Land cover classification using freely available multitemporal SAR data (work in progress)
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 | 237 | |
| 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 | 10.5446/57209 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
|
FOSS4G Argentina 202147 / 237
12
15
16
23
26
36
44
52
53
54
59
72
90
99
114
121
122
123
124
127
129
130
139
154
155
166
203
204
219
223
224
230
00:00
Sheaf (mathematics)TrailApplied mathematicsCovering spaceStudent's t-testDegree (graph theory)Medical imagingArithmetic progressionMathematicsObservational studyUniverse (mathematics)Meeting/Interview
00:52
Game theoryMusical ensembleMilitary operationPolarization (waves)Duality (mathematics)Asynchronous Transfer ModeObservational studyVector potentialSample (statistics)ForestComputer-generated imageryLevel (video gaming)GradientAngleIncidence algebraGamma functionSocial classSampling (statistics)Exception handlingMathematical analysisMedical imagingCharacteristic polynomialDuality (mathematics)Object (grammar)AreaDifferent (Kate Ryan album)Latent heatTime zoneMereologyDenial-of-service attackPairwise comparisonLink (knot theory)Electric generatorWater vaporFlow separationArithmetic progressionCovering spacePhysical systemPolarization (waves)Multiplication signMatter waveStability theoryHypothesisMechanism designElectromagnetic radiationSpektrum <Mathematik>State observerProfil (magazine)GeometryKernel (computing)SurfaceType theoryStandard deviationHypermediaScatteringObservational studyMotion captureSource codePoint cloudBitIntegrated development environmentSoftwarePresentation of a groupSelf-organizationComputer virusNoise (electronics)InformationPreprocessorLevel (video gaming)OrbitExpert systemForestTotal S.A.Workstation <Musikinstrument>ExergieLine (geometry)Network topologySystem callRoutingInteractive televisionFamilyEndliche ModelltheorieSign (mathematics)Metropolitan area networkMachine visionVirtual machineDecision tree learningGame theoryOrder (biology)Green's functionElectronic mailing listArmTable (information)SpacetimeData storage deviceComputer animation
09:32
Arithmetic meanSimilarity (geometry)Polarization (waves)PixelDuality (mathematics)AlgorithmForestEndliche ModelltheoriePrice indexCohen's kappaMultiplication signArmResultantProfil (magazine)Different (Kate Ryan album)InformationContext awarenessNetwork topologyForestMoment (mathematics)CoalitionArithmetic meanGroup actionVideo gameEndliche ModelltheorieLevel (video gaming)Self-organizationDenial-of-service attackRing (mathematics)Materialization (paranormal)Set (mathematics)Social classWebsiteSoftware bugInternet forumPoint (geometry)Texture mappingAreaObject (grammar)MeasurementImplementationTheoryFamilyCollisionLine (geometry)MathematicsTotal S.A.Complete metric spaceSelectivity (electronic)Polarization (waves)Wave packetWater vaporHand fanMusical ensembleContrast (vision)Vapor barrierCross-correlationBasis <Mathematik>Office suiteComa BerenicesExtreme programmingAnalogyDressing (medical)Goodness of fitPixelVariable (mathematics)StatisticsPattern languageMedical imagingCohen's kappaMathematical analysisCASE <Informatik>Standard deviationInclusion mapCorrespondence (mathematics)RandomizationLibrary (computing)Software testingEntropie <Informationstheorie>Subject indexingPreprocessorCalculationCovering spaceType theoryMatrix (mathematics)GradientSimilarity (geometry)Distribution (mathematics)Computer animation
18:12
Domain nameExecution unitNetwork topologyMedical imagingTerm (mathematics)InformationCellular automatonLattice (order)Materialization (paranormal)Self-organizationCondition numberLibrary (computing)Characteristic polynomialContext awarenessForest19 (number)CASE <Informatik>Multiplication signComputer clusterTrailGoodness of fitDifferent (Kate Ryan album)Combinational logicAuthorizationWeightBitGame theorySatelliteQuicksortVirtualizationBit rateObservational studyOpticsSoftwareFocus (optics)Cartesian coordinate systemoutputTexture mappingMatter waveType theoryPixelVariety (linguistics)Frequency1 (number)Online chat
Transcript: English(auto-generated)
00:14
Hello and welcome back to the academic track of Phosphor G 2021. We are about to start the last talk of this session.
