An evaluation of seasat SAR imagery for urban analysis

REMOTE SENSING OF ENVIRONMENT 12:439-461 (1982)

439

An Evaluation of Seasat SAR Imagery for Urban Analysis*

FLOYD M. HENDERSON Department of Geography, State University of New York at Albany, Albany, New York 12222

Digitally processed Seasat SAR imagery of the Denver, Colorado, area is analyzed with regard to the types of urban data that can be detected and/or inferred from satellite-borne L-band systems. Black-and-white images of the scene were generated at three scales to determine the advantages and detail discernible at each level of display. The large-scale imagery was density-sliced to evaluate the feasibility of producing a semiautomated land-cover classification from the SAR data. Gray level classes were assigned colors to aid interpretation and subsequently compared with the black-and-white images to assess the contribution of each technique and benefits of combining the data from both procedures.

Introduction

Many new methods are being explored and developed to improve the urban data base. In planning for and monitoring the constantly changing urban milieu information is needed more rapidly and at a higher level of consistency, quality, and detail. Of the techniques being examined with this objective in mind, remote sensing systems are receiving considerable attention. Although much of the remote sensing effort to date has focused on visible and near infrared sensor systems operating from aircraft and spacecraft platforms the potential of radar imagery also merits attention. In certain instances radar may prove to be the only sensor capable of providing data. Equally if not more important is an assessment of the potential of radar imagery as a complement to other sensor systems. As such, the question arises as to

*Portions of this study were presented at the 14th Congress of the International Society of Photogrammetry, Hamburg, Germany, 13-25, July 1980. ©Elsevier Science Publishing Co., Inc., 1982 52 Vanderbilt Ave., New York, NY 10017

what distinct contribution to urban data collection can be made by radar as a function of its sensitivity to texture, surface roughness, spatial orientation, contrasts, and other system/environment-related parameters. In short, what can radar offer? The purpose of this study is to explore the capability of digitally processed Lband Seasat SAR imagery for mapping urban land cover and the compatibility of the SAR derived land-cover classes with those described in the United States Geological Survey land-use/land-cover classification system designed by Anderson et al. (1976). Study Area The 6 August 1978, 100 × 100 km digitally processed SAR scene (revolution 580, ascending pass) of the Denver, Colorado, metropolitan area served as the study area. Six subareas of the Denver SAR scene representative of Level II urban land-cover categories extant in the area were selected for detailed analysis of land cover types (Fig. 1). They were 0034-4257/82/439-23502.75

440

F . M . HENDERSON . . . . . . . . . . . . . . . . . . . .

7"

\

FL

¢

Z,

i

-%

~cAt~

T

1

2

\

3

K t L O M ~ [E~5 .

FIGURE 1.

.

.

.

j

. . . . . .

Location of large scale, subscene study areas,

selected by examining aerial photography and locating areas which contained examples of the following urban land cover: older residential area in the city interior; new residential areas on the urban fringe; single- and multiple-family housing; industrial; commercial and service; recreation and open space; and transportation.

Methodology Black-and-white imagery of the study area were generated from the digital tapes at three different scales using a General Electric Image 100 interactive processing

system and applying a piecewise linear contrast stretch to the appropriate subset: (1) 1:500,000, (2) 1:131,000, and (3) 1:41,000. The first scale depicted the flail SAR scene and the second scale provided an intermediate scale and area of coverage comparable to that of a high altitude aerial photograph. Black-and-white images at both scales were manually/visually interpreted. The third scale was the maximum enlargement possible (512×512 pixels) on the Image 100 system without data resampling. Two distinct approaches were employed in examining these d a t a - - a visual interpretation of the black-and-white

SAR IMAGERY FOR URBAN ANALYSIS

FIGURE 2.

