Daniel L. Civco

Artificial neural networks for land-cover classification and mapping

Abstract. Artificial intelligence approaches toward image processing and pattern recognition are perceived as an alternative to, and an improvement over, traditional statistically-based procedures. Of particular interest to the satellite remote sensing community are artificial neural networks. This article describes the application of such an approach to the problem of deriving land-cover information from Landsat satellite Thematic Mapper (TM) digital imagery. The techniques being developed are ones that will provide more accurate and useful data for use with geographical information systems.

Development of a geospatial model to quantify, describe and map urban growth

Abstract In the United States, there is widespread concern about understanding and curbing urban sprawl , which has been cited for its negative impacts on natural resources, economic health, and community character. There is not, however, a universally accepted definition of urban sprawl. It has been described using quantitative measures, qualitative terms, attitudinal explanations, and landscape patterns. To help local, regional and state land use planners better understand and address the issues attributed to sprawl, researchers at NASAs Northeast Regional Earth Science Applications Center (RESAC) at The University of Connecticut have developed an urban growth model. The model, which is based on land cover derived from remotely sensed satellite imagery, determines the geographic extent, patterns, and classes of urban growth over time. Input data to the urban growth model consist of two dates of satellite-derived land cover data that are converted, based on user-defined reclassification options, to just three classes: developed, non-developed, and water. The model identifies three classes of undeveloped land as well as developed land for both dates based on neighborhood information. These two images are used to create a change map that provides more detail than a traditional change analysis by utilizing the classes of non-developed land and including contextual information. The change map becomes the input for the urban growth analysis where five classes of growth are identified: infill , expansion , isolated , linear branch , and clustered branch . The output urban growth map is a powerful visual and quantitative assessment of the kinds of urban growth that have occurred across a landscape. Urban growth further can be characterized using a temporal sequence of urban growth maps to illustrate urban growth dynamics. Beyond analysis, the ability of remote sensing-based information to show changes to a communitys landscape, at different geographic scales and over time, is a new and unique resource for local land use decision makers as they plan the future of their communities.

Mapping urban areas on a global scale: which of the eight maps now available is more accurate?

Eight groups from government and academia have created 10 global maps that offer a ca 2000 portrait of land in urban use. Our initial investigation found that their estimates of the total amount of urban land differ by as much as an order of magnitude (0.27–3.52 ×106 km2). Since it is not possible for these heterogeneous maps to all represent urban areas accurately, we undertake the first global accuracy assessment of these maps using a two-tiered approach that draws on a stratified random sample of 10 000 high-resolution Google Earth validation sites and 140 medium-resolution Landsat-based city maps. Employing a wide range of accuracy measures at different spatial scales, we conclude that the new MODIS 500 m resolution global urban map has the highest accuracy, followed by a thresholded version of the Global Impervious Surface Area map based on the Night-time Lights and LandScan datasets.

Road Extraction Using SVM and Image Segmentation

Accurate road information is vital for transportation applications, including as part of geographical information systems (GIS). This article reports on the development of a two-step approach for road extraction that utilizes pixel spectral information for classification and image segmentation-derived object features. In the first step, support vector machine (SVM) was employed merely to classify the image into two groups of categories: a road group and a non-road group. For this classification, support vector machine (SVM) achieved higher accuracy than Gaussian maximum likelihood (GML). In the second step, the road group image was segmented into geometrically homogeneous objects using a region growing technique based on a similarity criterion, with higher weighting on shape factors over spectral criteria. A simple thresholding on the shape index and density features derived from these objects was performed to extract road features, which were further processed by thinning and vectorization to obtain road centerlines. The authors conclude that the proposed approach worked well with images comprised by both rural and urban area features.

A competitive pixel-object approach for land cover classification

This paper describes a novel remote sensing land cover classification approach named competitive pixel-object classification, based on Bayesian neural networks and image segmentation. This approach makes use of both pixel spectral features and object features resulting from image segmentation through a competitive mechanism to resolve the problem of spectral confusion caused by reflectance similarity of some land cover types that traditional pixel-based classification cannot resolve. The competitive pixel-object method reduces the unreliability of object feature information produced by over- or under-segmentation of the image through a competitive mechanism. The experiment shows that the competitive pixel-object approach produces higher classification accuracy than either pixel-based classification or object-oriented classification.

