Inter-comparison of the Global Land Cover Maps in Africa Suplemented by Spatial Association of Errors
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 | 295 | |
Author | ||
Contributors | ||
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 | 10.5446/43554 (DOI) | |
Publisher | ||
Release Date | ||
Language |
Content Metadata
Subject Area | ||
Genre | ||
Abstract |
| |
Keywords |
00:00
3 (number)Student's t-testUniverse (mathematics)FamilyLecture/Conference
00:21
Covering spacePairwise comparisonAssociative propertyPrototypeNumberImage resolutionProduct (business)Texture mappingSet (mathematics)MetrePrototypeAssociative propertyMappingCovering spaceResultantNumberSystem callCartesian coordinate systemBitDifferent (Kate Ryan album)InformationState observerProduct (business)MereologyValidity (statistics)Order (biology)Mixed realityComputer animationXML
01:42
Covering spaceMereologyValidity (statistics)Order (biology)Cartesian coordinate systemMultiplication signProduct (business)Texture mappingNumberMixed realityMeeting/Interview
02:14
Covering spaceNumberImage resolutionProduct (business)Pairwise comparisonSubject indexingSimilarity (geometry)Maxima and minimaCross-correlationSocial classPrototypeSample (statistics)PreprocessorPhysical systemPixelRaster graphicsTrailSingle-precision floating-point formatOverlay-NetzGrass (card game)PixelImage resolutionPrototypePairwise comparisonTexture mappingSocial classGrass (card game)Covering spaceReference dataTable (information)CountingProcedural programmingDerivation (linguistics)Subject indexingMappingOverlay-NetzProduct (business)MetreSimilarity (geometry)PreprocessorPattern languageProcess (computing)Order (biology)Mathematical analysisExecution unitResultantSlide ruleCross-correlationNumberPhysical systemPosition operatorTotal S.A.Statistical hypothesis testingPoint (geometry)TrailValidity (statistics)Field (computer science)Raster graphicsMatrix (mathematics)Digital photographyInterpreter (computing)SatelliteDialectSet (mathematics)Term (mathematics)Mixed realityMereologyPresentation of a groupRadical (chemistry)Series (mathematics)SpacetimeProjective planeBridging (networking)Degree (graph theory)Time zoneComputer animation
08:28
Subject indexingSimilarity (geometry)PixelWeightPairwise comparisonCross-correlationSurfacePrototypeForestForm (programming)Sign (mathematics)Associative propertySocial classFocus (optics)Covering spaceImage resolutionStudent's t-testMappingAssociative propertyPrototypePairwise comparisonSocial classTable (information)Row (database)Procedural programmingTotal S.A.Multiplication signPixelTexture mappingGrass (card game)Figurate numberCross-correlationPoint (geometry)Similarity (geometry)ResultantNumberCombinational logicNetwork topologyContext awarenessBit rateGraph (mathematics)ForestProgramming languageFunctional (mathematics)BitPresentation of a groupTransportation theory (mathematics)Mathematical analysisInformationFood energyDemosceneDreizehnGene clusterGraph coloringUsabilityMixed realityOrder (biology)Morley's categoricity theoremProcess (computing)Metropolitan area networkAreaImage resolutionWeightDifferent (Kate Ryan album)Covering spacePattern languageSet (mathematics)Numeral (linguistics)Validity (statistics)WhiteboardSoftware developerSubject indexingSoftwareComputer animation
14:42
Point (geometry)Procedural programmingMathematical analysisMultiplication signDescriptive statisticsNeuroinformatikCombinational logicCross-correlationHypothesisAreaSampling (statistics)MereologyProjective planePairwise comparisonMappingSubject indexingCovering spaceReference dataPrototypeInstance (computer science)Moment (mathematics)Control flowSocial classBuildingProcess (computing)Latent heatReading (process)UsabilityJSONXMLLecture/Conference
Transcript: English(auto-generated)
00:07
Hello everyone, my name is Goritsa Bratic, I'm a Ph.D. student and temporary research fellow at Polytechnic University of Milan and it's my pleasure to present to you today our work that I did together with Professor Maria Brevely and engineer Daniela Oxoli
00:27
that is called Intercomparison of Global Land Cover Maps in Africa, Supplemented by Spatial Association of Errors. So first of all I would like to give you some clues what I will talk about today.
