)Moment (mathematics)Pairwise comparisonPresentation of a groupInteractive televisionEvent horizonSymbol tableComputer animationSource code
02:00
AlgorithmMedical imagingMathematical modelType theoryAngular resolutionCategory of beingIntegrated development environmentTemporal logicFunctional (mathematics)Maxima and minimaMatter waveMultiplicationResultantSatelliteSpektrum <Mathematik>Goodness of fitWater vaporCASE <Informatik>Connectivity (graph theory)Field (computer science)Musical ensembleExterior algebraInterior (topology)Greatest elementMetreMultiplication signImage resolutionVolume (thermodynamics)Mathematical analysisOrder (biology)Computer programmingElectric generatorUniverse (mathematics)Group actionGodInternetworkingOperator (mathematics)Parallel portSummierbarkeitSet (mathematics)Element (mathematics)Single-precision floating-point formatMathematical modelComputer animation
06:05
AlgorithmPerformance appraisalPairwise comparisonSatelliteAreaWater vaporGreatest elementFreewareImage resolutionCellular automatonObservational studySoftware developerComputer animation
06:48
AlgorithmMedical imagingDynamic rangeGraph (mathematics)ScatteringInformation retrievalMathematical modelPlotterType theoryLevel (video gaming)Angular resolutionRegression analysisBand matrixReduction of orderBitExtrapolationPearson product-moment correlation coefficientPerformance appraisalLinear regressionMatter wavePairwise comparisonResultantSatelliteTable (information)Term (mathematics)NumberAreaGoodness of fitNichtlineares GleichungssystemDivisorWater vaporWeightCASE <Informatik>Connectivity (graph theory)Electronic signatureResidual (numerical analysis)Musical ensemblePoint (geometry)Social classSet (mathematics)Open setPixelKernel (computing)CircleCoefficientEstimatorRadiusRadiometryGreatest elementAddress spaceGene clusterSupervised learningGrass (card game)Coefficient of determinationDifferent (Kate Ryan album)Single-precision floating-point formatMultiplication signMathematical analysisData managementNetwork topologyIncidence algebraDialect3 (number)Group actionStructural loadPhysical systemExecution unitSurface of revolutionSubsetCellular automatonMatching (graph theory)Operator (mathematics)Parameter (computer programming)Process (computing)Metropolitan area networkInstance (computer science)Shared memoryExistenceQuotientShape (magazine)Student's t-testEvent horizonSymbol tableStandard deviationMessage passingPosition operatorComputer animation
15:42
Mathematical modelCategory of beingBitTime zoneShared memorySocial classMultiplication signScatteringPlotterProfil (magazine)Regression analysisPoint (geometry)MetreDifferent (Kate Ryan album)
16:15
Mathematical modelLine (geometry)GodStudent's t-testAlgorithmRegression analysisBitPoint (geometry)MetreComputer animation
16:52
Mathematical modelResultantFundamental theorem of algebraElectric generatorNumberAreaPoint (geometry)EstimatorMetreTwitterIncidence algebraLimit (category theory)Standard deviation
17:38
ScatteringPlotterProfil (magazine)Cross section (physics)TwitterComputer programmingMathematical modelAreaSocial classCoefficientPoint (geometry)Multiplication signDemosceneComputer animation
18:35
AlgorithmSparse matrixScatteringMathematical modelPlotterMathematical optimizationAxiom of choiceInternetworkingPresentation of a groupKernel (computing)Observational studyEstimatorConcentric1 (number)Dynamic rangeProfil (magazine)Regression analysisBitResultantNear-ringGoodness of fitDiameterModule (mathematics)Computer animation
20:18
Well-formed formulaForm (programming)NumberDiameterWordRadiusWebsiteBand matrixGoodness of fitPoint (geometry)PixelMetreComputer animation
21:37
Computer animation
Transcript: English(auto-generated)
00:05
Good afternoon everyone. My name is Vinay. I am a PhD candidate from Osaka City University, Japan. So basically my research is about estimating near shore bathymetry from remote sensing
00:23
imageries, basically optical remote sensing imageries. So we were testing so many algorithms and improving the algorithms and modifying the existing algorithms and all. So here I will talk about the geographical weighted regression model which is using to
00:41
estimate the water depth which is basically using for all kind of spatially related estimation which is a linear regression actually. So this is the outline of my presentation today. In the introduction part I will briefly talk about how optical remote sensing can be used
01:06
for near shore depth estimation and the advantages of using optical remote sensing for depth estimation and the physical principle behind that. And in materials and method I will talk about the correction methods which I have
01:23
been used to correct the components which is already there and contained in the satellite imageries which should be removed for better estimation and then the global regression method which is a conventional method which is a simple multiple linear regression method
01:41
which is used by many others to estimate the depth and then the proposed method which is called GWR model. So finally we will evaluate the results and compare the results between by the ground truth data is collected.
