Considerations in the Use of High Spatial Resolution Imagery

ABSTRACT

With the advent of improvements in remote sensing technology, higher spatial and spectral resolution images have become more widely available for applications research purposes. New sensors are capable of sub-meter resolution and fine spectral resolution, opening the field to applications previously not considered with digital image data. Very high spatial (1-5m) and hyperspectral imagery bring with them, however, a whole new set of issues associated with the resolution of individual scene elements. Mapping activities in the Chesapeake Bay Watershed being conducted by the mid-Atlantic Regional Earth Science Applications Center (RESAC) have examined some of these issues and attempted to develop methods to compensate for them. We illustrate a specific research application, impervious surface mapping, and provide an overview for consideration by the potential user community, particularly resource managers within Local Government agencies.

INTRODUCTION

The advent of high spatial resolution imagery (1-5m resolution) has potentially important implications for a wide range of resource management issues, including Local Government planimetric maps, land cover assessments, transportation planning, growth management, and numerous other applications. A variety of managers are hoping to make use of these new image data sets, providing they can provide cost effective solutions to their needs which have traditionally been met with manual interpretation of aerial photography. There are, however a number of practical considerations in the use of high resolution digital imagery. Besides the training required for managers to work with digital data, image processing systems, and geographic information systems, high resolution imagery frequently requires substantial upgrades in computer processing power and data storage requirements. Provided these needs are met (they are less limiting as personal computers have attained the power of previous generation high-end workstations), there are a number of other considerations that come with the powerful advantages conveyed by high resolution imagery.

The objective of this paper is to provide a brief overview of some user considerations associated with the use of this new source of geospatial information. We provide a specific case study using a variety of high spatial resolution image data for impervious surface area estimation, and include high spectral resolution imagery for an additional dimension with potential utility to similar applications.

ADVANCES IN LAND SURFACE MAPPING PERMITTED BY HIGH RESOLUTION IMAGERY

Fine scale mapping of many features and complex landscape mosaics, including suburban development and coastal zones, have become possible with improved image data and associated geospatial technologies. Suggestions and uses for high resolution airborne and satellite imagery have ranged from the extraction of soil conditions to measurement of agricultural forest buffer zones, extraction of geographic information system (GIS) data for new road systems, and mapping of newly developed impervious surfaces. High-resolution imagery also carries another significant advantage: textural identification of ground features. Whereas previous coarse resolution images are difficult to use without a-priori knowledge of ground features, the increased textural information available in finer resolution imagery allows for improved interpretation based on the shape and texture of ground features. For example, buildings become identifiable features represented in multiple image elements (pixels), rather than in a single pixel. Sparsely vegetated fields appear as identifiable exposed soil parcels with traces of vegetation throughout, rather than as a complex mixture of spectral signatures. As a result, methods previously utilized on panchromatic air photos for feature delineation purposes can be applied to digital imagery to supplement, for example, spectral land cover and land use type classifications.

Difficulties associated with mixed pixels (e.g. pixels that overlap two or more significantly different surfaces and thus show spectral properties of each) may be reduced with higher resolution imagery, of which fewer will cover distinctly different surfaces. Riparian buffers strips and agricultural field boundaries may be clearly delimited as separable scene elements. Increased spatial resolution also means that individual parcels, such as buildings and roads, become more clearly visible and interpretable to the human eye. This same quality allows for more heterogeneous areas to be distinguished, thus farmers are able to identify exact locations of stressed or diseased crops, and small scale ecosystem managers (e.g., golf course administrators) can comprehensively assess the extent of biological invasion (weed growth) or change in wetlands. Moreover, fine resolution imagery is often useful for interpreting coarser resolution imagery, and can be used in effective sampling strategies. The increased spatial perspective available at finer scales has the potential to revolutionize the methods by which ecological, geological, managerial, and developmental operations are pursued.