00:41
So in the introduction I will present you what is our motivation behind this work. Then I will describe a dataset, these datasets are global N30 and S2 prototype LC 20 meter map of Africa that I will call CCI Africa prototype for the simplicity. Then I will present methodology that was a bit different than traditional methodology of intercomparison
01:07
and then I will show the results and finally draw some conclusions. So in the past years the number of land cover maps have been increasing and that's because the high demand for these maps, because these maps provide very useful information for different applications.
01:27
On the other hand, earth observation technologies that are behind these land cover maps are also improving so they allowed technology to increase the number of land cover maps.
01:43
One part of production of land cover map is validation and it is crucial in order to determine if map is accurate enough for certain application. However, by increasing the number of land cover maps we are also increasing the efforts that we need to put to validate them.
02:02
And traditional way of validation is quite time consuming and also like costly and by intercomparison we can reduce this time and also money that we spend for validation. In general for validation the biggest problem is to collect reference data
02:25
and we usually need to do the field work or photo interpretation of high resolution satellite imagery while with intercomparison we can reduce this by comparing one land cover map with another land cover map and then in this way we can identify which are the most critical regions
02:44
which are the regions which don't agree on both of the maps and then focus validation on these parts. So I would call intercomparison as a preparation for validation. The scope of our work was to compare two recent high resolution land cover maps in Africa
03:02
and we restricted the region to the country of Rwanda. So what we did here, we used traditional accuracy indexes but actually we didn't determine accuracy we just determined similarity between the maps but the indexes in literature you will find them called accuracy indexes
03:27
that's why I call them like that. Besides we explored spatial variabilities and patterns of this agreement. Then data set processing was suited to obtain numerical values of the disagreement
03:46
because in general the values related to land cover maps are categorical but we needed numerical values to perform spatial analysis that we did here because spatial analysis methods are very few for categorical data.
04:05
Then the results showed that similarity between the two maps is not very high and also that the spatial correlation of the disagreement is strong. So the two data sets we used are global N30 and CCI Africa prototype.
04:27
Global N30 is produced by National Geometric Center of China. It has resolution of 30 meter because it is produced from Landsat imagery that has a 30 meter resolution. Imagery is from 2015 and it has 10 land cover classes.
04:46
Then we have CCI Africa prototype that is a product of climate change initiative of European Space Agency. It has 20 meter resolution for year 2016 and like global N30 it also has 10 land cover classes.
05:06
We needed some pre-processing before doing proper inter-comparison. The two things were common for pre-processing of the data. The complete pre-processing was done with GRASS GIS. This is the first similarity and the second similarity is that they were both projected coordinate system
05:30
in order to have map units in meter because originally they were in WGS 84 and in degrees. Then further pre-processing steps were done for CCI Africa prototype.
05:42
This was resampling to 10 meter resolution and then reclassifying to match the classes of global N30 so that we can compare them. This slide is explaining how we did inter-comparison.
06:01
Intercomparison we did is, as I mentioned at the beginning, is different from the traditional or common intercomparison because what we did here is we compared single pixel of one map to the multiple pixel of another map and common procedure is to compare pixel by pixel, one by one.
06:23
So we call this procedure sub-pixel comparison and we compared one pixel of global N30 to the nine pixel of CCI. Here it is visible why we resampled CCI Africa prototype to 10 meters so that we can have exactly nine pixel to compare with one pixel of global N30.
06:49
Then another interesting point here is that we needed additional, I would call it artificial raster, to keep the track of the position of each global N30 pixel because with our procedure we couldn't have that simply without this raster.
07:08
This raster had the same size and resolution as global N30 but the values were from one to the total number of pixels of global N30 distributed row-wise. Here we can see illustration of the procedure on example of one pixel of global N30.
07:30
So what we can see here, we overlay these three maps to compare them. We have value of ID 15 which means that that is order 15 of this pixel of global N30.
07:51
Pixel of global N30 has value 30 while we have five pixels of CCI Africa prototype that are exactly the same class as the global N30 pixel and then four pixels that are different classes.
08:05
We wrote all of these in the first table there, then we aggregated this table by the ID numbers, by the same ID and we got a count of each class in the second table for one ID value.
08:25
This second table when we did the procedure for whole map, it was useful to derive error matrix for computing let's say similarity indexes. The third table is representing each row we divided by the total of the
08:45
row to obtain percentages which are numerical values which we used for spatial association analysis. We called the last table sub-pixel disagreement table and the whole
09:02
procedure was done in combination of GRAZ GIS functionalities and Python programming language. So the traditional accuracy indexes that we computed which now maybe doesn't seem good because they are actually in this context they are similarity indexes, however they are called overall accuracy, user's accuracy and producer's accuracy.