441

Raw data Seasat SAR scene of Denver, Colorado. North is to the right.

imagery and an automated machine/visual interpretation of study areas. The purpose was to evaluate the feasibility of producing semiautomated land-cover classifications with SAR data and to assess the contribution of each approach (i.e., visual or optical versus density-sliced imagery). The land-cover classification system described by Anderson et al. in U.S. Geological Professional Paper 964 (1976) was adopted to provide a basis for systematic comparsison of the data. To determine accuracy all SAR interpretation results were compared with aerial photography (October, 1978) and existing USGS land cover maps (1971, 1976) of the area. Visual interpretation A black-and-white positive transparency of the entire 100 × 100 km scene was generated from the digital data at a

scale of approximately 1:500,000 (Fig. 2). An examination of the product indicated no meaningful land-cover patterns could be discerned from this essentially raw data due to excessive image noise inherent in the data. Consequently, a second image was produced by computing the average for all nonoverlapping three by three windows in the original image. That is, the image was resampled with one third of the original number of pixels in each direction, each new pixel being the average of nine pixels. The three by three averaging appeared to reduce the noise and land cover patterns were more apparent (Fig. 3). Level I land cover categories (i.e., urban, agriculture, forest, rangeland, and water) were located and delimited on this transparency. This synoptic view of the entire metropolitan area permitted an assessment of the detectability of general land cover; particularly the

442

F.M. HENDERSON

FIGURE 3.

3 × 3 window-averaged Seasat SAR scene. North is to the

right. visibility of the rural-urban fringe, satellite communities, and the distinction between urban and nonurban land cover. The intermediate scale image (1: 131,000) was generated from the data tapes averaged in the same manner as the small scale (1:500,000) scene. This 367 km 2 area (Fig. 4) was interpreted at Level II land cover detail and a confusion matrix generated. Black-and-white images of each of the six large-scale subscenes (1:41,000) were generated and visually interpreted using Level II land-cover categories. The averaging was not applied to the analyses of the six large-scale subscenes as the image noise presented no problem at this scale (Fig. 5). That is, at this scale the physical size of the pixels on the transparencies was large enough that the eye could discern patterns appearing to correspond to urban land-cover patterns as opposed to perceiving only a display of random gray

tones and dots. The resulting SAR landcover maps were analyzed and confusion matrices generated to determine the accuracy of the SAR interpretation, instances, and sources of error, and the level of detail and type of classification possible by this approach. Effects of system/environment variables such as specular reflection, incidence angle and system wavelength, street orientation, and land-cover type on image classification were also examined.

Machine/visual interpretation The objective of the digital analysis was to determine the feasibility of discriminating land-cover classes using density level slicing of the image frequency histogram. Color images of each of the five density-sliced large scale subscene study areas were produced, and land-cover categories delimited at Level II detail.

SAR IMAGERY FOR URBAN ANALYSIS

FIGURE 5.

Large scale black-and-white subscene image of Study Area 1.

443

444

F.M. HENDERSON

Confusion matrices were generated and compared with the results obtained from the optical interpretations of the corresponding black-and-white images of each study area. Analysis and Results 1:500,000 (small-scale imagery) At this scale the boundaries of urban brat-up areas could be easily delimited due to the high return of suburban housing in contrast to the darker tones of agriculture, rangeland, and other open space. Agricultural land was identifiable only when several rectangular, cultivated fields were juxtaposed. Rangeland, comprised of bare ground with sparse, short grasses and some scrub vegetation, generated a smooth, low return response (dark gray to black). Given the L-band wavelength and 25 M system resolution this was not unexpected. The predominant surface features would be below the Rayleigh scatter threshold for rough surfaces (5.72 cm) and the few potential rough surface scatterers would tend to be lost in smooth response (low return) of the dominant low cover types within the resolution cell. As a result of the combined effects of surface roughness parameters and the land cover pattern, composition, and morphology, the rangeland, pasture, and bare fields produced similar tone-texture responses at this scale. Forest land was discernible in the mountains, along stream banks, and on lowland hills as tex~red, medium to light gray toned areas. Water was not consistently detectable at this scale. Small water bodies (less than 2 km2) could not be delimited consistently from surrounding land

