Resolution and Resampling Effects of Raster Data in Global and Regional Models

1
Resolution and Resampling Effects of Raster Data
in Global and Regional Models
NRC Board on Earth Sciences and Resources Mapping
Sciences Committee Roundtable
E. Lynn Usery
usery_at_usgs.gov
U.S. Department of Interior U.S. Geological
Survey
http//cegis.usgs.gov
2
Outline

• Objectives
• Landscape and Regional Models
• AGNPS
• Continental and Global Models
• Asia Land Cover
• Sea Level Rise
• Conclusions

3
Objectives

• Provide examples of spatial data effects on
  models
• Provide examples of data processing effects
• Provide sample model results
• Demonstrate model results through animation
• Provide basis for discussion of resolution and
  accuracy effects

4
Agricultural Non-Point Source Pollution (AGNPS)

• Operates on a cell basis and is a distributed
  parameter, event-based model
• Requires 22 input parameters
• Elevation, land cover, and soils data are the
  base for extraction of input parameters
• Generates 51 output parameters

5
Comparison of Watershed AreasLittle River
Watershed, Georgia, USA (Area in hectares)
Resolution (m) NAWQA GIS Weasel

30 33423.8 34885.8
60 33702.5 35089.2
120 34076.2 35493.1
210 34631.7 35986.1
240 34859.5 36241.9
480 36426.2 37739.5
960 39444.5 40458.2
1920 45711.4 46418.9
6
(No Transcript)
7
Results Resolution Effects

• Tested with two independent collections
• Elevation at 3 m and 30 m resolution
• Land cover at 3 m and 30 m resolution
• Comparison of values

8
Elevation
9
(No Transcript)
10
Sampling of Points for Land Cover and Elevation
Comparisons for Little River, GA
11
Statistical Testing

• Selected 500 random points over the watershed
• Compared elevation, slope, and land cover values
  at the 500 points
• Computed R2 and pseudo R2 between resolutions
• Plotted R2 and pseudo R2 against resampled
  resolutions from 30 m data

12
Regression Results

• 3 m to 30 m comparison
• Elevations -- R2 of 0.81
• Land cover McFaddens pseudo R2 of 0.139,
  meaning little correlation
• Derived parameters, e.g., slope, problematic
  because of degraded data source

13
Results Resampling Effects

• Analysis uses DEM, slope, and land cover at 30,
  60, 120, 210, 240, 480, 960, 1920 m cells
• Starting point is 30 m DEM and land cover
• Calculate slope at 30 m cell size from DEM
• Resample land cover
• How to generate slope at 60 m and larger cell
  sizes? How to aggregate land cover?

14
Results -- Slope
Slope 30 to 480m Pixels 7.8816 7.8232
7.5870 7.8251 8.1604 8.5415 8.2065
7.9530 7.7434 7.7092
Slope 210 to 480m Pixels 7.9514 7.8969
7.6244 7.7855 8.1263 8.5087 8.2157
7.8606 7.6390 7.6081
Regression Output Constant 0.2762 Std Err
of Y Est 1.1626 R Squared 0.7690 No. of
Observations 500 Degrees of Freedom 498
X Coefficient(s) 0.8860 Std Err of
Coef. 0.0218
15
Results -- Slope

• Slope
• Method of calculation affects results
• Higher resolution aggregation directly to large
  pixel sizes yields better results than multistage
  aggregation (e.g., 30 m to 960 m is better than
  30 m to 60 m to 120 m to 240 m to 480 m to 960 m)
• Even multiples of pixels hold results while odd
  pixel sizes introduce error

16
Results - Land Cover -- 210 m Pixels
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
Effects on Model OutputsTotal Soluble Nitrogen
22
Effects on Model OutputsTotal Soluble Phosphorus
23
Conclusions

• Resolution affects results
• Elevation retains values well because of
  averaging methods of resampling
• Derivative datasets do not retain values as well
• Land cover (categorical data) is inconsistent
  across resolutions because of nearest neighbor
  resampling
• Model outputs follow input degradation with
  resolution, but indicate a threshold based on
  model formulation with respect to areas of
  aggregation
• Resampling retains values better with even
  multiples of original pixel sizes
• Aggregation directly from higher resolution to
  lower retains values better than multiple
  intermediate resampling

24
Asia Land Cover

• Areas resulting from projection and resampling

25
Resampling Asia Land Cover

• Land cover data (21 categories) at 1 km pixel
  size for Asia
• Resample to 2, 4, 8, 16, 25, and 50 km pixels
• Tabulate land cover percentages at each
  resolution to assess scale effects
• Aggregate in various ways and retabulate to
  assess aggregation effects

