Abstracts of Presentations

Symposium Home | Presentations | References | Schedule | Photos

Keynote Address by Thomas Blaschke, University of Salzburg

Presenters:
Solomon Dobrowski, Tahoe Environmental Research Center

"From Pixels to Patches: Seeing the forest through the trees"
View:Presentation as PDF.

Solomon Dobrowski, Tahoe Environmental Research Center, University of California, Davis
Jonathan Greenberg, Center for Spatial Technologies and Remote Sensing, UC, Davis

Resource management practices are becoming progressively more complex, and require more sophisticated types of spatial information that are of higher thematic and spatial resolution. The availability of high resolution satellite imagery (pixel resolution ≤ 1m2), in conjunction with novel pattern recognition algorithms, is changing the dynamic between the remote sensing sciences and the resource sciences. I discuss the application of these geo-spatial techniques for modeling and mapping forest composition and structure over complex terrain in the Lake Tahoe Basin and American River watershed, CA. Our results suggest that image segmentation is effective at identifying patches of vegetation that are physiognomically or structurally unique. Similarly, spectral variability associated with species composition differences are likely to play a secondary role to variable physiognomy and image texture in driving region-growing segmentation processes. This suggests that OBIA approaches to vegetation mapping are effective at identifying patches of vegetation that represent higher level physiognomic units of the federal vegetation classification hierarchy as opposed to lower level floristic units. Additionally, an image object (polygon) may represent variable floristic composition of similar physiognomy. If we assume that image objects representing vegetation patches with variable species composition are likely to have higher spectral variability than image objects representing patches with homogeneous floristic composition, then the difficulty of attributing these image objects will be greater in the former group than the latter, particularly where spectral data alone are used in classification. I enumerate on the unique capabilities and challenges associated with mapping vegetation to existing federal standards using OBIA techniques.

Kass Green, Alta Vista, Inc

"Object Oriented Classification of Digital Airborne Data for Benthic Habitat Mapping"

Kass Green, The Alta Vista Company
Chad Lopez, Earthdata LLC
Bill Stevenson, NOAA Coastal Services Center
John Wood-Harte Research Institute at Texas A&M University-Corpus Christi
Jim Simons-Coastal Fisheries Division, Texas Parks and Wildlife

NOAA’s Coastal Services Center is working cooperatively with the Texas Parks and Wildlife Department (TPWD) and the Texas A&M University Center for Coastal Studies to develop benthic habitat data, primarily Submerged Aquatic Vegetation (SAV) for 2500 square miles of the east Texas Coastline. Inputs included ADS40 imagery, ancillary data, and field sample data. To develop consistent methods, three object oriented methods were compared in a pilot project: Feature Analyst wall to wall, Feature Analyst class by class, and Definiens image segmentation with CART analysis for polygon labeling. While the accuracies of the three methods were almost identical, the cost and polygon delineations varied considerably. This paper reviews the results of the pilot project and presents the final methodology and accuracy results for the entire project.

Jonathan Greenberg, University of California, Davis, Center for Spatial Technologies and Remote Sensing

"Individual Tree Crown Recognition and Remote Sensing Aware Allometry: Biomass Mapping Using Hyperspatial Remote Sensing"
View: Abstract, Presentation as PowerPoint, Presentation as PDF.

Jonathan A. Greenberg, Center for Spatial Technologies and Remote Sensing, U.C. Davis
Solomon Z. Dobrowski, Tahoe Environmental Research Center, U.C. Davis
Peter Tittmann, Geography Graduate Group, U.C. Davis
Vern Vanderbilt, Ames Research Center, NASA
Susan L. Ustin, Center for Spatial Technologies and Remote Sensing, U.C. Davis

Mapping biomass levels in forests is an ongoing problem in remote sensing due to the widely published “saturation problem,” that increasing amount of biomass result in exponentially decreasing optical responses that saturates at relatively low biomass levels. We present an approach to detecting biomass levels using a combination of individual tree crown delineation of spaceborne hyperspatial sensors and “remote sensing aware allometry.” We demonstrate a technique for estimating crown areas and leaf types using computer vision techniques as applied to IKONOS imagery of the eastern Lake Tahoe Basin. A multiscale template matching approach was applied using representative trees, identified in the field and extracted from the imagery. This analysis produces a vector coverage of every tree in our area of interest and includes estimates of crown area for each tree. We then adapted existing allometric equations to predict biomass as a function of crown area. The combination of these techniques circumvents the saturation problem and allows for accurate, unconstrained estimates of biomass in single-strata forests.