cover of grass, beach, bare soil, and rangeland. Larger reservoirs often produced a salt-and-pepper response rather than the expected black, no return response owing to the L-band signal sensitivity to rough and choppy water. These observations are not meant to imply that Seasat SAR imagery cannot be of use in small-scale synoptic land-cover inventory efforts. It is believed that this particular environment is a limiting factor and that more detail may be visible in other regions. The cultivation and land-use patterns of the Denver area reflect the semiarid climatic conditions. At this scale open space, bare ground, sparse range vegetation, sandy soils, and areally extensive agricultural field patterns do not offer sharp tonal contrasts at L-band wavelengths (Inkster et al., 1980; Morain and Campbell, 1974; Waite et al., 1980). However, Level I land cover can be mapped (Fig. 6). While the distinction of nonurban land-cover types is confined to broad generalizations the contrast between large concentrations of urban builtup land and nonurban areas is quite distinct. In fact, the generalization afforded by such a synoptic view is a prime objective of such small scale imagery. In more humid environments a n d / o r areas that are intensively cultivated more contrast and surface roughness (manifested as texture and gray tones on the imagery) would be expected as a function of crop diversity, field patterns, and spatial distribution of land cover (Waite et al., 1980). In such cases much more detailed land-cover information should be obtainable from the Seasat SAR imagery. Also, the development of more sophisticated preprocessing algorithms should increase the level of

SAR IMAGERY FOR URBAN ANALYSIS

445

SCALE N /

0 i 0

FIGURE 6.

10 I

I

20 I

10

30 I I 20

KILOMETERS MILES

LOOK

DIRECTION

l e v e l I land cover m a p of small scale (1:500,000) SAR scene.

information obtainable from the digital data and enhance subtle land cover contrasts that at present remain indistinct in all environments. 1:131,000 (medium-scale imagery) Recently constructed residential subdivisions were easily identified by optical

interpretation of the black-and-white image as were most older, interior residential areas (Fig. 7). The commercial/ industrial core of the city and the concentration of downtown commercial activity were also apparent. However, small commercial blocks in residential areas and the distinctions between the commercial

446

F.M. HENDERSON

R/F~

O/O

R/O l~

R/O~

P/P

O/O

::::'"'"'"'" "m .......

ChC ,I

R/O

R/F

-

O!O

R/P

~

R/R

~0/0 R/F

F/F

R/I ~

RC IS

R/F

R/R O/C R/(

SCALE SAR

CLASS

i

2I

I CIR

[ ]

CLASS

O

2

CORRECT RADAR CLASSIFICATION

~I

I

I

4

6

MILES

LOOK DIRECTION ~_

KILOMETERS

INCORRECT RADAR CLASSIFICATION

FMH:maa

FIGURE 7. Comparison of 1:131,000 SAR classification with actual land cover. Key: (CI/I) commercial-industrial; (CS) commercial-services; (FOP) recreational-open-public; (R) residential; (W) water.

core and the interior residential areas were ambiguous. Confusion between commercial/service land-cover and residential land cover was the cause of most of the error at this scale. Portions of major arterial roads could be inferred from the dark linears travers-

ing the urban area but complete road networks were not visible. Open space was distinguishable owing to the sharp contrast between its dark, low return and the surrounding medium to light gray tones of built-up land cover. However, the exact use of the open space could not be

447

SAR IMAGERY FOR URBAN ANALYSIS TABLE 1 Confusion Matrix of SAR Interpretation Accuracy at 1:131,000 Scale a SAR Im'~P~TED LANDC o v ~ ACTUALGnOUNDCovEn CI/I CS FOP R

CI/I

CS

FOP

4,435.5 ha (10,960 a)

121.8 ha (301 a)

410.4 ha (1,014 a) 66.4 ha 188.2 ha (164 a) (465 a) 898.0 ha 1,197.5ha (2,219 a) (2,959 a)

5,500.2 ha (13,591a) 121.8 ha (301 a)

R

5,399.9 ha (13,343 a)

1,796.1ha (4,438 a)

5,743.8 ha (14,193 a)

TOTALACTUAL A~A

388.1 h (959 a) 388.1 ha (959 a)

4,712.7 ha (11,645 a) 1,297.5 ha (3,206 a) 6,686.0 ha (16,521 a) 24,638.6 ha (60,882 a) 388.1 ha (959 a) 37,722.9 ha (93,213 a)