09:26
For spatial association we computed global moron eye from sub-pixel disagreement table and for computing global moron eye we used weights based on eight nearest neighbors.
09:41
So first result is the result of so-called producer's accuracy that is agreement between global N30 and CCI Africa prototype in each class measured with respect to the total number of pixels per each class of global N30. So in this graph we can see that similarity between the classes in two
10:08
maps is highest when it comes to the classes of cultivated land and water bodies. We had a bit of a problem in the classes of tundra and berlin because they were not at all present in the global N30 but this is probably the problem of matching different classification legends.
10:26
And we didn't simply have enough information to do this matching better. Then the two classes that have low similarity are shabland and wetland and they are present in both of the maps.
10:40
And shabland in the CCI is mostly confused with grassland in GL30 while wetland is mostly confused with forest in GL30. Furthermore we have user's accuracy that is agreement between global N30 and CCI Africa prototype in each class measured with respect to total number of pixels per each class of CCI Africa prototype.
11:07
So here again the results for user's accuracy are quite consistent with the results of producer's accuracy again we have high similarity for cultivated land and water bodies and low similarity for shabland and wetland.
11:21
But this time shabland and wetland both are confused with cropland in the CCI Africa prototype. Overall accuracy was accounted for 66% which means that two maps are not very similar and the classes with the highest confusion which were wetland and shabland are analyzed further.
11:44
Global Moroni was 0.74 which indicates strong spatial correlation of the disagreement and also we identified significant spatial clustering in the mismatching pixels.
12:01
This is the figure that is showing total disagreement between GL30 and CCI Africa prototype and this figure is showing that stronger patterns are on the east side and the northwest side of this map.
12:23
And the darker is the color the stronger is the disagreement between the maps. And then this is a map of disagreement for the two classes. Wetland that is in blue and shabland that is in green and we can see that probably this is also shown on
12:42
the map of total disagreement because again here we see the strong patterns on the east and on the west of the map. And finally I will draw a few conclusions. So our work was focused on intercomparison of two land cover maps of Africa that are GL30 and CCI Africa prototype.
13:11
Accuracy assessment techniques were applied in order to determine similarity between the maps and spatial association of disagreement between the two maps was analyzed.
13:23
Methodology applied here allowed comparison of the data set with different resolution and also we found a way to turn categorical data into numerical data. Similarity between the two of the maps is rather low and wetland and shabland are the most critical classes with the highest disagreement.
13:42
And the areas we identified with our analysis we see it as areas which require further attention, further analysis maybe to do validation there or maybe also some other analysis. That's it.
14:09
Personally I was happy to see Gras again. As a board developer it's nice to see that the software is useful and used. So please I hope that you have questions.
14:42
It was just, it doesn't have any special reason we just chosen one country to analyze because also I forgot to mention this procedure that we applied is quite like time consuming so we couldn't apply for whole Africa so we have chosen one country which didn't have, not for a specific reason.
15:07
Enough small also yeah.
15:20
Yeah this was a tricky point because for global N30 we have some description of the classes not very detailed but some and for the CCI Africa prototype we had only names. So we were forced just to use the names and do our best to match because we couldn't find and this is often a problem with dealing with comparison of land cover data.
15:51
Yes.
16:30
We didn't do analysis for that specifically but this is always true for land cover data like they will be correlated because they are already only in one portion let's say but we didn't do specific analysis here to compute indexes or something like this.
17:40
Well further work beside this paper was to analyze further the data so I did for each confusion for the let's say for each confusion for the class combination I tried to analyze with some let's say I compared it with Bing maps and everything but so far I didn't find anything peculiar but this were analysis from the point of spatial correlation.
18:08
And I didn't compare it with any reference data hypothesis for this work were that if the two maps have agreement in the same area it is high probability that it is true.
18:21
And then the part of disagreement is part that we would recommend to apply further analysis but we didn't do this.
19:16
It is difficult to find the ground truth for a certain region because for other
19:22
projects we try to find other data like roads in Africa where you find the data. At the end of the day reliable comparison is probably the only way because even if they are validated the sample is very limited.
19:51
It's just that for instance in the region like Rwanda it could be 200 points so you have to decide if it is valid or not.
20:09
So there is at the moment there is not as it is worth it is good to try to share
21:07
our experiences outside yeah so to be informed in such a way that because we are we are going there.
21:26
So thanks a lot we have a few minutes for a break that you can change the rooms and for the newcomers this is academic track.