02:02
So these are the both types which mainly used to collect depth from the near shore or ocean. So one is LIDAR airborne LIDAR bathymetry and the other one is multibeam echo sounding. So both are very good accuracy and high resolution datasets but there are disadvantages
02:23
because due to inaccessibility of such kind of big data and time consuming and then of course cost is also a problem. So there are availability of some datasets of depth which are already there and some
02:41
of them are free and not all among that Jepco is giving good resolution which is not high resolution which is only 900 meter resolution. I don't think it can be used for any kind of coastal dynamic modeling. So so many others were trying to have an alternative to get depth in an easy way.
03:06
So the alternative method we have been used that multispectral imageries. So there are lots of advantages and some disadvantages also. The advantage is that there are wide availability of datasets like so many satellite imagers
03:24
which have lots of bands can be used to estimate effectively maybe reliable quality and things like that. It is relatively low cost when you compare with the field depth collections and large
03:41
spatial coverage and high spatial resolution and also temporal data accessibility and our case of satellite imager is such a way you can have and you can derive bathymetry it is 20 meter I have written here which is which can be changed according to water quality mainly the water quality the stuff how can the light can penetrate to the bottom
04:05
of the water. If in case of coastal environmental water is very complex so it is not easy to penetrate up to the bottom. So the advantage is it is relatively low accuracy and comparing with the other multi
04:23
beam echo sounding methods and you need a water depth anyway you have to collect water depth data to calibrate or compare or evaluate your results. So this is the physical principle of the algorithm which actually attenuation of light in water
04:44
is a function of wavelength if you take a short wavelength light in the electromagnetic spectrum which can penetrate maximum water depth which is written here 200 meter which is not a cannot be applied in a coastal complex water which can be very less maybe
05:02
30 or some 30 meter or something like that and if you when your wavelength increases your attenuation speed decreases. So if you go to the high wavelength I mean long wavelength components of the electromagnetic spectrum which cannot penetrate to the water bottom which will attenuates from the shallow
05:24
water itself. So other concept is that there are four main components which is contained in a satellite imagery that is one is water surface reflectance and the other one is inner water volume which
05:42
is water column properties when the light penetrates to the bottom and you will of course have a bottom reflectance if it could reach to the bottom and by going back to the signals it is the atmospheric scattering. So in my case my interest is only bottom reflectance I have to remove all three other
06:01
components. So this is the steady area pure authorica which is lies on north eastern Caribbean sea. There are two reasons to select a pure authorica as the steady area one is that free availability
06:22
of high resolution LiDAR depth data which could be used for calibration and comparison of the vessels and the water quality of the area is very good which is a coral reef area and even you can see some satellite imagery the bottom itself that means clear water we have in this region.
06:42
So we have selected pure authorica as our steady area to apply our algorithm and evaluate it. These are the data sets used to estimate depth. I have collected two satellite images one is Landsat 8 which is freely available open data and Rapidae which has high spatial resolution and high radiometric resolution
07:06
of 12 bit dynamic range and even the LiDAR depth data which is also freely available from NOAA. So I will briefly talk about the correction method I have been used to remove which is
07:27
an existing method which I have made modification in terms of in terms of in case of Landsat 8 data there are two new bands available compared to Landsat 7 which is called
07:42
costal band which can be used for estimation because it is very short wavelength can penetrate and short wave infrared band which is high wavelength band and this band is using
08:00
for correction and the other bands are using for estimation. This much bands are using for estimation and short wave infrared band using for correction because short wave infrared cannot penetrate to the water bottom it will activate when it reach to the water. So all the other components are acquired by the short wave infrared and only bottom
08:25
and bottom reflectance will not be there. So you can effectively use to remove the unwanted components from the image.
08:41
So this is the water depth retrieval models. Global regression model is a conventional model which is being used by so many others and which was giving good resolution results but when you examine the results you can
09:03
find that it is not able to address the heterogeneity in the data which is due to different bottom types and different water quality especially spatial problems of having bottom types and different water quality. So I was thinking of having a new model which can be effectively address the problem
09:26
due to the spatial heterogeneity. So this is the residual map which is shown on the left side and this map is generated
09:41
by just subtracting the estimated depth by global regression model and from the LIDAR depth. So the result is showing that the residuals are not uniformly distributed on the map and it has a cluster of residuals that means that some of the local areas have influence
10:02
on estimating depth. And here I forgot to tell you one thing, in the global regression model it is a simple multiple linear regression model which is having single set of coefficients which is you develop a single set of coefficients and you apply for all the area.
10:24
So it is generating a depth and so that will not be a good idea to have single set of coefficients to estimate all over the data. So I use these special residual clusters to produce a classified map.
