|
|
|
| Home | Projects | Members | Publications | Committee | Data | RSPA | Meetings | Contact |
| Introduction | |
| Remote Sensing Imagery | |
| Land Cover Classification |
|
| LandPro99 Dataset | |
| Land Use Classification | |
| References | Return to Environmental Assessment Home |
| Return to NCRST-E Home |
|
The Sope Creek watershed is a highly urbanized watershed on the northern edge of Marietta, Georgia, just north of the Atlanta Metropolitan area (Figures 1 and 2). This is a rapidly growing area that has preserved ecological diversity and natural water resources in the midst of urbanization. The Sope Creek was selected as a challenging land cover and land use classification project as the watershed has a diverse combination of urban features for which high resolution multispectral remote sensing data were available. Also, the watershed's close proximity to major transportation arteries and the availability of remote sensing and existing land use data were primary considerations.
Land cover for the Sope Creek watershed was classified using high spatial resolution (10 meter) multispectral (15 bands) remote sensing imagery from an airborne sensor, the Advanced Terrestrial Land Applications Sensor (ATLAS). ATLAS is a 15-channel multispectral scanner that possesses the same bandwidths as the space borne Landsat Thematic Mapper (TM) instrument and additional bands in the middle reflective infrared and thermal infrared (TIR) range that is flown on a Lear 23 jet aircraft operated by NASA. The 10 m spatial resolution permitted the discrimination of discrete surface types (e.g., concrete, asphalt) as well as individual structures (e.g., buildings, houses). Moreover, TIR data collected with the ATLAS at high spatial resolution are particularly important for providing fine scale resolution of individual surfaces from which the thermal energy responses characteristic of these surface can be derived. These data have been extremely useful for providing quantitative information on what the thermal attributes are of the urban landscape, as well as what the distribution is of thermal energy responses. Additionally, the employment of high spatial resolution TIR data for analyses of surface thermal responses is both innovative and unique, and offers research and application capabilities that have proven to be extremely useful for developing a better understanding of how urban surface heating affects UHI development (see Quattrochi and Ridd, 1994, Lo et al., 1997, Quattrochi and Ridd, 1998, Quattrochi and Luvall,1999).
A classification of land use is not a direct aggregation of land cover classes. For example, grass, a typical land cover class, can be associated with golf courses, recreational parks, cemeteries, residential lawns, or surround commercial buildings. To convert land cover maps to land use maps requires additional ancillary data and processing. For this purpose, we utilized the LandPro99 dataset, a collection of land use classes compiled for the Atlanta Regional Commission. The classification is is a vector product with 28 classes. These data were used to define training data for guidance in reclassification of the image for land use classes.
|
|
The effects of atmospheric path radiance and transmittance were corrected for to permit accurate analyses of thermal energy responses from ATLAS' TIR data (cf. Anderson and Wilson; 1981; Quattrochi and Goel, 1995; Quattrochi and Luvall, 1999). In order to correct for atmospheric effects, it is necessary to know what the structure of the atmosphere is through measuring various atmospheric characteristics. This information for input into atmospheric correction algorithms is obtained from atmospheric soundings, launched during the time of TIR data acquisition, that records data on humidity (dew point), temperature, barometric pressure, and height above the ground. Hence, these weather balloon data provide a profile of the atmosphere between the ground and the airplane, that are in turn, used as input into computer models that correct for atmospheric effects on TIR data. Sounding data were obtained commensurate with ATLAS data collection.
LOWTRAN6 has been shown to provide robust results in modeling out atmospheric path radiance and transmission, the primary constituents that effect the derivation of accurate measurements from TIR data. The output from LOWTRAN6 is combined with calibrated ATLAS spectral response curves and blackbody information recorded during the overflight, using the Earth Resources Laboratory Applications Software (ELAS) module TRADE (TIMS RADiant Energy) (Graham et al., 1986) to produce a look-up table for pixel temperatures as a function of ATLAS values (Anderson, 1992). A NE?T was then calculated for each channel from image "housekeeping" data. (NE?T or "noise equivalent delta temperature", is a measure of the energy of that input radiation falling on the sensor detector that would give an output signal equal to the noise generated by the detector. The smaller the NE?T value, the better the detector [Anderson and Wilson, 1984]). A NE?T of 0.2 was obtained for ATLAS channel 13 (9.6-10.2 *m, centered at 10 *m) used in the analysis of data for Atlanta.
