Generating a 2m resolution DEM of South Africa · aerial imagery of South Africa on a regular (3-4...
Transcript of Generating a 2m resolution DEM of South Africa · aerial imagery of South Africa on a regular (3-4...
Generating a 2m resolution DEM of South Africa
Adriaan van Niekerk
Centre for Geographical Analysis
Stellenbosch University
1
Outline
• Terminology (DTMs, DSMs, nDSMs, DEMs)
• Rationale
• Overview of research
• Examples
• Conclusions & challenges
• Current work
2
SOME IMPORTANT TERMINOLOGY
3
DTMs, DSMs and nDSMs
• Digital terrain model (DTM)
– Height of earth's surface excluding land cover
• Digital surface model (DSM)
– Height of earth's surface including land cover
• Normalized DSM (nDSM)
– Height of land cover
– nDSM = DSM - DTM
Digital Elevation Models (DEMs)
• Regular array of elevation points (cells)
• Can represent:
–DSM
–DTM
–nDSM10m
30m
0m
RATIONALE
6
Shuttle Radar Topography Mission (SRTM)
• 30m version of the SRTM DEM of Africa was released in 2014
• Big jump in resolution from the 90m version
• Quality not necessarily better
– Voids
– Noise
• But, best freely available dataset available
• MANY applications
7
8
9
10
SRTM DEM not suitable for:
• Hydraulic (flood) modelling
• Pre-processing of high resolution (< 10m) satellite imagery
• Detailed (<30m) land cover mapping
• Construction & telecommunication planning
• Air traffic routing and navigation
• Precision agriculture & forestry
• Environmental management and impact assessments
11
What we need is LiDAR data
• Can be used to extract DSMs, DTMs and nDSMs
• Highly accurate (10-100cm vertical)
• Provides additional (returns) information that can be used to infer land cover and vegetation structure
• Preferably this data should be updated every 1-3 years and it should cover the entire SA
• Viable? Maybe, but not in the near future
12
Alternatives?
• WorldDEM (DSM)– 12m resolution, 4m absolute vertical accuracy– Price 10 Euro per km2 (~R185million for SA)
• Elevation 10 (DSM)– 10m resolution, 5m absolute vertical accuracy– Price 17 Euro per km2 (~R310million for SA)
• Elevation 1/2 (DSM)– 1/2m resolution, up to 1.5m absolute vertical accuracy– Price ? (a lot!)
• Costs not viable for research, especially over large areas (societal benefits)– Started with developing our own products @ SU
13
14
2000-2008: Interpolation methods (contours and spot heights) [SUDEM Level 1 DTM]
2008-2012: Fusion methods (contours, spot heights and SRTM DEM) [SUDEM Level 2 DTM]
2012-: Automated extraction of elevation data using digital photogrammetry [SUDEM Level 3 DSM]
2014-: DSM to DTM conversion [SUDEM Level 4 DTM]
Overview of DEM research @ SU
15
2000-2008: Interpolation methods (contours and spot heights) [SUDEM Level 1 DTM]
2008-2012: Fusion methods (contours, spot heights and SRTM DEM) [SUDEM Level 2 DTM]
2012-: Automated extraction of elevation data using digital photogrammetry [SUDEM Level 3 DSM]
2014-: DSM to DTM conversion [SUDEM Level 4 DTM]
DEM research @ Stellenbosch University
Photogrammetry
• Use parallax in stereo images to
estimate distance (height)
• Has been around since 1867!
• Was used to (manually) produce contours and (most) spot heights on topographical maps
• Digital photogrammetry automates the extraction process on a per-pixel basis resulting in very high density of elevation points ("point cloud")
16
CD:NGI – A fantastic source of data!