10:50
So I thought that class can be used for better estimations. So in GRASS we have used ImaxLik to estimate supervised classification to use estimate
11:09
the bottom classes and this signature we have selected the area where the changes in the residuals. And the graph here, bivariate scatter plot here showing that on the x axis transformed
11:26
each band and the LIDAR depth and you can see that each class is having different type of scattering so that means it could be effectively addressed by having a different
11:40
coefficient instead of having a single set of coefficient. So after that after estimating the depth I have set the residual map again by the class based model which is better than the previous which is uniformly distributed and some of
12:02
the areas having still problem but it is far better than the previous and the correlation coefficient and RMS is shown on the table. The global model is giving 0.88 correlation coefficient but the class based model which is giving 0.94 correlation which is good and keep in mind that idea we have introduced
12:27
it is not it is an existing algorithm which is being used for so many land based regression analysis which is called GWR model geographically weighted regression. In case of GWR model it is a weighted multiple linear regression the other case it was just
12:45
multiple linear regression. So here we make a kernel like this and each kernel in the kernel you will have a centroid point you estimate coefficient for this centroid point by weighting your adjacent points.
13:02
If you the weight from this point will come as more weight for this point and if you go away from the centroid point the weight will be less so that if you have a spatial correlation you will get a good correlation from the spatially weighted regression.
13:23
So bandwidth is the main factor here if you have more denser data sets the bandwidth can be less I mean it can be a small circle which can give good accuracy results in GWR model. If your ground truth data to calibrate the results is very sparse then very less number
13:43
of then your kernel will be radius of the kernel or bandwidth of the kernel will be big and it will not give very good resolution but not very significant reduction in the results and this is the equation used to weigh the ith point I mean ith point here.
14:07
So that way it will create estimate coefficients for each pixel in the data so that you will have depth for each pixels that is what the algorithm is mainly doing.
14:24
So in results and comparison we have just made different scenarios to compare the results and in that mainly we have focused the availability of the in-situ depth to estimate or calibrate the depth. In that case you can have lower resolution in-situ depth to estimate some time you will
14:44
get lower resolution and the other time you will get more good dense data and it will be having only for small area and you want to estimate for the other area kind of extrapolation. So these kind of scenarios were tested and evaluated the result.
15:03
So in the first scenario there are randomly distributed data at that point for in case of Landsat 8 we have used 10,000 points and in case of Rapido 60,000 points were used and the evaluation is carried out by different points not the points we are used
15:23
to estimate the depth and the correlation coefficient we are showing is global modeling is not giving very good results but GWR modeling is improved significantly in terms of RMSE and R square and all.
15:44
And this is the bivariate scatter plot for each model which is also showing that GWR model is far better than the global regression model and the profile also showing in the green profile is global model and blue one is the Lidar depth 1 and GWR is very closer
16:07
and global model is little bit, it is around 2 meter, 3 meter difference you can see everywhere. This is the second scenario which we have selected 1750 points to examine the performance
16:29
of the algorithm and the interval between each point is 300 meter, you have 300 meter interval points so that but the global regression model will not change much but
16:42
even though the GWR model gives good accuracy here, it is little bit reducing the accuracy compared to the scenario 1. Still the GWR model is giving far better results compared to the global model in
17:00
scenario 2 also. This is the third scenario which is used 600 meter interval and which is also giving the same kind of trend of results and the main thing is only 450 points we were used to estimate the depth and this 450 point means it is very less because the coastal
17:24
area is around 14 kilometer long so 450 meter is very very less number of in-situ depth you need to generate reliable depth estimation. So, this is the bivariate scatter plot still showing same trend in the bivariate scatter
17:45
plot and cross-section profiles and here we need to extrapolate the estimation because if you have small area in-situ depth you want to estimate for all the scene where you have a coastal area that then I have made classes for all area like that and each
18:07
classes estimated coefficients from this points and that coefficients I applied all the other areas. So, that is also giving better model which is applied at a time you run global model
18:21
and you have the other one you have different different coefficients for all the area so that is also giving better accuracy than the global model so this is the bivariate scatter plot of that study and then we wanted to compare the cross profiles of which is
18:46
drawn on near shore cross profile spaced 2.5 kilometer away and these profiles also showing good results in the GWR model and the conclusion is that 12 bit dynamic
19:06
ranges of rapid eye and Landsat data can be effectively used for good estimation of depth from the near shore and global regression model is not able to address the heterogeneity
19:22
of the data and meanwhile the GWR can address the problem effectively and here after the algorithm which have been developed will be created as a module in the future studies.
19:42
Thank you. Question? Actually, I will ask a question. I am a member of GAPCO so I really enjoy your presentation.
20:04
I wonder what is the optimum size, optimum diameter of your kernel? It depends upon your in-situ depth which you have been used for calibration.
20:27
If you have good denser data sets, you can have, now it is 50 meter for this data sets, 50, not 50 meter, 50 pixels means 1 pixel is 5 meter then 50 into 5 that much is the
20:46
radius of the kernel and if you have large number of data set, this is just 10,000 point I have used. I have lots of points if you have large sets, the bandwidth will be reduced. I think your phrase can be converted to any formula or simple word, simple formula
21:10
then might be clear because if it is denser, the shoulder diameter may be okay so there should be some mathematical relationship.
21:25
So that is my comment. Is there any more? No more questions? Okay. Thank you, Vigner. Thank you.