Data processing and calibration were performed using the ELAS software. Flight lines of ATLAS TIR data for each city were first pre-processed to remove any abnormalities present in the data (e.g., dropped scan lines). Each flight line of ATLAS TIR data was then run through the TRADE algorithm to produce an atmospherically corrected data set. All of the ATLAS TIR multispectral data sets collected for Atlanta were then georeferenced to UTM map coordinates to produce a geometrically corrected data set for each city. These processes ensured the generation of the high quality data sets that have been used to calculate accurate surface temperatures for a variety of surface types, to produce a land cover classification, to derive thermal color and true color map products, and to provide a data set that can be imported into geographic information systems and used for further analysis.
|
|
The Landcover classification was performed with ENVI 4.0 software using the supervised parallelepiped technique. This is a standard technique readily available for use by transportation professionals. Parallelepiped classification uses simple decision rules that define boundaries based on standard deviation thresholds from the mean of each class to classify multispectral data. Bands 4 (visible), 6 (near infrared), and 13 (infrared) were used to develop the land cover classification. Polygons with a minimum of 500 pixels per land cover class were selected training. The total number of pixels in training polygons among all classes ranged from 657 to 4,356 and comprised approximately one percent of all pixels in the Sope Creek watershed.
The Sope Creek watershed spans two flight lines of ATLAS data. Each flight line was classified separately and then mosaicked to produce the land cover classification product (Figure 3).
Table 1: Classification results.
These classes are consistent with the heavily vegetated character of the study area. The majority of the low-density residential areas are comprised of forest and grass classes. Low-density residential rooftops comprise a significant portion of the unclassified pixels too. Urban areas have been classified into low albedo and high albedo areas. The high albedo areas are primarily light building roofs and concrete surfaces. Low albedo areas are represented by asphalt surfaces on roads and parking areas plus dark building roofs. Land areas in transition for urban development comprise most of the bare soil class.
Overall, the results are a good representation of the physical surface of the Sope Creek watershed. Some mixed pixels between the high albedo urban and bare soil class remain a problem. The mature forest canopy made identification of urban features challenging, especially in areas remote from urban concentrations. Also, with more detailed ground truth data a delineation of evergreen forest areas from the predominant deciduous forest could have been evaluated.
|
|
LandPro99 is a land use/land cover dataset developed by the Atlanta Regional Commission (ARC) for the 13-county Atlanta metropolitan region. The LandPro99 dataset was created by photo interpretation of color aerial photography at four-foot pixel resolution and color infrared digital orthophoto quarter quads with one-meter pixel resolution provided by the U.S. Geological Survey (USGS). Using both sources of imagery, landcover polygons were delineated. Land use was added after the landcover was completed and is based primarily on property ownership information available in the Region. The land use/land cover classification system is adapted from the USGS Anderson system and includes a total of 28 classes. These data were used to define training data for guidance in reclassification of the image for land use classes.
Within the Sope Creek study area, the LandPro99 dataset contained seventeen land use classes. These classes were available in vector format and provided an optimal data source for use in developing training classes to produce an automated land use class from the remotely sensed ATLAS data source. The LandPro99 land use classes are defined as follow with abbreviations as noted on the map legend describe in parentheses:
Medium Density Residential (Res_Med): Areas that have generally been developed for single-family residential use, with or without a significant mix of forested or agricultural landcover.
High Density Residential (Res_High): Areas that have predominately been developed of concentrated single-family residential use, usually found in urban neighborhoods.
Multi-family Residential (Res_Multi): Residential areas comprised predominately of apartment, condominium, and townhouse complexes where net density generally exceeds eight units per acre.
Commercial and Services (Commercial): Areas used predominately for the sale of products and services, including urban central business districts, shopping centers in suburban and outlying areas, commercial strip developments, junk yards and resorts.
Industrial/Commercial (Ind/Com): Industrial and commercial areas that typically occur together or in close functional proximity with one another.
Transportation, Communication, and Utilities (TCU): Various land use types associated with transportation, communication, and utilities such as electrical substations, pumping stations, and airwave communications.
Institutional-intensive (Inst_Intensive): The built-up portions of institutional land holdings, including all building, grounds and parking lots that compose educational, religious, health, correctional and military facilities.
Cemeteries (Cemeteries): Public and private lands devoted to burial grounds, including primary and secondary buildings and associated infrastructure.
Golf course (Golf_Courses): The "green space" areas of golf courses, including tees, fairways, greens and intervening land.
