OBJECT-ORIENTED LAND COVER CLASSIFICATION OF SPATIALLY-ENHANCED SATELLITE REMOTE SENSING DATA

• Sponsoring Institution: National Institute of Food and Agriculture
• Project Status: TERMINATED
• Funding Source: HATCH
• Reporting Frequency: Annual
• Accession No.: 0218964
• Project No.: CONS00848
• Project Start Date: Oct 1, 2009
• Project End Date: Sep 30, 2011

Project Director
Civco, D. L.

Recipient Organization
UNIV OF CONNECTICUT
(N/A)
STORRS,CT 06269

Performing Department
Natural Resources & the Environment

Non Technical Summary
Over the past several years, researchers at the University of Connecticut?s Center for Landuse Education and Research (CLEAR) have used Landsat TM and ETM imagery to map land use and land cover change for five dates over a 21 year period for the state of Connecticut. While this type of information has proven very valuable to land use decision makers at a regional scale, the spatial and thematic detail has been limiting at the more localized level. Now that powerful computing processing technology and disk storage space has become more affordable, and with the abundance of available image types, researchers at CLEAR have begun to look at the creation of a finer spatial, more thematically rich, land use and land cover map for Connecticut. In order to detect, identify, classify, and delineate urban land use features with improved spatial resolution and thematic distinction over the current Connecticut's Changing Landscape land cover two elements are required: (1) remote sensing data with better spatial resolution, while maintaining the spectral resolution afforded by multispectral instruments, such as the TM and ETM on Landsat and,(2) intelligent image processing and classification procedures, which are object-based, rather than myopic per-pixel techniques, and allow for the inclusion of spatial operators and human expertise. Development of a new protocol for mapping Connecticut?s land use and land cover will enhance efforts to distinguish among different types of urban development, and at a finer level of detail. We aim to extract at least the following categories of urban land use: (1) residential, (2) commercial and industrial, and (3) roads. Further, we expect to be able to characterize the density of development, such as single-family rural residential versus high density urban and suburban residential. Our overall goal is to extract land cover and land use information with a finer spatial resolution and deeper thematic distinction than has been done to date in Connecticut. Whereas there are high resolution (< 2m) remote sensing data available, they typically are panchromatic, or, if multispectral, lack either near or middle infrared bands, which are invaluable in discriminating among many different land cover types. Conversely, Landsat TM and ETM data, while possessing those critical reflective infrared bands, lack the spatial resolution of aerial or commercial satellite remote sensors. Our objective is to ?marry? the two, preserving the attributes of both in their union, and to segment and classify this enhanced imagery for the purpose of extracting meaningful land use classes. Further, we will employ advanced remote sensing image analysis techniques that exploit both the spectral and spatial properties of the image data. The approaches to be investigated will leverage recent innovations in remote sensing image exploitation - notably spatial resolution enhancement, image segmentation, and object-oriented classification - in the development of a protocol to derive land use information with a high level of spatial detail.

Animal Health Component

(N/A)

Research Effort Categories

Basic

30%

Applied

40%

Developmental

30%

Classification

Knowledge Area (KA)	Subject of Investigation (SOI)	Field of Science (FOS)	Percent
608	7210	2060	100%

Knowledge Area
608 - Community Resource Planning and Development;

Subject Of Investigation
7210 - Remote sensing equipment and technology;

Field Of Science
2060 - Geography;

Keywords

land use

land cover

remote sensing

image segmentation

object-oriented classification

resolution enhancement

data fusion

connecticut

Goals / Objectives
We propose to perform an object-oriented land cover classification of spatially-enhanced satellite remote sensing data for a sample of towns in Connecticut spanning the urban-suburban-rural gradient representing different levels and patterns of development. Specific objectives are to: (1) perform a needs assessment of Connecticut's land use information user community in to ascertain desired levels of spatial and thematic resolution for the state's land use and land cover data layer; (2) assess the effectiveness of different algorithms for data fusion of high spatial resolution with high spectral resolution remote sensing data; (3)develop resolution-enhanced image segmentation and object-oriented classification procedures that reflect the user community's land use and land cover needs; and (4) develop protocols for statewide land use and land cover classification from spectrally-enhanced, high spatial resolution remote sensing imagery. Results of this study will be used to guide future statewide land cover mapping efforts, which, to date, have been done with Landsat data with a 30-meter (per pixel) ground resolution.