Qinghua Guo, University of California, Merced, Department of Engineering

"Mapping from land cover to land use: an object-based classification approach"

Qinghua Guo, School of Engineering, University of California at Merced, Merced, CA 95344
Maggi Kelly, Department of Environmental Science, Policy, and Management, University of California, Berkeley, CA, 94710
Yu Liu, School of Engineering, University of California at Merced, Merced, CA 95344

Land use (LU) and land cover (LC) mapping are important for understanding landscape dynamics and their relationship with global environmental change. Remote sensing techniques are important tools for LULC mapping: for example, conventional pixel-based classifiers are effective in mapping land cover, which is directly related to physical properties of earth objects. However, the challenge for conventional pixel-based classifiers is in mapping land use: LU classes not only represent physical properties of earth objects, but also represent economic, cultural, and social aspects of earth objects. Moreover, mapping from LC to LU is not often straightforward. For example, one land cover class can correspond to different land use classes, and the same land use class can consist of several land cover classes. There has been considerable literature devoted to the interaction between LU and LC for course-spatial resolution imagery; the recent availability of high-spatial resolution imagery requires a re-examination of LULC mapping. In this study we examine LULC mapping with high-spatial resolution imagery, and evaluate the utility of object-based, rather than pixel-based classifiers to aid the transition from mapped land cover to land use. We first segmented the high-spatial resolution images into objects, which are defined as a group of homogeneous neighboring pixels; then features such as area, shape were extracted from the objects; in addition, we developed a framework of constructing spatial relation among objects, which together with the object features was used to classify high resolution image, and map LU. The results indicated that features and spatial relation are effective in differentiating different LU classes. For example, shape information is useful in differentiating between built lands from barren soils; spatial relations are useful in differentiating between commercial and residential areas; and area information is useful in differentiating forest land from urban trees. We also discuss the challenges and future prospects of object-based methods in mapping land use.

Andrea Laliberte, USDA ARS Jornada Experimental Range, Las Cruces, NM

"Approaches for mapping and monitoring arid rangelands with object-based image analysis and hyperspatial imagery"
View: Abstract, Presentation as PowerPoint, Presentation as PDF.

At the USDA Agricultural Research Service Jornada Experimental Range (JER) in southern New Mexico, remote sensing research is focused on finding new methods for mapping and monitoring rangelands, and on relating ground-based surveys to remotely sensed information. This presentation will give an overview or recent research at JER involving object-based image analysis and hyperspatial imagery ranging from QuickBird satellite imagery, aerial photography, imagery acquired with unmanned aircraft, to ground-based plot photography. The QuickBird image was segmented at multiple scales to map shrubs at a fine scale and other vegetation communities at coarser scales. We were able to identify and map 87% of shrubs greater than 2 m2. A decision tree was used as an effective tool for sorting through the numerous input features and reducing them to a manageable rule set for classification. The overall accuracy was 80% for the optimal segmentation scale. In this case, we used ground-based plot photography as ground truth for the QuickBird remote sensing analysis. We compared image-based estimates of vegetation cover with line-point-intercept (LPI) measures for 50 plots (2.5 m x 3.5 m). The images were transformed from the RGB (red, green, blue) to the IHS (intensity, hue, saturation) color space. Object-based image analysis was used to classify the images into soil, shadow, green vegetation, and senescent vegetation using a masking approach and combination of membership rules and nearest neighbor classification. The correlation coefficients between LPI- and image-based estimates for the four classes ranged from 0.88 to 0.95. The object-based image approach was less labor and time intensive than the LPI method and has the potential to be incorporated into rangeland monitoring protocols. Our latest tool for rangeland monitoring is an unmanned aerial vehicle (UAV), capable of acquiring 5 cm resolution imagery from a 150 m above ground flying height. While the imagery presents some challenges for orthorectification and mosaicking, object-based image analysis has already proven to be highly successful. The very high-resolution imagery allows for identification of individual plants, patches, gaps, and patterns not previously possible, and will allow for assessment of rangeland health and ecosystem change at multiple scales. While this is a project in progress, initial mapping results indicate accuracies in the high 90% range. Based on this recent research, it is apparent that object-based image analysis will play a major part in future rangeland mapping and monitoring at multiple scales. It is also apparent that new tools in object-based image analysis will be needed, specifically in the area of object-based accuracy assessment.

Marguerite Madden, Center for Remote Sensing and Mapping Science (CRMS)
Department of Geography, University of Georgia

"A Comparison of Object-based Image Analysis and Manual Interpretation of Vegetation Communities"
View: Abstract, Presentation as PowerPoint, Presentation as PDF.

The Center for Remote Sensing and Mapping Science (CRMS) at The University of Georgia Geography Department has worked cooperatively with state and federal resource management agencies for 23 years to develop digital data in support of decision making for the preservation of natural and cultural resources in 21 National Park Units of the southeastern United States. Originally created from manually interpreted large-scale color infrared (CIR) aerial photographs using geospatial tools such as geographic information systems (GIS), softcopy photogrammetry, Global Positioning System (GPS) field surveys and GIS modeling procedures, there is a desire to develop efficient methods for vegetation data updating, maintenance of current databases and the preservation of expert knowledge. Object-based image analysis (OBIA) is being explored as a new and exciting method for combining the knowledge of experienced manual interpreters and contextual, object-based vegetation classification techniques. The study area is Great Smoky Mountains National Park located in the southern Appalachian Mountain region of the eastern United States. Mapped as part of the National Park Service and U.S. Geological Survey National Vegetation Mapping Program, digital data sets of association-level vegetation classes developed for the 2,000 km2 park were correlated with physical terrain characteristics of elevation range, slope and aspect to develop models of vegetation patterns. These models, in turn, were used to derive context rules for object-based classification. Progress on the success of this transition from manual to object-based automated techniques will be reported, along with geovisualizations used to assess vegetation patterns and provide information that can be used for thematic attribution and geometric quality control. It is hoped that contributions from these studies will provide resource managers with geospatial tools and spatial information needed for creating sound management plans.