00:24
The first day of the conference and the next speaker that I have the pleasure to introduce is Mariela Reinberg. She has a degree in applied mathematics. Currently, she is a doctoral student at the laboratory of remote sensing and ecology at the University of San Martin in Argentina and
00:42
She's currently working on the classification of vegetation cover in wetlands using SAR images and her talk Is titled land cover classification using freely available multi-temporal SAR data work in progress Mariela over to you for the presentation
01:03
Hello. Thank you very much, Marco. First of all, I want you to thank the organizers, the Phosphor G organizers Well, this is today I'm going to present a joint work made together with Professor Grinson, Juan Luca Hvali, Priscila Minotti and Patricia Candos. We work in the Institute of investigation and
01:24
environmental engineering at the University of San Martin in Argentina. As Marco told before, I'm going to present this work Which title is land cover classification using freely available multi-temporal SAR data. So first of all, I wanted to start
01:42
Telling you a little bit about how SAR synthetic aperture radars work. These are radar systems that usually go over Aircraft or spacecraft and the way they Generate the images is by sending an electromagnetic wave that interacts with the earth's surface and
02:04
after interacting with the surface a portion of this electromagnetic wave Returns to the SAR and the SAR captures this echo These are active systems. This means that they do not need from external sources
02:21
to get information so Sorry So they can take images during the day and also during the night. Also these electromagnetic waves operate in the microwave part of the electromagnetic spectrum and
02:42
this makes them possible to take images even when the Earth's presence of clouds or rains These electromagnetic waves Will be Will interact with the media and will have different
03:03
scattering mechanisms according to the presence of water through the geometry of the Of the observed surface and other characteristics. Here I wanted to show that perhaps we are seeing some vegetation and when we have Flooded, the same vegetation, the mechanism is going to be different and we are going to have
03:25
differences in the images. The images, the SAR images that I'm going to use are Sentinel-1 that are offered by the Special Agency, sorry, the European Special Agency
03:42
They are available since 2014 and 2016, the Sentinel-1b It's the first time that we have freely SAR available data with a short revisit time This instrument is a C-band. This means that the
04:00
the wavelength is about 5 GHz And it supports a dual polarization Vb and Vh is what I'm going to be using If you want to download or get more information, you can visit this link in the comparison
04:20
So which is the objective of my study? I'm going to Study how we can use these SAR multi-temporal data to generate vegetation cover maps in a specific place in Argentina that is characterized by being a wetland and very densely vegetated wetlands
04:42
My study area here on the left, I'm showing the southern part of South America and Here is Argentina. Here is located my study area and on the right side We can see a portion of the lower belt of the Parana River
05:00
This wetland is characterized by being having a humid and temporary climate and it's also, this zone is submitted to tidal regime so the floodings are hardly to predict. We cannot predict when they are going to occur. There are some exceptions
05:26
Specifically, I'm going to work with these over these islands and taking into account previous works, we expect the profile like the one that I'm showing you on the right
05:42
What I want to say here is that if we walk from the water to inside of the different islands, what we expect is first to have amount of sediments and have a leaves and over that leaves have different types of forest. We suppose or we expect that the dominant class of years are going to be willow plantations
06:04
then when we go through the island we can find other types of forest and like stable forest. Stable forest can be find isolated or it can be find with underneath vegetation like for example, Corpadera marshes that are these that I'm
06:23
showing you here. And when we get the center of the island where the soil is usually saturated we can find these Corpadera marshes. Also in the borders of the water courses we usually find these Hunko marshes that with this vertical structure.
06:46
So the hypothesis of this work is that in this area there are five classes that dominate the sea. The water and the four vegetation classes that I mentioned before.
07:01
To continue with this analysis what we did was we took about 100 samples per class. This means that I have 100 places that are being labeled with each of the classes and I'm going to use this information. To get these labeled places what we did was to use previous works in the area, experts of invigidation in the area and we also use a
07:27
two planet scope images. So I have the level data or the sample data and I'm going to use also this Sentinel-1 data.
07:40
Here in the images I'm showing two examples of different polarizations the VH and the VB that I'm going to use. I'm using a total of 76 images corresponding to October 2016 to April 2019. In both polarizations and here are more details.
