1139 
BLUNDER ELIMINATION TECHNIQUES IN ADAPTIVE AUTOMATIC TERRAIN 
EXTRACTION 
Fengliang Xu*, Neil Woodhouse, Zhizhong Xu, David Marr, Xinghe Yang, Younian Wang 
ERDAS, 5051 Peachtree Comers Cir, Suite 100, Norcross, GA 30092 - (fengliang.xu@erdas.com) 
COMMISSION III, ThS-3 
KEY WORDS: Automatic Terrain Extraction, Digital Surface Model, Digital Terrain Model, Image Registration, Blunder Removal, 
Object Filtering, ERDAS, LPS 
ABSTRACT: 
This paper introduced a new automatic terrain extraction (ATE) module inside ERDAS’s photogrammetric software LPS. This 
method uses a global DEM to initialize a surface model and iteratively refines it with image registration on different pyramid levels. 
Search range used in image registration is adaptively controlled by elevation range of matched feature points. However, mismatches 
and elevation blunders may cause search range to be out-of-scope and fail ATE process. We used three blunder elimination 
techniques to ensure the convergence of search range: positional cross-correlation, PCA-based blunder elimination, and object 
filtering. The method is tested with images from various sensors including frame cameras, satellites, and Leica’s ADS40. 
1. INTRODUCTION 
1.1 Overview 
Automatic terrain extraction (ATE) is used to extract digital 
surface model (DSM) from triangulated stereo images (Zhang, 
2006; Zebedin, 2006). DSM can be either vector or raster 
format. Its density ranges from 1/3 to 1/10 of density of original 
image pixels, e.g., 10-meter resolution for satellite sensors such 
as Quickbird, 1-meter resolution for ADS40 imagery, etc. DSM 
can be used to generate digital terrain model (DTM), contour 
map, 3D building model, orthophoto, and true-ortho. 
As lidar technology becomes popular and affordable, ATE 
begins to lose market on airborne platform. Lidar is superior in 
providing DSM faster, denser, more accurate (Hodgson, 2004; 
Ma, 2005): centimetre-level vertical accuracy from lidar vs. 
meter-level accuracy from ATE, 1~7 points per meter 2 from 
lidar vs. 1 point per meter 2 from ATE, 2-hours of filtering for a 
1000-mile 2 by lidar vs. much longer time in matching, filtering, 
and manual post-editing by ATE. However, lidar alone cannot 
provide orthophotos. To generate quality orthophotos, images 
from other platform will need to go through automatic point 
measurement and triangulation before being integrated with 
lidar point cloud, and this adds to the cost and complexity of 
map generation, so ATE is still active in low-resolution and 
low-cost map generation. Furthermore, lidar can not generate 
dense and accurate enough points on satellite platform, which 
still relies on ATE for map generation. 
The future of ATE is still ambiguous right now: it may fade out 
in next 5~10 years, or it will further develop in new directions, 
such as: 1) points from ATE (high horizontal accuracy) and 
lidar (high vertical accuracy) may be triangulated together to 
achieve centimetre level accuracy in both horizontal and 
vertical direction; 2) feature-extraction-based ATE and lidar 
point cloud may be integrated to provide buildings structures 
and true-orthos at real-time; 
3) ATE may become more popular in map resynchronization 
and vector-to-raster registration for updating road map or 
detecting changes, etc. 
In order to compete against lidar in both point density and 
accuracy, ERDAS’s next generation ATE moves toward pixel- 
by-pixel and feature-based matching. Adaptive ATE is an 
intermediate step towards this goal. 
1.2 Adaptive ATE 
In ERDAS’s traditional ATE, customer needs to set up search 
range for image registration manually by identifying terrain 
types such as high-mountain or rolling hills and then selecting 
corresponding strategies. To use a wider search range on 
mountains and a smaller one on hills within the same image pair, 
customer needs to manually digitize area-of-interest (AOI) for 
mountain regions and hill regions separately and assign 
different strategies to them. ATE will go from high-pyramid 
levels to low pyramid levels, at each level the search range used 
stays the same on image space but is actually reduced to half on 
object space, which means search range goes smaller and 
smaller by brute-force. This method is proven to be a reliable 
solution, but it requires much human operation and is slow on 
production line. 
The performance of traditional ATE relies on customer-defined 
search ranges. To free customer from such kind of overhead, we 
developed an adaptive ATE module to define search range by 
terrain variation. The basic idea is: at the beginning of a 
matching circle, highest and lowest points (from terrain range) 
along an image ray from first image are projected to second 
image as starting and ending points, which defines a search 
range; and then, features are matched along epipolar-line within 
this range; finally, elevation range of matched points are used to 
update terrain range, ending this matching circle. Matching 
starts from a high image pyramid and terrain range is initialized 
with a global DEM generated with 3-second SRTM DEM 
(Slater, 2006); at each pyramid level, terrain range updated 
from matches on higher pyramid is used to limit search range at