European Forum of Geostatistics Bled October Downscaling population density with CORINE Land Cover Warning: This presentation is very similar to the one given in Helsinki for the Nordic forum, but: There are some new elements This time you can criticize directly the author (no swear words, please) NOTES 1. PLACE, DATE AND EVENT NAME 1.1. Access the slide-set place, date and event name text box beneath the JRC logo from the Slide Master Do not change the size nor the position of that text box Replace the mock-up texts for the place (“Place”), the date (“dd Month YYYY”) and the event name (“Event Name”) with your own texts Set it in MetaPlus Book Roman, if you own the typeface. Otherwise, keep the original typeface – Arial Keep the original flush-left justification Keep the original font colour (white) Keep the original font body size (7 pt) and the text on one single line. 2. SLIDE NUMBER 2.1. The slide number on the banner’s lower right-hand side is automatically generated. 3. SLIDES 3.1. Duplicate the first slide as needed Do not change the size nor the position of the slide’s text box Try not to place more text on each slide than will fit in the given text box Replace the mock-up heading text (“Joint Research Centre (JRC)”) with your own text heading Set it in Eurostile Bold Extended Two or in Helvetica Rounded Bold Condensed, if you own one of these typefaces. Otherwise, keep the original typeface – Arial Keep the original flush-left justification Keep the original font colour (100c 80m 0y 0k) Keep the original font body size (28 pt) and the heading on one single line whenever possible. Reduce the font body size if needed Replace the mock-up text (“The European Commission’s Research-Based Policy Support Organisation)”) with your own text Set it in MetaPlus Book Roman, if you own the typeface. Otherwise, keep the original typeface – Arial Keep the original flush-left justification Keep the original font colour (100c 80m 0y 0k). Use black if you need a second colour Keep the original font body size (22 pt) or reduce it if unavoidable Replace the EU-27 map mock-up illustration with your own illustration(s) Try to keep your illustration(s) right- and top- or bottom-aligned with the main text box whenever possible. NOTES 1. PLACE, DATE AND EVENT NAME 1.1. Access the slide-set place, date and event name text box beneath the JRC logo from the Slide Master Do not change the size nor the position of that text box Replace the mock-up texts for the place (“Place”), the date (“dd Month YYYY”) and the event name (“Event Name”) with your own texts Set it in MetaPlus Book Roman, if you own the typeface. Otherwise, keep the original typeface – Arial Keep the original flush-left justification Keep the original font colour (white) Keep the original font body size (7 pt) and the text on one single line. 2. SLIDE NUMBER 2.1. The slide number on the banner’s lower right-hand side is automatically generated. 3. SLIDES 3.1. Duplicate the first slide as needed Do not change the size nor the position of the slide’s text box Try not to place more text on each slide than will fit in the given text box Replace the mock-up heading text (“Joint Research Centre (JRC)”) with your own text heading Set it in Eurostile Bold Extended Two or in Helvetica Rounded Bold Condensed, if you own one of these typefaces. Otherwise, keep the original typeface – Arial Keep the original flush-left justification Keep the original font colour (100c 80m 0y 0k) Keep the original font body size (28 pt) and the heading on one single line whenever possible. Reduce the font body size if needed Replace the mock-up text (“The European Commission’s Research-Based Policy Support Organisation)”) with your own text Set it in MetaPlus Book Roman, if you own the typeface. Otherwise, keep the original typeface – Arial Keep the original flush-left justification Keep the original font colour (100c 80m 0y 0k). Use black if you need a second colour Keep the original font body size (22 pt) or reduce it if unavoidable Replace the EU-27 map mock-up illustration with your own illustration(s) Try to keep your illustration(s) right- and top- or bottom-aligned with the main text box whenever possible.

European Forum of Geostatistics Bled October The study area of this example is EU27+ (Croatia….) In most countries, population data are available for public use only per administrative unit (commune) In many cases this may be insufficient for geographical analysis. –It depends on the size of the geographic units and the scale of the event under assessment. –Population hit by a flood –Population in the 65 decibel contour of airports –Population at a distance > 2 km of the closest primary school. In some countries, population data exist for 1 km grids – Bottom-up approach (much better..) – for public use? (problem of confidentiality) Rationale

European Forum of Geostatistics Bled October Worldwide downscaled density exist: Landscan-ORNL (Oak Ridge National Laboratory). 30” resolution –Different concept: “Ambient population” ~ average of the number of people in each cell along the day/week/year Gridded population of the world (GPW) of the World Resources Institute (CIESIN). 2.5’ resolution. Monitoring changes ? Much better with bottom-up data, but something can be done with top-down data. Rationale

European Forum of Geostatistics Bled October Population density downscaling Starting data: Population per commune and CORINE Land Cover. Result: Approximate population density with 1 ha resolution (GIS grid) + =

European Forum of Geostatistics Bled October CORINE Land Cover Land cover map from photo-interpreted Landsat-TM images 44 classes – Urban dense – Urban discontinuous – 9 other “artificial” land cover classes For this exercise, simplified nomenclature of 9 classes Minimum mapping unit: 25 ha. – Smaller patches swallowed by dominant class – heterogeneous classes if no one is dominand (~10% of the total area)

