Accuracy Characteristics of ALOS World 3D 30m DSM Karsten Jacobsen Leibniz University Hannover, Germany Institute of Photogrammetry and Geoinformation jacobsen@ipi.uni-hannover.de 1
Introduction Japanese space agency JAXA generated with all usable ALOS PRISM images height model ALOS World 3D (AW3D) covering landmass from 82 southern to 82 northern latitude with point spacing 0.15 arcsec (~5m) as commercial version - nominal SZ = 5m ALOS PRISM active from January 2006 May 2011, 2.5m GSD, tri-stereo, max b:h=1:1 Reduced version with 1 arcsec (~30m) is available as AW3D30 free of charge limit of SRTM available AW3D30 gaps due to clouds Possible replacement of SRTM DSM (limited to -56 +60 latitude) 1 x1 Bogota, 28% gaps 2
Large area covering free of charge height models DSM from Shuttle Radar Topography Mission (SRTM) imaging 11 days in February 2000 since September 2014 with exception of Near East available with 1 arcsec point spacing (~30m) SRTM DSM currently standard for several applications ASTER GDEM2 based on all usable ASTER PRISM Images, point spacing 1 arcsec, similar range as AW3D Bright area: SRTM only 3 arcsec point spacing (~90m) 60-56 82-82 ICESat only satellite laser profiler, footprint 70m, SZ=0.2m but satisfying flat part required No height model, but usable for orientation of height models 3
Quality Information of AW3D30 Square mean AW3D30 against SRTM, ASTER GDEM2 and ICESat within 1 x 1 unit 5 used test areas Sainte-Maxime: France at Mediterranean Sea, small part not dense urban area, agriculture, mountains, mostly covered by forest Zonguldak: Turkey at Black Sea, rough mountains, partly forest + dense urban area Karaburun: Turkey at Black Sea, mountains, partly forest, quarries, gravel pits, dumps Jordan: Dead Sea up to Amman, soft mountains, nearly no vegetation Mausanne: France, larger parts of flat agriculture, mountains covered by forest Characteristics of test areas explaining the varying root mean square differences 4
Quality Information of AW3D30 Test area marked in 1 x 1 distribution unit Mask of available data no data: Black Sea + reservoir - only test area Frequency distribution of correlation coefficients Number of images/object point Percentage including sea 5
Images/point Test area Sainte-Maxime, France 15 Reference DSM from digital aerial images (IGN) SZ as number of images/point) SZ = 4.29m 0.14m x images/point number Images/point Frequency distribution of DZ with overlaid normal distribution as F(SZ) and F(NMAD) Kurtosis = 1.03, skewness = 0,50 DZ NMAD = normalized median absolute deviation - in case of normal distribution NMAD = SZ Normal distribution based as F(NMAD) usually fits better to frequency distribution as F(SZ) NMAD is characterizing the frequency distribution of DZ better as SZ 6
Test area Saintes-Maxime, France whole area slope<10% slope>10% DZ Color coded height differences AW3D30 against reference-dsm Influence of vegetation visible reference-dsm imaging ~July 2014 ALOS-PRISM 2006-2011 Google AW3D30 ~ 25% more accurate as SRTM DSM As usual flat areas (<10% slope) more accurate as whole area including areas with slope > 10% 7
Test area Zonguldak, Turkey Google mountainous, partially harshly, partially forest, partially mountainous city area Color coded number of images/point in average: 6.7 images/point images/point SZ as F(images/point) SZ=9.20m 0.90m x images/point 8
Test area Zonguldak, Turkey Google DZ Whole area slope<10% relative DZ AW3D30 against reference DTM But only points on ground determined directly in aerial model Comparison of AW3D30 with other free available DSM; AWD3D30 clearly better as other, SRTM C (3 arcsec) and SRTM X (1 arcsec) similar, ASTER GDEM2 not as good, only relative accuracy < as for SRTM 9
DZ Test area Karaburun, Turkey DZ AW3D30 against LiDAR DSM without quarries, gravel pits and dumps whole area slope < 10% slope > 10% Difference LiDAR 2014 to phogrammetric survey 2005 Comparison of AW3D30 with LiDAR DSM; AWD3D30 clearly better as SRTM DSM AW3D30 only based on 3 stacks 10
number of images/point Test area Jordan SZ as F(images/point) SZ = 4.08m 0.18m x images/point In average 5.3 images/point DZ AW3D30 against reference DSM DZ Images/point Whole area slope<10% relative SZ and NMAD of AW3D30 ~ 60% von SRTM DSM ASTER DGEM2 not as good, only relative accuracy better as for SRTM Jordan results better as for Zonguldak nearly no vegetation, surface not as rough 11
number of images/point Test area Mausanne SZ as F(images/point) SZ = 5.02m 0.14m x images/point average 12.7 images/point DZ Bilder/Punkt Whole area slope<10% slope>10% SRTM and AW3D30 filtered to DTM due to reference DTM DZ AW3D30 against reference DTM 12
conclusion Accuracy varying from test area to test area vegetation different, different urban areas, different terrain slope, different roughness, different number of images/point AW3D30 quality file includes accuracy trend NMAD should be preferred as accuracy figure instead of SZ Accuracy depends upon terrain slope Accuracy depends upon number of images/point in average: SZ = 4.46m 0.15 x images/point ASTER GDEM2 not as good as AW3D30 and SRTM, only relative accuracy of GDEM2 better as for SRTM Up to now GDEM2 used for area -56 up to -82 southern latitude and for 60 up to 82 northern latitude as standard reference DSM 13
conclusion Average of 5 test areas Accuracy relation ~ 1.5 (SRTM) : 1 (AW3D30) SRTM: NM=4.51 + 3.68 x tan(slope) AW3D: NM=2.83 + 3.67 x tan(slope) SZ= 5.63m 0.34m x images/point without Zonguldak: SZ = 4.35m 0.15m x images/point Advantage of AW3D30 against SRTM: more accurate, coverage +/-82 latitude, point spacing in general1 arcsec (~ 30m); nevertheless there are some areas with gaps due to cloud coverage Of course there are commercial DSM / DTM more precise and with smaller point spacing as AW3D30 - e.g. World DEM (TanDEM X) and AW3D 14