Jeff Milliken, US Bureau of Reclamation

"Current Projects Using OBIA at the U.S. Bureau of Reclamation"
View: Abstract, Presentation as PowerPoint, Presentation as PDF.

The U.S.B.R. is using Object Based Image Analysis for a number of current and planned projects. OBIA is used for a variety of purposes including the delineation and mapping of riparian and marsh vegetation attributes, automated signature generation for crop classification purposes, change detection analysis, and delineation of vegetated surfaces in urban areas. Current and planned projects include the Lower Colorado River Accounting System, Central Valley Habitat Monitoring Project, Klamath Marsh vegetation mapping with U.S. Fish and Wildlife Service, and urban vegetation mapping with the California Department of Water Resources.

Jeff Milliken has an M.A. in Geography from San Francisco State University, and B.S. in Geology from Colorado State University. He has worked in the private and public sectors as a geologist and remote sensing/GIS specialist for 30 years. He is currently a Remote Sensing and GIS Specialist with the Bureau of Reclamation doing work for both the Lower Colorado and Mid Pacific Regional Offices. His current work primarily focuses on using and developing Remote Sensing and GIS applications for mapping and monitoring land cover types in support of water resource related projects.

Doug Stow, San Diego State University, Department of Geography

"Object-Based Monitoring of Fine-Scale Vegetation Changes in Shrubland Habitat"
View: Abstract, Presentation as PowerPoint, Presentation as PDF.

Most remote sensing-based studies of vegetation and land cover change utilize per-pixel change detection approaches. However, object-based approaches may be more appropriate for change detection based on high spatial resolution imagery, particularly when attempting to delineate discrete land cover change features such as removal or morality of plants or formation of trails. We demonstrate the potential for an integrated segmentation and object-based classification approach to detecting change from multi-temporal data sets derived from very high spatial resolution (0.2 and 1 m ground sampling distances) visible (V) and near infrared (NIR) imagery.

Digital airborne imagery was captured over two shrubland study areas in San Diego County using a frame center matching approach that facilitates precise image registration. The multi-temporal V/NIR data sets were input directly to a segmentation routine to generate change and no-change image objects at a single (fine) scale. Segment-based (i.e., multi-pixel) classification was implemented in a manner that attempted to exploit spectral, temporal, shape, and contextual attributes. For the 0.2 m data set, the focus was on delineating new trails created by cross-border immigration and smuggling activities. Monitoring vegetation conditions in a mixed habitat and recreational reserve was the focus of the study based on the 1 m multi-temporal data set.

Results from both study sites demonstrate that detailed land cover changes were delineated and identified that corresponded to both direct human disturbance (e.g., urban edge, recreation, and immigration) and drought effects. Based on field validation data, habitat change maps from the object-based approach were more reliable than a pixel-based approach, and habitat change features were more realistically represented in terms of shape and size. New trail features can be delineated just as effectively with simpler per-pixel approaches.

Karin Tuxen, University of California, Berkeley, College of Natural Resources

"Multi-scale functional mapping of tidal marshes using OBIA"
View: Abstract, Presentation as PowerPoint, Presentation as PDF.

Karin Tuxen, University of California, Berkeley, College of Natural Resources
Maggi Kelly, Department of Environmental Science, Policy, and Management, University of California, Berkeley, CA, 94710

Salt marsh restoration projects seek to restore wetlands back to their natural function, which includes complex and heterogeneous vegetation pattern. As salt marshes are multi-scale, hierarchically structured, and contain many interacting components, the question remains about how to map vegetation pattern in a multi-scale manner. Restoration goals often target species or biological process with specific objectives for use or function of the marsh at specific scales. In this study, we map three salt marshes in the San Francisco Estuary (two restored and one natural) using color-infrared aerial photographs and multi-scale object-based classification, with respect to three wetland functions: salt marsh harvest mouse habitat, California clapperrail habitat, and nutrient cycling/carbon sequestration. Pattern and scale are important for each of these three functions, but important in different ways and at different scales. Object-based image analysis (OBIA) is applied to very high-resolution data (20 cm), for segmentation into functional patches at multiple scales in a semantic structure. We will share results of our multi-scale maps, as well as results of accuracy assessments and land cover pattern metric analysis.