Visualizing two decades of land use changes in Europe
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 | 57 | |
| Author | ||
| License | CC Attribution 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 purpose as long as the work is attributed to the author in the manner specified by the author or licensor. | |
| Identifiers | 10.5446/55260 (DOI) | |
| Publisher | ||
| Release Date | ||
| Language | ||
| Producer | ||
| Production Place | Wageningen |
Content Metadata
| Subject Area | ||
| Genre | ||
| Abstract |
| |
| Keywords |
11
32
38
40
50
53
54
57
00:00
Social classCovering spaceModul <Datentyp>Module (mathematics)Process (computing)ResultantRow (database)Personal digital assistantVirtual machineEndliche ModelltheoriePoint (geometry)System callVisualization (computer graphics)Recurrence relationOverlay-NetzFood energyInsertion lossoutputWordMultiplication signNetwork topologyMathematicsMereologySocial classInformationSign (mathematics)Temporal logicMusical ensembleSpacetimeLevel (video gaming)Variable (mathematics)CodeOpen setQuicksortComputer animation
02:47
Organic computingGroup actionBitDevice driverComputer animation
03:10
Organic computingMathematical analysisField (computer science)Helmholtz decompositionInternet forumGroup actionFamilyBitMultiplication signPosition operatorTwitterForestPixelScaling (geometry)Graph (mathematics)MetreSound effectLevel (video gaming)Type theoryPrice indexEndliche ModelltheorieSquare numberLinear regressionAverageMedical imagingImplementationArithmetic meanFood energyCausalityShared memoryGastropod shellForcing (mathematics)Vector potentialGrass (card game)Product (business)Execution unitMathematicsAmsterdam Ordnance DatumComputer animationDiagram
06:21
Landau theoryModule (mathematics)Router (computing)Covering spaceMathematical analysisSign (mathematics)Visual systemGoogolForestDataflowCASE <Informatik>AreaInformationData storage deviceInstance (computer science)ForestForcing (mathematics)Position operatorBitSocial classWebsiteComputer animation
08:06
Scaling (geometry)Axiom of choiceWordAreaWater vaporSocial classMereologyMathematicsEndliche ModelltheorieRight angleComputer animation
08:42
Civil engineeringForestMathematicsThermal expansionDifferent (Kate Ryan album)AreaMathematicsReference dataComputer animation
08:59
Workstation <Musikinstrument>MathematicsBitFilter <Stochastik>
09:16
Connected spaceMathematicsWorkstation <Musikinstrument>BitMathematicsSocial classComputer animation
09:32
MathematicsTemporal logicMedianSimilarity (geometry)Thermal expansionWorkstation <Musikinstrument>MereologyNegative numberAreaMedical imagingInstance (computer science)BitInformationMatching (graph theory)Rule of inferenceBit ratePhysical lawWave packetWhiteboardPattern languageAuthorizationSocial classMathematicsEndliche ModelltheorieResultantArmMultiplication signFood energyHypermediaView (database)ForestArithmetic meanBlock (periodic table)PixelMetreFilm editingMappingShooting methodNegative numberCASE <Informatik>Covering space
15:18
Mathematical analysisInsertion lossOrder (biology)Point cloudMappingComputer animationDiagram
15:48
Module (mathematics)Covering spaceLandau theoryArm
16:15
Mathematical analysisIdeal (ethics)AreaWeb pageRow (database)Image resolutionFood energyBitMathematicsHelmholtz decompositionComplex (psychology)TwitterPoint (geometry)Linear regression1 (number)AverageArithmetic meanOptical disc driveNetwork topologyPlanningMachine visionTheory of relativityDirection (geometry)Mathematical analysisVisualization (computer graphics)PixelData miningGraph (mathematics)Different (Kate Ryan album)Sound effectStatisticsComputer animationDiagram
Transcript: English(auto-generated)
00:39
Hi, I'm Chris, I'm with Open
00:40
This is a nice job of actually showing all the results showing off all the results so that's that's pretty cool. But the most of the work is actually done by Leandro and Martijn you heard speak today already. So here I want to point out the EU map where we also publish most of the code that was used to produce these things that I'm showing you.
01:03
And the preprint where some of the visuals that I show will be there in this in this preprint of the work that Martijn has been doing. So I want to talk about, sorry, I want to talk about three main things.
01:23
So the NDVI slopes because I think it is a metric that most people understand that is used for quite a while. So it's kind of a part of the input of our models, but it shows us also information about it's it's understandable and yeah it's a nice metric to look at.