• 0.5m resolution colour and near-infrared stereo aerial imagery of South Africa on a regular (3-4 year) basis
• Spatial accuracy excellent (<1m error)• Freely available!• Grossly underutilised (mainly used as backdrop in
GIS)• CD:NGI also has very accurate triangulation data,
which means that time-consuming image orientation procedures can be fully automated
17
DSM extraction process
1. Calculate surface – camera distance
(For each pair of pixels in overlapping aerial photographs)
2. Use camera's position to calculate height above geoid (mean sea level)
3. Store heights in a point cloud
4. Analyse point cloud
5. Generate a DEM
18
Sources: https://www.e-education.psu.edu/geog480/node/444
19
0.5m
2m 16 height measurements per 2m area
20
Red
Green
Blue
Pan
Filter
Elevations are extracted from 5 bands
21
Flight path
Individual photo
Overlap = 8
Overlap = 2
Overlap = 4
Normally 4 or more overlapping images
Result
• HUGE point cloud
– 20 – 80 points per m2
– 80 – 320 points per 2x2m pixel
• Allows for statistical analyses
– Remove outliers (anomalies)
– Can use correlation score to assess points
– Estimate VERY accurate elevations
22
http://www.opentopography.org/index.php/news/detail/srtm_version_30_global
SUDEM Level 3 (DSM) workflow
23
Preparation & Setup
Epipolar pairs and DSM
extractions
(PCI)
DSM geocoding
(PCI)
DSM merging
(PCI)
Automated gap filling
(custom)
Automated error corrections
(custom)
DSM extracted using fully automated methods
Manual error corrections
Final product
24
Contours & spot heights SUDEM Level 1
SRTM DEM, LiDAR
SUDEM Level 2
Interpolation
Fusion
0.5m Aerial Images Photogrammetry
SUDEM Level 3
DSM to DTM conversion
SUDEM Level 4
SUDEM DEVELOPMENT
OVERVIEW
Update strategy
• 5m SUDEM (L2) product forms basis (wall-to-wall coverage)
• 2m DSM (L3) is continuously generated (prioritizing based on demand)
• DSM to DTM (L4) conversion
• DTM (L4) fused into 5m product
25
SOME EXAMPLES OF LEVEL 3 PRODUCT
26
2730m SRTM DEM of Ceres region
282m SUDEM L3 of Ceres region
292m SUDEM L3 of Ceres region
3030m SRTM DEM of Mitchell's pass
312m SUDEM L3 of Mitchell's pass
3230m SRTM DEM 2m SUDEM
3330m SRTM DEM 2m SUDEM
34
2m SUDEM L3
0.5m aerial photograph
35
S U D E M
S U D E M
Vertical absolute accuracy?
36
PRODUCT MAE (m) STD DEV (m) 90 percentile (m)
30m SRTM DEM 3.22 2.88 6.42
5m SUDEM Level 1 1.57 1.76 3.25
5m SUDEM Level 2 1.77 1.21 3.06
2m SUDEM Level 3/4 0.35 0.25 0.66
Based on surveyed reference points supplied by City of Cape Town
CONCLUSIONS
37
SUDEM Level 3 product is:
• Much more detailed than any other existing DEM
– 2m resolution (1m also available)
– Very accurate: 0.66m (90th percentile)
• Comparable with LiDAR
– Slightly less detailed/accurate
– No penetrative power (DSM not DTM)
38
Challenges
• Automated DSM to DTM conversion
– Many applications require DTMs
• Failed height extractions (blunders) in areas of shadow (e.g. trees) and water bodies (e.g. dams)
• HUGE datasets!
39
Challenge: Dealing with blunders
• Minimize by using elevations from multiple bands
• Identification (using correlation score)
• Small blunders: interpolate from surrounding elevations
• Large blunders: use other data to fill the gaps – SUDEM Level 2
• Manual edits
40
41
Aerial photograph
42
Geocoded DSM extracted from the green band
43
Geocoded DSM extracted from the blue band
44
Geocoded DSM extracted from the blue band
45
Geocoded DSM extracted from PC1
46
Unreliable data(areas with lowcorrelation scoresin all bands)
47
Small unreliable areas (3.5%) are removed by interpolating from surrounding reliable elevations. Only gross errors remain.
48
Gross errors are corrected by making use of L2 product (2.3%) and minor manual editing (<0.7%).
Challenge: Processing requirements
• Each 1:50 000 tile requires 8 days of CPU time (8 cores) and 1TB of disk space
• If run on one machine it will take 43 years to complete the entire SA!
• Completed only 3% of SA so far!
• Currently have 6 dedicated machines
• Limited resources to acquire additional hardware
49
Work in progress…
• Improve workflow
– Reduce errors (automated corrections) & streamline/reduce manual editing
– Optimization
• DSM to DTM conversion
• Increase processing capacity
– Parallelization
– High performance computing (HPC) cluster (funding!) –hope to have 40 machines by end of November
50
Applications
• Current research projects– Digital soil mapping
– Salinity modelling
– Hydrological modelling
• Potential applications– Wetland mapping?
– Visual impact assessments?
– Setback lines?
– ?
51