Parks (Parks): Active recreation areas identified from aerial photography, including baseball and other sports fields, tennis courts, swimming pools, camp grounds, parking lots, structures, drives, and trails.
Agriculture Crops and Pasture (Ag_Crops): Agricultural land regularly used to grow field crops or to pasture animals.
Forest (Forest): All forested areas of coniferous and/or deciduous trees.
Urban Other (Urban_Other): Open land in intensive or indeterminate urban uses that do not require or do not have structures.
Limited Access (Ltd_Access): All highways or portions of highways that are considered "limited access", their right-of-ways, ramps and interchanges.
Transitional (Transitional): Recently cleared or altered land in transition from one land use activity, either built-up or non-built-up, to another unknown or undeterminable land use.
Reservoirs, Lakes, and Ponds (Reservoirs): Man-made impoundments often referred to as "lakes or "ponds", which are persistently covered with water.
Bare Exposed Rock (Exposed_Rock): Naturally occurring areas of exposed bedrock with little or no vegetative cover.
|
|
The major objective of this task is to evaluate the utility of remotely sensed multi-spectral data for urban land use classification in major transportation corridors. ATLAS multi-spectral 10-meter data was used to perform the Sope Creek land use classification with ENVI 4.0 software. Polygons matching landpro99 land use classes were selected as training classes. Standard supervised classification algorithms including parallelepiped, maximum likelihood, and minimum distance were evaluated for use in this study. All these algorithms are standard techniques readily available for use by transportation professionals. Parallelepiped classification uses the threshold of each class signature to determine if a pixel is within a class or not (J.A. Richards, 1994, Remote Sensing Digital Image Analysis, Springer-Verlag, Berlin, p. 340). Advantages are speed and potential problems are low accuracy if a large number of pixels cannot be classified. The maximum likelihood classifier uses the Gaussian threshold in each class signature to determine the pixel's class. The maximum likelihood technique is slower compared to parallelepiped technique due to extra computations required, however, results may be more accurate if classes in the input data have a Gaussian distribution. Finally, the minimum distance algorithm compares distances between the pixel to be classified and each class center and the pixel is assigned to the nearest class center. For this study, the parallelepiped classifier was used. The advantage of a Gaussian distribution in classes was not evident in preliminary classifications thereby limiting the advantages offered by the maximum likelihood classifier and the wide variation in class sizes was a factor in not selecting the minimum distance algorithm. The parallelepiped technique resulted in a low percentage of unclassified pixels, alleviating a common concern when utilizing this classifier. Because each study area has a unique combination of physical features and surface types, testing is necessary to identify the spectral bands that carry the majority of the information given the spectral and spatial resolution of the available remote sensing data. For the Sope Creek study area with ATLAS data, we determined that bands 3 (visible), 6 (near infrared), and 13 (infrared) were sufficient to develop the final landcover classification. Land use compared to land cover classification presents some unique problems. The land use classes represent a mixture of land covers with varying spectral signatures, which makes classification of land uses more difficult. For example, the Low-density Residential land use class is comprised of a mixture of surface types (i.e. land covers) including grass, forest, concrete, and asphalt materials. Commercial and industrial land uses typically contain concrete, asphalt, and grass interspersed with small forest stands. These "mixed pixel" classes limit the ability of standard classifiers to accurately characterize some land use types. As a result, more specific land use classes must be aggregated into a more generalized class. For example, Multi-family, High-density Residential, and Institutional classes may have similar spectral features as Commercial and Industrial classes and be considered part of such classes. Also, grass and selected recreational areas such as parks and golf courses may be considered part of the Low-density Residential land use class. Another limitation that stems from using standard classifiers to produce land use classes from multi-spectral data is the small spatial extent of some of the land use classes in the Sope Creek region. This impacted classes with less than 5,000 total pixels, such as Parks/Recreation, Limited Access, Cemeteries, and Transitional. The heavily shaded character of the study areas also presented classification challenges in the interpretation of spectral signals in these areas (ATLAS data were processed for geometric, atmospheric, and shadow corrections prior to use in this study). Land use classification produced four classes as shown in Figure 4 and Table 2. Although the LandPro99 data set identified 17 land use classes within the Sope Creek watershed, due to mixed pixel challenges and the small extent of some classes as noted above, only a rather generalized classification with three principal classes and a small percentage of unclassified pixels could be achieved. A number of land use classes simply were not spectrally unique. Ultimately, LandPro99 Forest and Cemeteries and Transitional classes were combined as the Forest class. Commercial and Services, Industrial/Commercial, Reservoirs, Lakes, and Ponds, Transportation, Communication, and Utilities (TCU) were aggregated as Commercial/Industrial. Medium Density Residential was significantly unique and in large enough spatial extent to be isolated as Residential.