Project Methods
The approaches to be investigated will leverage recent innovations in remote sensing image exploitation - notably spatial resolution enhancement, image segmentation, and object-oriented classification - in the development of a protocol to derive not only land cover but also land use at a resolution finer than that afforded by 30-meter Landsat, which is the basis for current Connecticut statewide land cover mapping. To gain a better understanding of land use data needs in the state, a survey will be conducted of a representative sample of land use managers and planners at the state and local levels to determine current usage of land use data and perceived requirements - spatial, thematic, and temporal. Guided by the recommendations for land use spatial and thematic characteristics, appropriate remote sensing data will be acquired from existing inventories for a set of towns representative of Connecticut's land use diversity. A resolution enhancement algorithm will be applied to the selected high spatial resolution and moderate resolution multispectral remote sensing data. The optimal balance between spectral and spatial preservation will be assessed using root mean square and correlation as statistical measures. Spectral quality of the derived resolution-enhanced imagery will be determined by numerical, as well as visual, comparison with the source multispectral data. The PI has developed a unique approach for assessing the spatial quality of resolution-enhanced imagery by examining the Fourier spectral of original and derived imagery. A commercial-off-the-shelf image segmentation and object-oriented analysis software tool will be used in land use classification. Selection of training areas representative of the desired land use classes will be performed in a desktop GIS with those data reformatted and exported for use in a data mining tool, which will generate a classification tree to be used in object-based land use classification. Accuracy assessment will be performed during decision rule development as well as after classification maps have been created, the latter by way of ground verification of a sample of stratified random points. Based on land use classification results, and their degree of satisfying user-requested land use thematic distinction and spatial resolution (minimum mapping unit), coupled with the accuracy of those results, a standard operating procedure will be formulated for recommended use in future Connecticut statewide land use and land cover classification projects. Seminars and informal workshops will be held with select representatives from municipalities, CT DEP, and other state, regional, and local interests.

Progress 10/01/09 to 09/30/11

Outputs
OUTPUTS: Land use and land cover continues to be a valuable source of information for resource managers, used in activities such as by the Connecticut (CT) Department of Energy and Environmental Protection (DEEP) in environmental decision-making, the CT Council on Environmental Quality (CEQ) for its annual report on state of CT's environment, the U.S. Environmental Protection Agency (EPA) Long Island Sound Study (LISS) for tracking progress of its long-term management plan, and the National Oceanic and Atmospheric Administration (NOAA) as the model for its national web-based forest fragmentation analysis tool (csc.noaa.gov/digitalcoast/tools/lft). Maintaining and expanding current land cover offerings, therefore, remains a high priority. CLEAR's Connecticut's Changing Landscape (CCL) land cover data and derived products are one of the most utilized resources on the CLEAR website with more than 1,000 hits a month. The CCL land cover products are based on Landsat moderate-resolution (30 meter) satellite imagery. Current CCL land cover has been expanded to include Westchester County and northern Long Island, NY to be current and consistent with land cover data covering Connecticut. These data are for the years 1985, 1990, 1995, 2002, and 2006. A 2010 land cover product is currently being developed, consistent with the other dates, and covering the same areas of Connecticut, Westchester County and northern Long Island, NY. Furthermore, data for all six dates are being derived for southern Long Island, NY. The CCL land cover is also used for forest fragmentation analysis and impervious surface estimation, which have recently been completed for the entire land cover region. Data can be accessed through the CLEAR website (clear.uconn.edu), and a new interactive mapping webpage is under development, taking advantage of new technology which will provide users with improved access and query abilities. PARTICIPANTS: Daniel L. Civco, Professor James D. Hurd, Research Associate Qian Lei, Graduate Assistant TARGET AUDIENCES: Land use managers and decision makers PROJECT MODIFICATIONS: Results to date will contribute to the development of a series of protocols for the extraction of detailed land use information from multi-source data.

Impacts
RapidEye satellite remote sensing imagery, with a five-meter spatial resolution, and Landsat Enhanced Thematic Mapper (ETM) imagery, with six reflectance bands, were acquired for the University of Connecticut Storrs campus. Various techniques were assessed for their effectiveness for merging high spatial resolution (RapidEye) with high spectral resolution (ETM) remote sensing data. A method based on principal components analysis (PCA) was selected as the optimal fusion algorithm due to the balance between spectral and spatial preservation of resolution-merge products. Spectral quality of the derived pan-sharpened imagery was determined by numerical, as well as visual, comparison with the source multispectral data. The PI has developed a unique approach to assess the spatial quality of resolution-enhanced imagery by examining the Fourier spectra of original and derived imagery. A hierarchical land use classification scheme system including two levels and 11 land use classes has been developed for image segmentation and object-oriented classification. Hierarchical image segmentation was performed using eCognition. Object-based spectral and spatial attributes were selected to discriminate best among the land use classes and exported into an attribute table. Training sample selection of each land use class was performed in ArcGIS. A decision tree was established using the See5 data mining tool based on the corresponding attributes of the training samples. The significance of each attribute for separating land use classes and accuracy assessment for the training samples was computed during decision rule development. An object-oriented classification was then performed in eCognition by applying the derived decision tree. Based on the visual interpretation of the classification results and confusion matrix of the training samples, Level 1 (6 general classes) of the classification hierarchy was found to have relatively high classification accuracy. Overall classification accuracy at Level 2 (11 classes) was found to be comparatively low with much confusion between classes. For example, there were many misclassifications between agriculture grasslands and lawns, which can be attributed to spectral similarity of these two classes. Several artificial turf athletic fields were misclassified as impervious surfaces, owing to their spectral similarity and similar regular geometry. The building and transportation classes had the poorest accuracy. Deciduous forests and coniferous forests were easily separated, as would be expected. These results indicate that the RapidEye imagery is insufficient to derive the type of land use and land cover information required by federal, state, and municipal users, especially the latter who need data at the Anderson Level III level (e.g., low-, medium- and high-density residential). Ancillary information, such as LiDAR-derived Digital Elevation Models, parcel, and building data, is required to achieve spatially- and thematically-detailed land use information. Research continues on the development of a set of protocols for multi-source data fusion and data mining aimed at achieving detailed land use information.