08:04
When we work with solar images we have to do a lot of pre-processing and in this study we followed the Philippines workflow proposed in her work Sentinel-1 GRD pre-processing workflow where he proposed first to apply the orbit to each of the images, then make some noise removal
08:27
associated to thermal and to border removal, do a calibration, speckle filtering, terrain flattening, terrain correction and the last place we've got for every
08:41
polarization and for every day pixels values that represent the backscatter value for each data and polarization. All this analysis was done in the SNAP software that is freely available by the European Space Agency. Here I wanted to show a little bit how these classes
09:10
this means scattered values classes so for each date and each of the sample data I calculated the mean scattered values and also the deviation. Here we can see that there is one class that is that
09:23
this over here that's a little bit separate from the others. The other three vegetation classes show some difference in some dates, for example here or we can find some difference over here. This difference
09:41
occurred during winter but when we plot also the standard deviation we see that they have a lot of overlap. So identifying these classes seems to be hard and difficult to do. At this point I was only using the temporal information
10:01
but I also wanted to use for each pixel the spatial relationship with their neighbors. Sometimes it can happen when we are studying two areas of interest and when we calculate some statistics like the mean values from each of the areas of interest they have the same mean value.
10:22
But the distributions of the values in each of them could be different and that is what we want to capture with the textures measurements. They are going to give us information about the context of each pixel. One way to calculate texture measurements are those based on the grade level co-occurrence matrix.
10:42
They are very used in remote sensing and they are already implemented in snap toolbox. There are a lot of texture measurements that can be calculated based on the GLCM matrix. We decided to follow Hallweyer's
11:00
works and she proposed to use the contrast, variance, entropy and correlation texture measurements. Once we did all these calculations for the pre-processing we create different data sets. We wanted to evaluate which data set was useful for
11:20
creating vegetation cover maps in this context. So we create a total of 30 data sets. Each data set has a corresponding set of dates, corresponding type of prioritization and a corresponding pixels value. This means that we could use for each data set just the backscatter values or the backscatter values
11:45
together with the GLCM values. What I mean here is together with the textures values. All this analysis was done using Python, using GDAL library in Python and what there were a lot of libraries like Numpy,
12:02
Pandas that were also used in this analysis. Then once we create all the different data sets, we want to make, sorry, we want to make classifications with each of them. So we decided to use the random forest classifier because it's
12:26
very well known that it gives good results if it's resistant to out layers and it can be trained with small data sets. So for each of the data sets we
12:41
create a random forest model. This random forest model was created by selecting by doing a grid search and then once when we have the best random forest classifier trained for each of the data sets, we evaluated over the same test set all the models and then we
13:03
the way that we evaluated was analyzing the overall accuracy and the kappa index values. All these in all these models adjusting and implementation were done using the scikit-learn library in Python. And as a result, one of our
13:26
one of our objectives was to analyze if it was useful to use textures. And what we noticed is that in all the cases using the textures give us better results. When we compare
13:41
data sets associated to intensity values or backscatter values, that gave us good results like kappa index is higher than a 0.9. And when we applied to them, or sorry, the textures, we noticed that they got better results, but the results were like at least a 3% better.
14:06
In the cases that we had using the backscatter values and we got kappa index values lower than 0.8, when we incorporate the textures, the increment in the kappa index values and overall accuracy was
14:23
like at least an 8%. Great. Here, I wanted to show you some classifications. In all the cases, I'm using just the BH polarization for the first, the summer dates, the winter and the complete set of dates.
14:41
In all the cases I'm using for the classification, I use the variables associated to the backscatter and also to the textures. And as we can see here, the summer data sets show a very noisy pattern. It doesn't seem to show what we were expecting in this profile that I showed you before.
15:04
On the other hand, in the winter and in the complete data set, we can find that pattern if we work from the water to the center of the islands, we can find first an area that is covered by different kinds of forests and then we have the corporate data marshes. And in most of the
15:25
water courses, we can find hunk. Also, we can observe that winter and complete data sets give us very similar kappa values and overall accuracy. Also, when we see the classification, it's hard to find some differences.
15:44
There are some differences, but there are a few of them. Another analysis that we decided to do is that when we use random forests, it is possible to make a variable
16:03
analysis, variable importance analysis. This means that it's possible to score the data sets variable according to its usefulness in predicting the target classes. So, we can use random forests to get a ranking and that ranking is about
16:22
the variables that were used in the model which were mostly useful to which are the less useful. And what we observed when we did this analysis over the complete data set in the BH polarization was that the most useful dates were associated to
16:45
intensity of backscatter values in the winter or autumn dates. And it's not in this slide, but the lowest of the less useful variables were associated all with the entropy texture. So, perhaps incorporating this texture is not so useful.
