High-accuracy Terrain Modelling for SoilMapping using ALOS-PRISM Imagery
S.J McNeill S.E Belliss D Pairman
Landcare Research New Zealand
Motivation
S-map is a national soil database, map and informationinference system for New Zealand providing:
A complete national digital soil mapAccessible data and inferred key informationProvides the best legacy data as well as new data
S-map methodology:
Gather legacy knowledge in modern data frameworkGather new data for areas poorly mapped at presentUse data mining methods to predict high country soilpropertiesPresent data in an accessible form for end-users
A good DEM is required for the high country
Low cost per unit areaBetter accuracy than existing DEMs generated from20-metre contoursSatisfactory for estimation of low-order derivatives
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Objectives
Investigate how better DEMs can be produced forhigh-country soil mapping
Emphasis is on use of existing ERDAS product suite, wherepossible
Consider DEM quality for generation of complex terrainattributes
Important considerations:
ALOS-PRISM imagery favoured due to low cost per unit areaHigh cloud cover means some areas cannot be covered withoptical imageryMethodology should use a variety of data sources whereadvantagous
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
ALOS-PRISM sensor
Forward, nadir, aft telescopes with constant 2.5 m resolution
Each telescope has 4×CCD line sensors camera, each with aseparate focal centreRaw data rate from PRISM sensor subsystem 960 Mbit/s
Far exceeds available downlink data rate of 120 or 240Mbit/sLossy JPEG compression implemented on-board, withconstant output data rate
JPEG block artifacts reduced by processing change in Oct'07
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Study data
ALOS-PRISM imagery
01 Jan 2008 (3 scenes), 20 Jan 2009 (3 scenes)Excellent sun elevation (59o & 56o), no cloud
ALOS-PALSAR imagery
Intended to �ll gap between PRISM DEMs using InSARDual-pol. data used as basis for vegetation height model
Field data
Di�erential GPS feature position estimates (σ = 0.25m)
Other supporting data
Raster DEM at 25m postings from 20m contours & spotheightsGeodetic marks for height validation
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Study area
166 168 170 172 174 176 178
−48
−46
−44
−42
−40
−38
−36
−34
PALSARPRISM
174.0 175.0 176.0 177.0
−42.0
−41.5
−41.0
−40.5
−40.0
−39.5
−39.0
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Scene orthophoto
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Subscene orthophoto
3141× 2448 subscene
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Imagine LPS DEM generation
Extensive facilities for data management and editing
Adaptive, multi-resolution, correlation stereo method
Generic line-array model for ALOS-PRISM sensorRange of output options limited:
ASCII �le of estimated DEM pointsTIN, 3D shape model, or interpolated TINOnly qualitative error measure available
(19× 19 km sun-shaded subscene) (8× 8 km sun-shaded subscene)
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Imagine LPS DEM point output
1820000 1830000 1840000 1850000 1860000
5560000
5565000
5570000
5575000
5580000
5585000
5590000
5595000
●
●
●
ExcellentGoodFair
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Imagine LPS DEM accuracy
Measured against geodetic marks (order 1�5)Quadratic trend in error minimum in scene centre
Due to simple single-CCD line sensor model (?)
Can be corrected using post-�tting of DEM data
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Accuracy of corrected DEM
Measured against independent contour-derived DEMEstimated σ = 6.48m for TIN-interpolated DEM95% equal-tail con�dence interval [4.46, 11.8]m
Measured against independent geodetic marks (order 1�5)Estimated σ = 2.90m for TIN-interpolated DEM
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Interpolation of point output
Abandon use of TIN-interpolated DEM output
Use interpolation of point output, with requirements:
Enforce smoothnessIncorporate other data, where available
Rational Basis Function (RBF) interpolation
s(x) = p(x) +N
∑i
λiΦ (x − xi )
The RBF is a weighted sum of a radially symmetric Φ at thecentres xi and a low degree polynomial p
Finding λi given xi , s(xi ) very di�cult for large N
Specialised software provides �tting, �ltering and surfacegeneration using approximation methods
Method used can �t successively �ner-resolution terrainsurface models
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Processing method
Reduce high-frequency JPEG artifacts and de-stripe
Import PRISM data for ERDAS LPS
Use GCPs to �t stereo model using ERDAS LPS
Generate point output using ERDAS LPS DEM generation
Apply cross-track model correction for DEM height
Build uncertainty model for corrected point output
Fit RBF surface from point output with speci�ed pointuncertainty
Generate terrain surface by evaluating RBF
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
TIN/RBF interpolation comparison
TIN-interpolated DEM, sun-shaded RBF-interpolated DEM, sun-shaded
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
TIN/RBF detailed comparison
TIN-interpolated DEM, sun-shaded RBF-interpolated DEM, sun-shaded
(730× 753) (730× 753)
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
RBF advantages & disadvantages
Advantages:
Terrain surface inherently smoothFitted RBF can be �t successively with additional terraininformationAccuracy of RBF-interpolated DEM not degraded comparedto TIN-interpolated DEM
Disadvantages:
Fitting process needs to be managed carefully to preservememorySome tuning of RBF �tting parameters required
IGARSS-2011, 25-29 July 2011, Vancouver, Canada
Conclusions
We have developed a pragmatic rather than the technically�best� solution for DEM generation
Need to provide pre- & post-processing to make exitingsoftware work satisfactorily
Additional processing requires little extra manual e�ort
For high country, results provide a useful improvement overexisting DEMs
Acknowledgements
Research funded by the Ministry for Scienceand Innovation (contract C09X0704).
IGARSS-2011, 25-29 July 2011, Vancouver, Canada