ISSUES ARISING FROM THE USE OF HIGH RESOLUTION IMAGERY

Acquisition Issues

In high spatial resolution systems, an additional difficulty is presented in scene element illumination. A high solar elevation angle reduces shadows, which become far more prominent in high-resolution images than in spectral mixtures within coarse resolution imagery. A 10-meter object will almost fully shadow at least two pixels in a 1-meter resolution image when the sun elevation shifts just 10o. These shadows may be difficult to identify and classify, and are thus the source of some error in image interpretation (Figure 2d). Moreover, because three-dimensional objects intercept light differently they appear non-uniform in illumination and spectral reflectance (Figure 2b). In the case study presented below, both houses and tree canopies appear significantly different with illumination conditions, sometimes causing misidentification.

Until the advent of the IKONOS satellite and similar orbiting high-resolution imaging systems, the preferred method of obtaining high-resolution images of small areas was with the use of airborne systems. The most prominent difficulty with airborne systems are those which affect actual flying for data acquisition. To map an extensive area (~50 square kilometers) in a low-flying aircraft can take several hours. In this time period, weather systems and lighting changes occur, and geo-referencing systems may become less accurate (see case study below). More importantly for management applications, it is difficult to obtain consistent multiple date images over the same area for direct comparison. In heavy air traffic areas, clearances for long term scanning may be difficult to obtain, and air traffic controllers may restrict the length of time allocated collecting data. Satellite systems are prone to a slightly different issue: to obtain specific ground images without passing immediately overhead requires scanning at an incidence angle to the desired site, which causes three dimensional objects to be misplaced laterally.

Interpretation Issues

Technological innovation and the commercial availability of such systems as the Charge Coupled Device (CCD) camera have increased sensor capacity for higher spectral resolution and multi-spectral imagery. These advances are useful for finer distinctions in vegetation type and function, including plant stress, disease, and pest detection in forests and crops. Higher spectral capacity systems are also important in geological applications, soil identification, and even paved surface quality control. The parameters controlling the remotely sensed image are more comprehensively defined by the presence of identifiable absorption and emission spectra. High spatial resolution imaging systems produce large data volumes that must be pre-processed during and after acquisition before they can be used for any given application. Systems with high spatial capabilities are often still limited in their capacity to record very fine spectral resolution data. Thus, the potential of identifying specific emission and absorption spectra may be limited. As a result, leaf-off vegetation may appear similar to impervious surfaces in shadows, and exposed soils may appear similar in every spectral band to some constructed impervious surfaces (Figure 2e). Systems that operate into the middle infrared portion of the electromagnetic spectrum may provide improved discrimination of these functionally different cover types.

CASE STUDY: MULTI-SENSOR COMPARISON OF IMPERVIOUS SURFACE AREA ESTIMATION

We focused a comparative study on a Public Park system within Montgomery County, Maryland. Rock Creek Regional Park is a closed deciduous canopy forest surrounded by residential neighborhoods, a well developed transportation system, and new land conversion projects. The park is continuous from mid-county into the District of Columbia, and includes increasingly rare contiguous closed forest within the Washington metropolitan area. The associated stream system drains much of the upper county watershed. Continued urban development in the lower county is leading to increased urban and suburban chemical pollution into the stream system, and requirements for improved flood control owing to an increasing proportion of impervious surface area within the watershed.

Although the single acquisition images used in this survey cannot be used towards a full understanding of watershed development impacts, they provide insight into the accuracy of the various remote-sensing systems measuring these changes. Three sets of multispectral imagery and a planimetric map derived from airphoto interpretation were processed through classification and re-sampling to provide a measure of the extent of impervious surfaces. Final products were produced as 30-meter resolution images classified as proportions of impervious surface area. Visual inspection of the data sources and statistical correlation of the final data products were used to identify the accuracy of the method compared to the actual surface condition as represented by the planimetric data.