Table 2: Results from Land Use Classification
The Forest class is a significant land use in the region. The Cemeteries and Transitional classes were too heavily forested to be spectrally unique resulting in being grouped with the Forest class. This, of course, is unique to this area with older, more established cemeteries. The urban core of industrial and commercial development is consistently captured in the Commercial/Industrial class. The one area identified as TCU in the area is primarily urban with very little vegetation in the class, so it also is accurately classified. Due to the low albedo of lakes and ponds, these small manmade water bodies also provide spectral signatures that result in being combined in the Commercial/Industrial class. In many instances, this is an inaccurate association as most of these water bodies are found in residential areas in the Sope Creek region. The majority of the Sope Creek region is in single-family or low-density residential development that corresponds well to the Medium Density Residential land use class that encompasses the majority of the study area. There is some inclusion of forest areas in the residential class where dense tree growth occurs. Bare Exposed Rock, Limited Access, Institutional-intensive, Agriculture Crops and Pasture, Golf Course, High Density Residential, Urban Other, Parks, and Multi-family Residential land use classes, were not consistently represented in the major land use classes delineated above. Most of these classes, with the exception of Institutional-intensive and Multi-family Residential, are very small in extent and often in one location. The Institutional-intensive and Multi-family Residential classes are typically confused with the Commercial/Industrial class due to the high percentage of common surface types. However, areas in these classes with significant vegetation tend to classify as Residential resulting in a mixture of pixels in these land use classes. Possibly with more evaluation and the use of advance textural classifiers, a Multi-family Residential class could be delineated. Limited Access classes tended to be classified as commercial in highly developed areas with asphalt surfaces and as residential in less developed areas especially with a high percentage of concrete and vegetative surfaces. Only one location each of Bare Exposed Rock, High-density Residential, Agriculture Crops and Pasture, and Golf Course classes were found in the study area, making evaluation of these areas impractical. With the opportunity to develop training classes, it is likely the High-density Residential class would be grouped with the Commercial/Industrial class described above. Development of unique classes for golf courses and some agricultural areas may also be possible.
Overall, the result is a good generalized representation of the major land use classes in the Sope Creek watershed. Further evaluation and the use of more advanced textural classifiers would likely enable two to three additional land use classes to be delineated.
|
|
Anderson, J.E., 1992. Determination of water surface temperature based on the use of thermal infrared multispectral scanner data. GeoCarto International, 7(3):3-8. Anderson, J.M. and Wilson, S.B., 1982. An evaluation of the effects of atmospheric absorption on the data required during an infrared survey. Proceedings of the Conference on Remote Sensing and the Atmosphere, Liverpool, U.K., pp. 92-99. Anderson, J.M. and Wilson, S.B., 1984. The physical basis of current infrared remote-sensing techniques and the interpretation of data from aerial surveys. International Journal of Remote Sensing, 5(1):1-18. Lo, C.P., Quattrochi, D.A. and Luvall, J.C., 1997. Application of high-resolution thermal infrared remote sensing and GIS to assess the urban heat island effect. International Journal of Remote Sensing, 18:287-304. Quattrochi, D.A. and Ridd, M.K., 1994. Measurement and analysis of thermal energy responses from discrete urban surfaces using remote sensing data. International Journal of Remote Sensing, 15:1991-2022. Quattrochi, D.A. and Goel, N.S., 1995. Spatial and temporal scaling of thermal infrared remote sensing data. Remote Sensing Reviews, 12:255-286. Quattrochi, D.A. and Ridd, M.K., 1998. Analysis of vegetation within a semi-arid urban environment using high spatial resolution airborne thermal infrared remote sensing data. Atmospheric Environment, 32(1):19-33. Quattrochi, D.A. and Luvall, J.C., 1999. Thermal infrared remote sensing data for analysis of landscape ecological processes: Methods and applications. Landscape Ecology, 14(6):577-598. Quattrochi, D., Luvall, J., Rickman, D., Estes, M., Laymon, C., Howell, B., 2000. A decision support system for urban landscape management using thermal infrared data, Photogrammetric Engineering and Remote Sensing, 66(10), 1195-1207. Richards, J.A., 1994, Remote Sensing Digital Image Analysis, Springer-Verlag, Berlin, 340 p.
|