European Forum of Geostatistics Bled October A simple model for downscaling X m : population in commune m S cm : area of land cover type c in commune m. Y cm : density of population for land cover type c in commune m. Inside each commune Y cm is assumed to be proportional to given coefficients U c for each land cover type: If we know U c, W m are computed to respect the total population of the commune Problem: estimating reasonable coefficients U c

European Forum of Geostatistics Bled October Version 1 of the downscaling method Estimating U c with an iterative algorithm: 1.Pretend for a moment that population is known only per region (not per commune) 2.Downscale with a provisional set of coefficients 3.Compute the population that would be attributed to each commune X * m 4.Compare each X * m with the known population X m and compute a disagreement index 5.Modify U c to reduce the disagreement (ask paper for details) 6.Turn to step 2 or stop if modification very small Algorithm run separately for 3 strata (types of communes) A.Communes with high density (> 2*regional average) B.Communes with low density, but some urban area in CLC C.Communes without urban area in CLC

European Forum of Geostatistics Bled October Coefficients with iterative method (version 1) Main type of inaccuracy: density overestimated for non-urban CLC classes in communes with an important urban nucleus (and under estimation in the nucleus)

European Forum of Geostatistics Bled October disaggregation: Commune coefficients. High (low) coefficients: high (low) density for a given CLC class

European Forum of Geostatistics Bled October LUCAS 2001/2003 (Land Use/Cover Area-frame Survey) Managed by Eurostat (Common specifications for EU15 ) nomenclature Land cover (57 classes) * Land use (14 classes) Two-stage – PSUs: Systematic sampling on a grid of 18 km (~9800 PSUs) sampling: –SSUs: 5x2 points 300m apart “Points” of 3 m Approximate location accuracy: 2.5 m. Carried out in 2001 (2002 for UK and Ireland) and The land Use “residential” gives information useful to assess the density of buildings in CLC non-urban classes. Introducing LUCAS data

European Forum of Geostatistics Bled October Using LUCAS 2001 LUCAS data allow to redo several steps on a more objective basis: Assessing grouping criteria of CLC classes. For example there LUCAS data suggest that there is no significant difference between arable land, permanent crops and pastures. Geographical tuning of coefficients: the density of scattered buildings in non-urban areas is not homogeneous in the EU and does not only depend on global population density. The residential area is used as proxy of the population density in CLC non-urban areas.

European Forum of Geostatistics Bled October Using LUCAS 2001

European Forum of Geostatistics Bled October % of LUCAS residential points for different CLC2000 classes

European Forum of Geostatistics Bled October Coefficients suggested by the % of residential area Version 3 of the disaggregated grid

European Forum of Geostatistics Bled October Further attempts with other methods Estimation of coefficients with the EM algorithm (Poisson model) Introduction of LUCAS information through a logit model Limiting variable

European Forum of Geostatistics Bled October Application of logit regression Assumption: the probability that a random point has residential land use depends on the CLC class and on the average population density of the commune The logit model assumes more specifically: Where Jc is an 0-1 indicator of the CLC class c

European Forum of Geostatistics Bled October Residuals of the logit regression (2001) The residuals of the logit regression can be used for the geographical tuning of the coefficients

European Forum of Geostatistics Bled October EM Algorithm Iterative algorithm:

European Forum of Geostatistics Bled October Limiting variable Steps: Attribute a homogeneous density to the commune. Apply a threshold to the least populated class (e.g. “natural vegetation. Distribute the remaining population to the other classes Same for the other classes, excluding the most populated class in the class.

European Forum of Geostatistics Bled October Limiting variable Main problem: computing reasonable thresholds for each land cover class. First approach: use density of the communes that have only one population density type. –Problem: few communes with only one land cover type in CLC. However the approach has been tested and the results are not bad. Promising if it can be improved…..

European Forum of Geostatistics Bled October Limiting variable Improving thresholds With the help of the 1km grids from Finland, Sweden, Denmark, Netherlands and Austria. Large number of “CLC-pure” 1 km cells Threshold as a function of the Commune average of population density. Ex: for the class “urban discontinuous” Y = * Dm ; Y<5000 anyhow.

European Forum of Geostatistics Bled October Validation in 5 countries A reliable reference grid available for 5 countries with 1 km 2 cells To be extended to other countries Disagreement index for map m: cell Disaggregated map Reference map disagreement of different disaggregated maps with reference data Austria DenmarkFinlandSweden Netherlands Communes (non disaggregated) CLC-iterative CLC-LUCAS simple CLC-LUCAS logit CLC EM CLC limiting variable

European Forum of Geostatistics Bled October Some conclusions Disaggregated population density maps with the help of CORINE Land Cover reduces by approximately 50% (not always) the disagreement with a reference map The logit model seems to give the best results among the approaches tested, but the differences are very small In communes that contain large urban and non-urban areas, all the disaggregated maps tested seem to over-estimate the density in non- urban areas. Reference maps available might be used also for calibrating models, not only for validation – Downscaling the reference maps to 1 ha resolution? New layers of geographic data should be tried, e.g.: – night time light – Tele-Atlas

European Forum of Geostatistics Bled October Further developments Adding Switzerland and Norway Producing a first version for 2006 Layer of changes Introducing more detailed data for large urban areas Papers Integrating bottom-up data? – Clarifying confidentiality and copyright issues. – Possibility: no integration, but information of how to get the bottom-up data