Publications

• Civco, D.L. 2012. Assessing spatial fidelity of resolution-enhanced imagery using Fourier analysis. Phtotogrammetric Engineerting and Remote Sensing (in review)
• Angel, S., J. Parent, D.L. Civco, and A.M. Biel. 2011. The Dimensions of Global Urban Expansion: Estimates and Projections for All Countries, 2000-2050. Progress in Planning 75(2):53-108.
• Angel, S., J. Parent, D.L. Civco, and A.M. Biel. 2011. Making Room for a Planet of Cities. Policy Focus Report, Lincoln Institute of Land Policy, Cambridge, MA. ISBN 978-55844-212-2. 72 p.
• Angel, S., J. Parent and D.L. Civco. 2010. Ten compactness properties of circles: measuring shape in geography. Canadian Geographer 54(4)(Winter):441-461.
• Civco, D.L., A. Chabaeva, and J. Parent. 2010. KH-series satellite imagery and Landsat MSS data fusion in support of assessing urban land cover growth. Proc. SPIE Remote Sensing Europe, Volume 7478, Berlin, Germany. 12 p.

Progress 01/01/10 to 12/31/10

Outputs
OUTPUTS: The overall purpose of this project is to perform a hierarchical, object&#8208;oriented land cover classification of spatially&#8208;enhanced satellite remote sensing data for a sample of towns in Connecticut. Expected results will be used to guide future land cover mapping efforts for Connecticut. To meet this purpose, four primary objectives are being pursued: 1) perform a needs assessment of Connecticut's land use information user community in order to ascertain desired levels of spatial and thematic resolution for the state's land use and land cover data layer; 2) assess the effectiveness of different algorithms for data fusion of high spatial resolution with high spectral resolution remote sensing data; 3) develop resolution&#8208;enhanced image segmentation and object&#8208;oriented classification procedures that reflect the user community's land use and land cover needs; and 4) develop protocols for statewide land use and land cover classification from spectrally enhanced, high spatial resolution remote sensing imagery. Objective 1) A draft land use/land cover user needs survey has been developed for distribution through an online survey system. The survey is currently be reviewed by these investigators and members of the CT Geospatial Information Systems Council's Land Use/Land Cover Subcommittee. The purpose of the survey is to gain a sense of the information needs of the land use/land cover user community in Connecticut. As such, the survey solicits information regarding: a) the user's employment sector, type of work performed, and current use of land use and land cover information; b) land use/land cover needs, that is, what type of land use/land cover information would be most useful to the user and the type of work performed; c) what classification scheme would the user prefer in a statewide land use/land cover product; d) what spatial detail would the prefer in a statewide land use/land cover product. Objective 2) Research is currently using RapidEye satellite imagery, collected in May 2010. The RapidEye sensor provides five band multispectral images (blue, green, red, red edge, and near infrared) at a resampled spatial resolution of 5-meters. Rapideye imagery has been acquired for the University of Connecticut Storrs campus area and for central Hartford to provide a range of land cover types for assessment. Upon completion of assessing the Rapideye imagery for classification, other sources of image data will be tested including the resolution merge of different image types to derive suitable imagery to meet the needs of statewide land use/land cover mapping. Objective 3) Using the RapidEye imagery, various image segementation layers are being generated based on numerous spectral and spatial input parameters required by the software. Different segmentation levels are being assessed to determine which provide optimal results based on the 5-meter spatial resolution, five-spectral band Rapideye data. Objective 4) Protocols will be developed based on the findings of the first three objectives. PARTICIPANTS: Not relevant to this project. TARGET AUDIENCES: Not relevant to this project. PROJECT MODIFICATIONS: Not relevant to this project.

Impacts
The investigators have initiated a land use/land cover user needs survey. Following final review, the survey will be made available to the user community for their input in early 2011. RapidEye satellite remote sensing imagery has been acquired for two study areas in Connecticut and a graduate student assigned to work on the image segmentation and classification aspect of the project. Following acquisition of the survey results, the information will be applied in the implementation of the image segmentation and classification process to develop a final protocol for the development of statewide land use/land cover for Connecticut. In addition to the RapidEye imagery, other types of geospatial data are being examined for their utility in land use classification, including USDA National Agriculture Imagery Program (NAIP) four-band (blue, green, red, near infrared), 1-meter resolution digital remote sensing data acquired in September 2010 and LiDAR-derived 3-meter resolution digital elevation data. The springtime, "leaf-off" RapidEye data and the summertime, "leaf-on" NAIP imagery will be assessed for their multitemporal content, enhancing the ability to extract both land use and land cover information. This effort will require an evaluation of data fusion techniques to ensure spatial registration among disparate spatial resolutions. Decision Trees incorporating these multisource data will be developed for use in object-oriented land use and land cover.

Publications

• No publications reported this period