17:08
So, as a conclusion, in this analysis, we saw that using the winter data set gave us results as good as using the complete data set, but what is good of using just the winter dates,
17:24
multi-temporal winter dates, is that we are using less and less amount of images. Also, the inclusion of textures give us better results and what we have to analyze is if it was necessary to include this
17:43
entropy texture measurement. Also, for future work, we would like to to work also with a multi-temporal L-band data that is very suitable for this type of wetlands. So, we hope now with SAOCOM, we could have this data and do this analysis.
18:05
Here, I put some information about the bibliography of your interest and I think that's all. Thank you very much. Thanks a lot, Mariela, for this very interesting talk and application of SIR images.
18:25
I invite the audience to ask questions to Mariela. In the meantime, I just read the first one that came a while ago, and it was about Random Forest. In which software did you run the Random Forest classifier?
18:43
Okay, I used the scikit-learn library in Python. Thanks. This, I think, answers the question. Second question, a new one.
19:01
Why do you think, so this is related to the to the conclusion, so why do you think that including texture features improves classification accuracy? I think that it improves because I was using a per-pixel strategy, so I was not taking advantage of the context of each pixel and it happens many times that when we
19:24
observe forests, the information is there, is that the difference between their neighbors? So, I guess it's the spatial relationship. Thanks a lot, Mariela. I hope this answers the question. This is not definitely my field, so I hope it does.
19:46
Otherwise, of course, we invite the people who ask questions to further ask questions in the chat. Okay, other questions coming. Can the snap pre-processing steps be included in your Python workflow?
20:03
Yes, I tried with a Python library called Pyrosar that wrapped that pre-processing. So, if you would like, it was really fast to do it with the library, so I guess it will help. Good, so that you might consider this for the future work.
20:25
Another question in the meantime. Did you consider to incorporate optic satellites or optical images, I guess, to the analysis? Yes, I was main focus in SAR and I came
20:41
with the SAR. Okay, I was thinking about SAR, but I will incorporate. They are sometimes like in summer, but the thing happens with short wavelengths like C-band that in summer there's a lot of vegetation and information about the underneath vegetation. It's very difficult to get with optical images. That was why I was
21:03
interested in SAR, but I guess that making a combination of a little bit of SAR and optical would be great. Yeah, indeed. So, there are no other questions, but I still invite the audience to ask if they wish. So, in the meantime, maybe I can ask something just out of curiosity.
21:25
You used the images from 2016 to 2019. I was just wondering whether there is a reason for that and not for using more recent ones or just to extend the interval and the second one is really curiosity. So, you use Sentinel-1 datasets, but I was wondering
21:45
what would happen if using other images and why this is just an experiment, if this is due to the specific characteristics like the high frequency rate or like the open access to the Sentinel-1 images or whether there are other reasons for that. Thanks.
22:03
I was interested in using Sentinel-1 in particular because first they are free. I work in a laboratory that explores or monitors wetlands. So, SAR images are very important and using some, perhaps if I use another type of image like ALOS-2, for example,
22:25
then I want to replicate this study in other places and I cannot find the available data. So, what is great, I think, with Sentinel-1 is that I can cover all the places that I want and I forgot, sorry, the first one and okay.
22:45
Well, something, it's like this. In this case, I download all the images and I started working with Sentinel-1 in 2019 and when I took to April, in April, June, July, I couldn't find images from this place. Then I could get again images,
23:07
I think, in September 2019. And so I did like stop there. I analyzed if just need winter or need summer or all those things. But well, I guess now I'm going to include 2020 that
23:21
there are the images and 2021 too. Good, thanks a lot for the answers. Mariella, looking at the chat, I do not see additional questions. So, I think we can wrap up and close not only the talk, but the whole academic track session.
23:43
It's been a pleasure for me to chair this session. I would like to thank Mariella and the audience, Mariella for the great presentation and for answering the questions and of course the audience for their good questions and input to the discussion. But let me also take the chance to thank all the other authors and whoever has
24:01
attended this academic track session of Phosphor-G. Once again, I think all the talks we have heard about prove how mature, how strong, how good Phosphor-G solutions are to support research in a variety of domains. This is the thing that I really appreciated the most.
24:22
So, if you, to the audience, if you like the talk, please put your hands together in a virtual applause for Mariella, as I'm doing right now and see you virtually in this exciting Phosphor-G 2021 and
24:44
enjoy your afternoon or evening or night or day or whatever, depending on where you are. Thanks a lot and bye-bye.