Using Satellite Images to Classify and Analyze the Health of Hemlock Forests Infested by the Hemlock Woolly Adelgid

A method is described to classify stands of eastern hemlock by health condition, at the landscape level, using remote sensing. The hemlock woolly adelgid has been a major cause of hemlock decline in Connecticut since 1985, resulting in varying degrees of defoliation in the region. A 1985 Landsat Thematic Mapper (TM) image was classified to develop a base line of once healthy hemlock stands. Radiance normalization and non-hemlock masking techniques were used to pre-process a 1995 TM image. Several techniques were used to transform the 1995 TM image; each was followed by cluster analysis to separate hemlocks into four levels of tree vigor. We evaluated 600 trees at 150 sites across the study area using the USFS Crown Condition Rating Guide. These field data were used to measure the accuracy of various health classification techniques. The Modified Soil Adjusted Vegetation Index-2 (MSAVI2) transform provided the best overall accuracy, 82.1%, for classifying hemlock according to tree vigor. Non-parametric statistics were used to determine if there were any significant variations in distribution of hemlock pixels by health class in association with features in the landscape. Several features were found to be statistically significant at a confidence level of 0.001. These were aspect of slope, hydrology group (infiltration rate), depth to bedrock, soil order, drainage class (hydraulic conductivity), and surface texture.

Relationships of salt-marsh plant distributions to tidal levels in Connecticut, USA

A three-year study of Connecticut, USA, salt-marsh vegetation was undertaken to determine the relationship of its distribution on the marsh surface to tidal levels, particularly mean high water (MHW) as measured on each of three sites representing different tidal amplitudes. Elevations and species present were measured on 1-m2 grids in 10x 70-m belt transects at each site. After the data were subjected to discriminant analysis and other standard statistical procedures, the results showed that 98.4% of all observations ofSpartina alterniflora Loisel. occurred at or below MHW. The data can aid in salt-marsh restoration by offering a reliable indicator of what species should be planted when restored elevations and on-site MHW are known.

Changes in Connecticut salt-marsh vegetation as revealed by historical aerial photographs and computer-assisted cartographics

Procedures are discussed for the interpretation of historical aerial photographs for salt-marsh vegetation mapping, as are techniques for computer-assisted analysis of digital vegetation maps. The mappings indicate an increase in the coverage by the low marsh speciesSpartina alterniflora Loisel. at three marsh sites studied in photographs from the period 1934–1981. It is hypothesized that changes in salt-marsh vegetation may be in response to natural tidal fluctuations or to management practices.

The fragmentation of urban landscapes: global evidence of a key attribute of the spatial structure of cities, 1990–2000

The fragmentation of urban landscapes – or the inter-penetration of the built-up areas of cities and the open spaces in and around them – is a key attribute of their spatial structure. Analyzing satellite images for 1990 and 2000 for a global sample of 120 cities, we find that cities typically contain or disturb vast quantities of open spaces equal in area, on average, to their built-up areas. We also find that fragmentation, defined as the relative share of open space in the urban landscape, is now in decline. Using multiple regression models, we find that larger cities are less fragmented, that higher-income cities are more fragmented, that cities with higher levels of car ownership are less fragmented, and that cities that constrain urban development are less fragmented. We recommend that making room for urban expansion in rapidly growing cities should take into account their expected fragmentation levels.

A Technique to Identify Changes in Hemlock Forest Health over Space and Time Using Satellite Image Data

The objective of this study was to develop a technique to classify health of eastern hemlock stands using historical satellite images. While remote sensing and geographic information systems have been used successfully to classify forest health using recent images, applying this process to older images is problematic because contemporaneous field data are not available to measure the accuracy of the classification of historical images. Data ranges were established for each hemlock health class using a contemporary image and field data. These ranges were used to level-slice archived images to create a series of health-class maps that show changes in forest health over time. By applying cross-tabulation procedures to pairs of classified images, it is possible to construct a transition map that indicates how the hemlock health class of each pixel in the images of the study area has changed over time. The resulting maps provide a look back at forest conditions of the past and can be used to identify areas of special interest.