26
Asia Land Cover Lambert Azimuthal Equal Area
Projection, 8 km pixels
27
Scale Effect ResultsAsia Land Cover
28
Aggregation Effect ResultsAsia Land Cover
29
Empirical TestingMultiple Datasets

• Global land cover, 30 arc-sec
• Global elevation, 30 arc-sec
• Global population, 30 arc-sec
• Global temperature, ½ degree
• Global precipitation, ½ degree
• Global vegetation, 1 degree
• Down-sample 30 arc-sec data to 1,4,8,16,25,50
  kilometers

30
Empirical TestingMultiple Projections
Primarily Equivalent

• Eckert IV
• Hammer
• Sinusoidal
• Mollweide
• Lambert Equal Area Cylindrical
• Wagner IV
• Robinson (non-equal area)

31
Empirical Test ResultsUse of Commercial GIS
Software

• Commercial GIS software is unreliable for global
  projection variety of problems
• Projections do not complete
• Work at some resamplings, but not others
• Inverse projections result is extension of raster
  areas to 0-degree lines
• Repeat areas at edges of projection
• Computation times extensive (100 to 200 hours on
  high end computers)
• May not use exact projection equations

32
Hammer from Commercial Software
33
Hammer from USGS Mapimg
34
Dynamic Projection of Raster Cells

• Compute areas of pixels in geographic coordinates
• Map each raster line to appropriate area
• Result is accurate for computation and analysis,
  but each raster line has different size
• For 30 arcsec data
• Cell at equator is approximately 1 km x 1 km,
  cell area 1 km2
• Near poles, a cell is 1 km x 0.006 m, cell area
  6 m2

35
Area Calculation for Equiangular Pixels in
Spherical Coordinates
36
Pixel Areas Computed from Spherical Coordinates
(in meters2)
Latitude 30 Arcsec 0.5 Degree 1.0 Degree
0 858,631 3,091,035,692 12,363,671,878
15 858,631 2,982,220,448 11,970,315,668
30 743,628 2,670,171,821 10,761,202,175
45 607,188 2,176,155,408 8,818,730,582
60 429,370 1,533,837,609 6,275,272,108
75 222,291 786,991,318 3,304,173,896
90 (-1 unit) 63 13,487,417 107,896,706
37
(No Transcript)
38
Conclusions

• Areas of land cover vary significantly (up to 30
  ) based on aggregation method
• Nearest neighbor resampling leads to inaccurate
  aggregations based on modal category concepts
• Continuous data (DEM and slope) retain values
  better through aggregation because of averaging
  (bilinear) during resampling.
• Continental land cover datasets shows significant
  effects on land cover areas resulting from
  categorical (nearest neighbor) resampling.

39
GIS Data Modeling of Sea Level Rise
40
Objectives

• Examine effects of sea level rise from global
  perspective
• Determine global population numbers in areas of
  inundation
• Determine population numbers in regional and
  local areas that would be inundated
• Model with simulation and animation of sea level
  rise

41
Sea Level Rise from Ice Melt

• Minimal over time
• Only now at 3.4 0.5 mm per year
• Not detectable in world wide GIS datasets
• Anticipate maximum of 6-7 m
• Slight rises can make high surges from storms
  more severe

42
Hurricane Katrina Storm Surge Effects
43
Hurricane Katrina Storm Surge
44
New Orleans MODIS Images Before and After Katrina
45
(No Transcript)
46
Hurricane Rita Storm Surge Effects
47
Peveto Beach Before and after Rita -- Note
destruction of houses from storm surge
48
The Indian Ocean Tsunami 2004

• Maximum run-up exceeding 30 m in Banda Aceh and
  10 m in several locations in Sri Lanka

49
Indian Ocean Tsunami Effects Banda Aceh, 2005
50
Pankarang Cape Jan 13, 2003
51
Pankarang Cape Dec 29, 2004
52
Modeling High Sea Level Rise or Surge and
Affected Population Numbers on Worldwide Basis
53
Data Sources -- Global

• Global land cover, 30 arc-sec, 1 Gb file, USGS
  source
• Global elevation, Gtopo-30, 30-arc-sec, 2 Gb
  file, USGS source
• Global population, Landscan 2005, 30 arc-sec, 1
  Gb file, Oak Ridge National Laboratory source

54
Data Sources Regional/Local

• SRTM Elevation Data, 90 m
• Global Land Cover, resampled to 90 m
• Global Population, resampled to 90 m
• For the U.S.
• USGS National Elevation Dataset, 30 m
• USGS National Land Cover Dataset, 30 m
• Census block population resampled to 30 m with
  dasymetric technique using residential land cover
  from NLCD