01:47
And then the probability slope because I really think that's something that hasn't been done so much before. And that is really now possible with using the Ensemble machine learning approaches and then class changes that are another result of our of
02:04
our approach and kind of show what what it means, what was modeled the sort of land use that was modeled over these decades. So this is a quick overview, just to remind you. So we have the temporal lens set, then we we can apply a gap filler. So we get a nice clean data.
02:27
And that is also part of this is the NDVI data. So that is what it will now be showing that also goes in to the space time overlay and also is a variable in the machine learning that we that we apply.
02:44
So it's kind of an input variable for this model. First to get a bit of an understanding of what what can we see in these in this NDVI values is is a little farm that's not so far from here. Takes you 30 minutes to bike there. We definitely recommend you to do that if you have some time.
03:05
It's a it's a farm that transition to organic agriculture in 2009. And if we look at the lens set values for for for for this, this field in particular, we know it was transitioned from crops to organic cattle.
03:25
And we use a decomposition of the seasonality. So this was what Leandro also talked about. That's the same approach that we use here and also shows the strength of the package where we can easily implement approaches from other packages on a large amount of data.
03:44
So here we see in red the average values for the NDVI for this field. And in blue kind of average mean. And below we see if we decompose the seasonal and the trend. So the red is just the trend and the yellow is just a seasonal seasonal effect.
04:05
And if you decompose it, you can quite quite clearly see this where we change from a from a crop type to a cattle kind of just grass, much more stable type of agriculture. So that also shows us a very clear indication of what's happening here.
04:27
And then we we went a little further and just met this trend for every pixel in Europe at 30 meter scale. And so here at the top, we see the same the same graph that I just showed you just for one pixel in in Sweden for a forest that is regrowing.
04:51
Let's see, there is a little map here. So it's over here. It's it's a little forest that just has been cut and then it's regrowing.
05:01
So we see kind of a increasing trend if we decompose this this signal into just the trends and we fit a ordinary lead squares or regression on this on these values. So we get a slope value that's positive. And if we if we do that for multiple pixels, we get this kind of image where green is a positive slope and red is a negative slope.
05:29
And then here I want to show you the the probability values for coniferous forest. So actually, here you see if we apply the model for this pixel for every year that we've got field
05:44
data, then we see that the the model the the probability value for this pixel to be a forest is increasing. And also this this we can map spatially. So I think that's quite interesting.
06:00
And then I want to focus a little bit on this one to show you. While let me skip this for sake of time. So that's, that's really just the, the decompose values map for every pixel in Europe at the 30 meter scale.
06:21
Then I want to go further to these slope values of the probabilities and show you that this impacts how here we can have some deeper understanding of how the landscape is changing because we can
06:42
see this for every class that we have that we that we use so 43 or something like that. So that gives us more information about what is happening in Europe. And then for me it's a bit the challenge to
07:00
To see how, how we can visualize that. So here's another example where you, you can check. If you want to have a thing like really what does this make sense in reality, you really see that that if we have, for instance here, positive slope for in this case, barely vegetated area, then here we see that in 2009, there was a forest and then in 2017
07:29
That forest had disappeared. So it makes sense that we have a positive slope for the probability of being barely vegetated area.
07:43
So that's just another example and then not an example for a reforested site. So if you look at the situation in 2016 and 2019, we see that the the probability, the slope for the probability values For coniferous forest is actually a positive. So that that makes sense.
08:07
So then going to the to the land use classes. So In a model, the, the, the class with the highest probability is chosen as as the class. And then this is if
08:21
you visualize this for 20 years. So here we see Rotterdam where they build a new part of the, of the port. Right. Well, yeah, the most clearly visible a change from water to land, I think. So, but how can we visualize this on a European wide scale that's
08:44
That's another challenge. So we have many different changes, of course. So this is just Kind of a simplified way we can Show all the changes that we have. And this is also in the reference data that I provided and then we can get something like this. If you apply it for larger areas. So we can see
09:08
That there is some red there. So there is, of course, disappearing. But as we have seen also here, some pixels, they kind of fluctuate is also something more time mentioned. So we, we, we try to apply a bit of a filter so that
09:28
Change to a class between to
09:58
And then one way we we imagined we could visualize this
10:03
Is to look at Look at the aggregated over five kilometer block. So just to check what is the main change within every five kilometer area and then we get maybe
10:23
Yeah, surprising image of Europe where green would be reforestation and brown would be deforestation. And then we get, of course, in a very interesting discussion with the people from Sweden. But that's, that's what we see in this image. So you might I ask, can you
10:44
Yeah, can you enhance that a little bit. So yes, we also saw that was 2020 kilometer blocks, actually. So we also look at five kilometer blocks. So it becomes already a little bit more detailed. Of course, we have 30 meter pixel. So we can go much
11:03
more detailed, but this this aggregates the results a bit and shows you more of where the change you're happening, but Of course, the if you just check the prevalence change, then even if there's three pixels changing,
11:20
then it could be one of the one of these pixels on left here will become green. So it's important to look at how much, how many of the pixels in in the area has actually changed. So how, how much of this area has actually changed. So we generate these change intensity maps to really see where the changes that we met are actually happening at a larger rate.