Image Data Sets

Four image or map data sets were used in this comparative study: (a) a planimetric map of Montgomery County transportation and built structures, interpreted by the county Parks and Planning Commission using 1990 air photos; (b) multi-spectral high resolution imagery from the Airborne Imaging Spectrometer (AISA) at 2 meter resolution covering an 11.2 square kilometer area; (c) a derived product of estimated percent impervious surface from Landsat Thematic Mapper (TM) imagery; and (d) recently acquired 4-meter resolution image from the IKONOS satellite. The planimetric map was provided by Maryland National Capitol Parks and Planning Commission (MNCPPC) and was used in this study as a field validation reference. The map was used to classify 3 meter pixels as impervious if they intersected a mapped structure, road, or parking lot. Sidewalks and driveways were not included in the map.

The TM data used in this study were provided by the Mid-Atlantic Regional Earth Sciences Applications Center (RESAC) (Goetz et al. 2000), and contained a 30 meter estimation of impervious surface percentage based on a decision tree classification from April and May 2000 images (Smith et al. 2001). The decision tree used the planimetric map as training data for mapping subpixel impervious surface area. The IKONOS imagery was acquired in April to June of 2000, primarily during leaf-off conditions. The AISA imagery was acquired in late October 1999 by the 3DI LLC at two-meter resolution. In the imagery analyzed for this study the spectral resolution was 30 bands at approximately 6.5 nm spectral resolution (Table 1).

Methodological Approach

Impervious surfaces were identified in the AISA imagery using a maximum likelihood supervised classification algorithm within the ERDAS Imagine software. Training data included roads, parking lots, and rooftops in shadow and in direct illumination. A total of 50 classes and subclasses were identified as separable (including vegetation types) and split into permeable or impervious surfaces. Soil surfaces often displayed similar spectra to some impervious surfaces. Errors associated with including exposed soil in construction and agricultural areas were deemed greater than those resulting from excluding these, thus impervious type classifications containing the exposed soil areas were not analyzed here. The remaining two-meter impervious surface areas were grouped and assigned a common value. Using ArcInfo, these pixels were aggregated over a 30-meter area with an averaging function to determine an estimate for the percent impervious surface for a given location. Visual inspection of geo-referencing errors suggested that the AISA data were within a reasonable margin of error (1-2 pixels) for the area analyzed.

The IKONOS data were similarly classified, but with 21 spectral groupings. Again, the soil surfaces were excluded. The 4-meter resolution pixels were also assigned a common value and similarly aggregated into 30-meter groupings for direct comparison with the other data sets.

Thematic Mapper data were pre-processed using the planimetric data to train a decision tree classification approach (Varlyguin et al. 2001). The percent of impervious surface area was based on two season (leaf-on / leaf-off) TM images and other ancillary data sets.

The field validation data, as noted above, included the location of all known buildings, parking lots, and roads in the county as of 1990. Classes identified as impervious were retained, whereas gravel and dirt roads were excluded. Using ArcView, these values in the polygon shapefile were converted to a 1-meter grid with a common value indicating impervious surfaces. The grid was aggregated to 30-meter pixels using the average function again to obtain a 1990 estimate of impervious surface percentage

All of these data sets were mapped to the same projection, and all pixels with detected impervious surfaces were given a known value (100) over the 11.2 km2 area common study area. The images were then aggregated to 30 meters to obtain percent impervious surface, and layered in ArcView to directly compare results on a per pixel basis. Since there was construction in the vicinity of the park used between 1990 and the latest data set in 2000, all 30-meter pixels with values of zero were excluded from the analysis. Shadowing in some of the data sets also proved to be problematic, and because these shadows often fell in forested areas (Figure 2d), all pixels that showed a value of zero for any of the data sets were also removed. The final results were plotted against the planimetric map, and the degree of correlation and slope of the line was used as an indicator of the accuracy of the remote sensing method and data sets to accurately identify impervious surfaces.