55
Projection of Data Sources

• All data originally in geographic coordinates
  (latitude and longitude)
• Projected to Mollweide for modeling and animation

56
Projection Solutions

• USGS developed mapimg projection software
• Handles global raster data, large file sizes
• Programmed new resampler for categorical data
  (land cover)
• Modal category, statistical selection, or user
  choice for output pixel value
• Programmed new resampler for population counts
• Additive resampler

57
Resampling

• Primary step in raster generalization, sometimes
  called down-sampling
• Geospatial data can suffer from great geometric
  distortions when being reprojected or transformed
• Errors associated with these distortions and
  scale changes affect resampling
• For categorical data, such as land cover, pixel
  gain and loss result
• For population counts, an additive resampler is
  required

58
Resampling Nearest Neighbor Method without
Generalization

• One point in the output image space maps to a
  corresponding point in the input image space (via
  the inverse mapping algorithm)

59
Resampling Nearest Neighbor Method with
Generalization (Down-sampling)

• If the resolution of the output image is reduced
  (generalized or down-sampled), adjacent pixels in
  the output may fall more than one pixel away in
  the input (via the inverse mapping algorithm)

60
Categorical Resampling

• New resampling algorithm treats pixels as areas,
  not points, Steinwand (2003)
• Four corners of each pixel are mapped into the
  input space
• Many pixels involved
• Apply simple statistical methods or user
  specification to determine output image pixels
  based on the area the pixel covers in the input
  image

61
Extreme Downsampling (64 pixels to 1) and
Reprojection with the Nearest Neighbor
Reduces data volume from 1 Gb to 16 Mb
62
Extreme Downsampling (64 pixels to 1) and
Reprojection with the New Algorithm
63
Generalizing Population Counts

• Data are numbers of people per pixel
• Generalize four pixels to one
• Must add the four pixels together
• Not available in commercial software
• Developed in mapimg

64
The Application to Global and Regional Data for
Sea Level Rise and Surge Modeling
65
Global Case, Reproject All Data
66
Regional Case, Subset Data to Desired Area
67
Reproject Data Subset
68
(No Transcript)
69
(No Transcript)
70
Model Results

• Model yields multiple maps
• Each map is sea level rise at specified level,
  e.g., 1 m, 2 m, 3 m, 5 m, 10 m
• Color mask shows inundation area
• Population counter tabulates number of people in
  inundated area

71
1 M Rise
72
2 M Rise
73
3 M Rise
74
5 M Rise
75
10 M Rise
76
20 M Rise
77
30 M Rise
78
Results -- World
Water Level Increase (m) Population Affected (Net) Area of Land Loss (km2)
5 669,739,138 5,431,902
10 870,751,960 (201,012,822) 6,308,676
20 1,176,709,476 (305,957,516) 7,888,233
30 1,405,824,876 (229,115,400) 9,459,562 (This is size of all of USA)
79
Results World Population Effects
80
Animation Process
Elevation
Land Cover Snapshots by Changing Elevation
Land Cover
Animation Software (Macromedia Flash)
AVI files
81
(No Transcript)
82
(No Transcript)
83
(No Transcript)
84
(No Transcript)
85
(No Transcript)
86
(No Transcript)
87
(No Transcript)
88
Conclusions

• Global sea level rise and storm surges
  potentially affect large numbers of people and
  various types of land cover
• Sea level rise can be modeled on a global scale
  with data at 30 arc-sec resolution
• Modeling requires specially designed projection
  and resampling software
• Better categorical resampler for global land
  cover
• Additive resampler for population counts
• Animation effectively portrays results and
  effects of potential sea level rise and storm
  surges

89
Caveats Resolution and Accuracy

• Resolution
• Global data 30 arc-sec resolution affects
  results small areas, e.g., river valleys smaller
  than twice resolution (60 arc-sec, 2 km at
  Equator) cannot be modeled effectively
• Regional/local data areas smaller than twice
  resolution (180 m for SRTM, 60 m for NED) cannot
  be modeled effectively
• Accuracy
• Errors in data reflect as errors in model (e.g.,
  in the gTopo30 data, the Fly River, New Guinea,
  is 30 km wide, but doesnt show in model because
  of data errors)

90
Resolution and Resampling Effects of Raster Data
in Global and Regional Models
NRC Board on Earth Sciences and Resources Mapping
Sciences Committee Roundtable
E. Lynn Usery
usery_at_usgs.gov
U.S. Department of Interior U.S. Geological
Survey
http//cegis.usgs.gov