11:46
And I think if we combine this becomes really more obvious where areas are stable and where areas are changing in whichever way that we have that come out of this model.
12:01
So I think that's a that's a very interesting result and really allows you for any any change case or Yeah, any change to really see okay what what classes are to go from a very broad view to zoom in to.
12:21
Yeah, what's actually happening here. So here we see for instance. Yeah, there's a lot of deforestation apparently here. And maybe reforestation in this area. So if for your particular research, maybe it would be interesting to look before you kind of start to look at these maps and see what you can expect in this area.
12:44
So I think that could be very useful to people. And then we can combine these three things. So we have the NDVI slopes, the probability slopes. And the actual land cover change to really get more detailed information about the area we're
13:01
looking at. So for instance, if you look at some area here in Sweden, we see that Yeah, some, some areas are probably being used for forest harvesting and some areas are really growing forest. So it's not so It's not so clear cut that it's just deforestation and the same in this area in France. And then some things we really cannot explain yet using our
13:31
approach. I think that's also interesting. So in the Alps, for instance, we see negative NDVI slopes, but we don't really see
13:42
clear signal in the probability slope. So that's really something we are looking into now, but we do see that these areas are Yeah, they kind of turning into deserts, according to the change maps. So that's something for me. It's very interesting to
14:02
look at these maps to see what is clear patterns and what can you tell me about the changes that are happening. I think that's all I got. So if there are any questions, please shoot.
14:22
Thank you. Yes.
14:59
Is there any value or do you feel like an area or one specific or new areas
15:06
which you are accomplishing and you have been averaging for a couple of weeks or how it was? Which one? Is it any? All data. Yeah, so
15:20
The maps, the most maps that I saw the question is, what do the values actually mean? Yes. So for all the maps that I'm showing. So the most of the maps that I showed, it's either NDVI slopes or NDVI.
15:41
Yeah, so we use the Landsat data. We apply a gap filling approach and we harmonize all these data so that there is no clouds or anything in it. And then the values, that's the values that come out is, yeah, that is kind of a clean
16:06
data set that in the workflow, we visualize it like this. And then on that data, I apply Yeah, I decompose the seasonal effect. Yeah, and then from that decomposed signal, I just take the trends and on that
16:29
I fit a regression. Yeah, and this that slope. That is what I'm visualizing. So yeah, it's a lot of steps. Yes, we did this for every pixel. That's correct.
16:48
This is just an example of one pixel. So there's one single pixel here where X marks it. And then this one pixel, this is kind of the signal for this pixel. So all the red dots, they are the NDVI values.
17:06
And then we decompose it using the stats model toolbox. And then we get this is just the slope, the trend. And then on that, we fit a regression of which we just visualize the slope here. So that's kind of a trend.
17:28
So also what Leandro discussed, we chose to this approach because it's relatively compute not so intensive. And so I'm, of course, there's other
17:45
approaches to do this to be fast, as mentioned, but I think it's a nice method to kind of show a direction of change. It's a very difficult approach because it's a very tedious one. So there was a very immediate area of the nature, it's different than all the cultures, and they even also organized the material and so on.
18:48
So the point you're making, I'm just repeating it for the people of mine. So the point you're making is that it's just really complex. So the graphs we're showing might be a bit too simple to look at, too simplified to visualize this complexity.
19:09
Yeah, I think that's a very good point. Thanks for making a point that I think, but I think I hope people like you or people here that they can
19:20
take this data and they can actually say, okay, I apply my own analysis on it and I can show you that it's either wrong or it's actually useful or we try to put this data out in the open so people can use it and discuss this.