Foreword
Dr. Raad A. Saleh, Guest Editor
Economics. That is a key thrust in our quest for automation of information extraction from digital imagery. Influenced by the specific field of application, economics morph from labor cost to time saving, to rapid response in support of combat theater, natural disasters, etc. Automation is the design, development, and implementation of operations with minimal human intervention. I start with a premise that the human eye-brain perception system is THE ultimate in information extraction, with respect to intelligence and resolving ill-defined cases. In this context, it follows that automation is an attempt to replace the human eye-brain system by machine. Thanks to atmospheric noise and other imperfections in the acquisition phase, aerial and satellite imagery fall into the category of “too complex” to automate. This is unlike industrial vision systems installed in controlled environments, thus lending themselves to near-perfect automation of information extraction. Cognizant of this difficulty, the goal has so far been to model the human eye-brain system, while capitalizing on only relevant aspects that maximize cost saving, i.e., economics.
Depending on many influencing factors, there have been distinct approaches in balancing the roles of the two contenders, machine versus human, or in less dramatic descriptors, automation versus manual. The influencing factors include the kind of imagery used, the intended application, the resources available, and the specifications imposed on deliverables. The common basic assumption, however, is that more automation means less manual; hence less labor cost, thus translating to better economics. To test this assumption, we draw from the chronological development of the costbenefit model for automated surface extraction. The model recognizes that manual editing is always required after automated extraction to correct erroneous or incomplete information. The model dictates that manual editing must not approach a threshold of high labor cost which may eventually defeat the whole purpose of automation. Simply put, economics fail if the human operator spends more time correcting erroneous data (that were automatically extracted), than the time that would have been spent to manually extract the same data.
What is clear, however, at least within the bounds of this Special Issue, is that we no longer confine ourselves to the restrictive classification of automated versus semi-automated. Instead, the competing roles of machine versus human can be classified into three scenarios. These are:
An example of the first is BridgeView (NCRST, 2002), a toolset developed to assist the operator extract new, and revise existing, roads and bridges using aerial and satellite imagery.
An example of the second is Feature Analyst (Opitz and Finlayson, 2004) developed by Visual Learning Systems, Inc. and described in www.featureanalyst.com. Feature Analyst is a suite of machine learning algorithms that extract the objectspecific features defined by the operator.
An example of the third approach is a software package developed for the Schafer Corporation (Gorkavyi et al., 2004) and was used with lidar data. The quality of automated extraction of roads and other linear features with this package was estimated as 80-90%. The general conclusion is that fully automated extraction of linear features is achievable, provided that false extractions are accounted for.
Quackenbush presents “A Review of Techniques for Extracting Linear Features from Imagery.” “ Many of the papers reviewed reported qualitative results that were based on a visual assessment. Those that provided more quantitative results were often very vague in defining how the evaluation was performed.” It is no surprise then to find research conducted primarily, if not exclusively, to evaluate the outcome of automation and assess its efficacy. This is evident in the papers by Harvey et al., Kaiser et al., Péteri et al., and by Hinz and Wiedemann.
In their paper titled “User-Centric Evaluation of Semi-Automated Road Network Extraction,” Harvey et al. describe a usability evaluation of a road extraction system, in both manual and semi-automated modes, and against a commercial photogrammetric workstation. Their main conclusion was that the performance of all three approaches was comparable. Kaiser et al. evaluated the utility of various extraction methods for mapping trail networks. They found that the most effective and efficient approach to trail mapping was a hybrid of an automated linear feature extraction routine, followed by manual interpretation, delineation, and editing.
The paper by Péteri et al. presents a description of a quantitative evaluation of roads extracted from high-resolution satellite images by means of automatic and semi-automatic methods. Hinz and Wiedemann suggest that the efficiency of road extraction can be increased by incorporating real time internal self-evaluation. They concluded specifically that the proposed self-diagnosis could be a helpful tool to guide the human operator through the road network by pointing to such parts where problems of uncertainties occurred during the automated extraction.
Other papers, by Doucette et al., Kim et al., and Chen et al., are geared toward high-resolution satellite imagery. Doucette et al. presented a methodology to automate road extraction while exploiting the spectral dimension in multispectral imagery. In their paper, “Tracking Road Centerlines from High Resolution Remote Sensing Images by Least Squares Correlation Matching,” Kim et al. reported that template matching could offer wide applicability of automated extraction of curvilinear features. Chen et al. reported that road edges can be extracted using two-dimensional wavelet transform across different scales. Post-processing is then used to delineate the road centerlines.
Most of the remaining papers proposed new or modified techniques to automate the extraction process. In their paper titled “Road Extraction Using SVM and Image Segmentation,” Song and Civco developed extraction techniques utilizing pixel spectral information for classification and image segmentation to derive the object features, including roads. Priestnall et al. present extraction framework based on object- oriented geodata model to allow contextual knowledge to be used to automatically differentiate between classes of linear features. Hu et al. describe a piecewise parabolic representation of the centerline by incorporating least square matching with deformable templates. An automated derivation of agreement index between extracted and reference road features using a line-length buffer is described by Hodgson et al. as a parameterization model. In their paper, “ Extraction of City Roads Through Shadow Path Reconstruction Using Laser Data,” Zhu et al. suggest combining image and laser data to extract roads using homoeomorphous mapping. They promote the potential of extracting linear features from the laser data and using them to guide the extraction from the aerial imagery.
This Special Issue was conceptualized to focus on Linear Feature Extraction from Remote Sensing Data for Road Centerline Delineation and Revision. The word “centerline” has since been found to be overly optimistic, given the state of science in this topic. It is therefore replaced by “Network” to reflect a more realistic state of affairs. The number of submissions was overwhelming, resulting in accepting only a portion of the peer reviewed manuscripts. The lessons learned are numerous. While operational packages are available in the commercial market as well as in academia, it is safe to say that viable full automation is still years away. In addition, we have witnessed a shift into evaluation aspects of the extraction process. We still need however to have better definitions, quantitative as well qualitative, on how to evaluate automation in comparison to manual extraction. Another issue is the substantial work that has been done on post-processing. This must be handled rather carefully because aesthetic presentation of the extracted features does not necessarily mean they are correct.
Acknowledgement
I thank all the peer reviewers who helped in evaluating the
many manuscripts. I also thank Dr. Stan Morain, Dr. James
Merchant, Kim Tilley, and Mike Renslow for their patience
and help in finally bringing this Special Issue to light.
References
NCRST, 2002, BridgeView, developed at the University of Wisconsin-Madison, member of the National Consortium on Remote Sensing in Transportation- Infrastructure, http://www.ncgia.ucsb.edu/ncrst/resources/easyread/cookbooks/BridgeView/BridgeView.pdf.
Opitz, David and Chris Finlayson, 2004, Extracting Road Centerlines: A Grand Challenge Problem. Personal Communication.
Gorkavyi , Ilya, Nick Gorkavyim, Hao Shen and, Tanya Taidakova, 2004, Automatic Extraction of Linear Features from 3-D LIDAR Data, Personal Communication.
Guest Editor
Dr. Raad A. Saleh
Associate Scientist, University of Wisconsin-Madison (At time
of editing this Special Issue)
Present affiliation:
Director, Global Sensing Group, LLC
Madison, WI, 53711-1021, USA
Email: rsaleh@charter.net
| Top | Home |