Results

There was substantial variability among the derived maps of impervious surface area. Overall, they depicted an average 24.4% + 2.7% impervious area. The three image maps averaged 18.5% error in overall estimation within the study area (see Table 2). The relationship between the impervious surface on a per pixel basis generally resulted in low correlation values and linear regression slopes. F-test statistics indicate that there were weak relationships between the various image-derived maps and the planimetric map.

The images all identified similar areas as impervious, as seen in Figure 3. The degree to which they were classified as impervious varied between maps. Statistics indicate that although the AISA map agreed more with the planimetric map on a per pixel basis, it had a higher error overall than the IKONOS map. The TM map had the lowest success rate in determining the impervious surface coverage, despite having initially been "trained" using the planimetric data. This was largely because the resolution of the TM imagery was not sufficient to discriminate between the fine features present in the scene, thus individual houses and narrow roads were completely missed.

IKONOS had improved spatial resolution over TM, but as with TM it is a four-band sensor system which had difficulty spectrally discriminating some soils and paved surfaces. The IKONOS map under-represented the actual impervious surface amount for roads, and a number of points in otherwise closed forest were misclassified as impervious because of similar spectral response. These results are partly because the IKONOS image was acquired in leaf-off conditions, thus the difference between near IR and red bands (which provide the primary discrimination between vegetation and non-vegetation) was lost. Nevertheless, leaf-off imagery are typically required for impervious surface area estimation because many areas would otherwise be obscured by canopy coverage. The total percentage impervious surface was close to the actual value estimated from the planimetric map ('GIS' in Table 2), but did not correlate well with it on a per pixel basis.

The AISA imagery was acquired during leaf-on conditions (late fall) and had fewer difficulties segregating impervious surfaces from soils and vegetation. However, the presence of shadows resulted in mis-classification of areas otherwise completely covered with vegetation (particularly near the edges of fields). In other areas, flightline errors (turbulence and sudden shifts) were responsible for mis-referencing the imagery, thus reduced correlation between the derived impervious area and the planimetric map. The TM and IKONOS data both had high geometric fidelity, thus were less likely to suffer from misregistation errors, but the IKONOS and AISA systems were more accurate than the TM map because of their ability to distinguish fine features and individual scene elements (before being degraded to TM resolution).

SUMMARY AND CONCLUSIONS

This study illustrates some of the considerations involved in the use of high-resolution multispectral imagery for research applications. New opportunities arise from the improved capabilities of advanced sensor systems and associated geospatial technologies (e.g., GIS, GPS, spatial models, etc.), but these must be realistically considered in the context of increased data handling and computing requirements, as well as complexities introduced by the discrimination of individual scene elements under varying illumination and phenological conditions. Additional considerations arise from the data collection methods (aircraft versus satellite platforms) and from the spectral properties of the various scene elements that can now be individually resolved. Some of these issues were demonstrated through the case study focused on mapping impervious surface areas with various sensors and comparing image products derived using a consistent methodology. Because such data sets and related products are now more widely available to local government agencies, resource managers and a wide variety of other potential users, it is important that the considerations and limitations (as well as the advantages) of their use be known.

REFERENCES

Goetz, S. J., Prince, S. D. Thawley, M. M., Smith, A. J. and Wright, R. (2000). The mid-Atlantic Regional Earth Science Applications Center (RESAC): an overview. Available at www.geog.umd.edu/resac and on ASPRS CD-ROM in. American Society for Photogrammetry and Remote Sensing (ASPRS) Conference Proceedings, Washington DC.

Smith, A. J., Goetz, S. J., Prince, S. D. (2001). Subpixel estimates of impervious surface cover from Landsat Thematic Mapper imagery, Remote Sensing of Environment, forthcoming.

Varlyguin, D., Wright, R., Goetz, S. J., Prince, S. D. (2001). Advances in land cover classification for applications research: a case study from mid-Atlantic RESAC. Available at www.geog.umd.edu/resac and on ASPRS CD-ROM in. American Society for Photogrammetry and Remote Sensing (ASPRS) Conference Proceedings, St. Louis MO.