155.4 ha (384 a) 887.1 ha (2,192 a) 931.2 ha (2,301 a) 22,421.3 ha (55,403 a)

W Total Area by SAR

W

24,395.0 ha (60,280 a)

aKey: (CI/I) commercial-indnstrial; (CS) commercial-services; (FOP) recreation-open-public; (R) residential; (W) water ha = hectares; a ffi acres.

consistently classified as recreation, cemeteries, or other open-space categories. Instances where institutions, public facilities, or commercial buildings were surrounded by open space were also classified as simply open space from the radar image. Table 1 provides a summary of omission and commission errors for this scale. 1 Given the complexity of land cover types and the relatively low total percentage of incorrect identifications (12.1%) the potential of such SAR imagery for urban data extraction appears promising. Familiarity with the urban area land cover locations and types might enable more precise identification but cannot be documented at present.

1:41,000 (large-scaleimagery) Six subareas of the entire SAR scene that encompassed a range of urban landcover types including: older, interior and 1Percent omission ffi 100%-% correct; percent commission ffi total number of commission errors × 100 total possible responses-total possible correct responses

new residential areas; single- and multiple-family housing; industrial; commercial and service activity; recreation and open space; and transportation were selected for analysis. Each study area encompassed some 64 krn~. As can be seen in Fig. 1, three of the study areas (3, 4, and 5) traverse the metropolitan area in addition to three areas located on the urban fringe (1, 2, and 6). Study area 2 was subsequently omitted as its land cover pattern as redundant to that of Study areas 1 and 6. On the black-and-white images generated from the raw data, older, interior residential areas were less distinct than the new residential areas and confused at times with commercial and service activity and some industrial fringe areas owing to similar tone and texture responses on the imagery. Separate categories of recreation, cemeteries, and open space could not be consistently defined other than as open space. Institutions, schools, and public land were also confused with open space due to the low return of their

448

grounds. The visibility of transportation elements was a flmction of their size, shape, orientation to flightline, and surrounding land cover. Only segments of major road networks were visible with sections of residential streets identifiable as dark lines or dashes in contrast to the higher return of surrounding housing structures. In commercial zones streets were generally obscured by the bright return and signal blooming (specular response) from buildings. The visibility of commercial and service activity was a function of size and location. The contrast between the central business district, industrial core, and surrounding residential land was detectable, but small commercial centers and commercial streets in residential areas remained a major point of confusion. Using the raw data at this large scale the distinction between the business and commercial activity versus other land cover, particularly residential, was more pronotmced than on the smaller 1:131,000 scale, smoothed data. However, isolated commercial/industrial buildings in open areas were still indistinct from residential development unless identification could be inferred from its spatial location and tmusually bright signal response--a function of building size, complexity, and orientation to the flightline. Given these general observations the next step was to examine the land cover maps created from the raw data and the density-sliced data for each of the five study areas. Confusion matrices were generated in each case using aerial photography and existing maps of the study areas as ground truth. Cells representing 1.6 hectares (4 acres) on the ground were selected as the recording unit of measurement as (1) it was the smallest cell from

F. M. HENDERSON

which land cover could be consistently and accurately recorded, (2) the area encompassed by each cell is the smallest areal unit for which the majority of urban land cover is recorded by urban planners and other users, and (3) it is a unit recommended by Anderson et al. (1976) for remote sensing systems. A systematic aligned sampling system (center dot) was employed to identify the land cover in each cell (Berry and Baker, 1968). It should be noted that there is some disagreement in the confusion matrix tables as to the number of hectares among the five study areas. This is a result of limitations of the image generation and reproduction process and does not detrimentally affect the quality or validity of the data. Preliminary examination had indicated that it was not possible to distinguish precisely specific types of land cover from each other. In compiling the confusion matrices the decision was made to combine some open space categories into one inclusive category as indicated in the tables. A second, modified group category was employed for commercial service and commercial/industrial land cover. For purposes of brevity the interpretation and resulting land-cover maps generated from the black-and-white prints of the raw data will be referred to as the optical interpretation and the results obtained from analysis of the density-sliced, color-coded data will be referred to as the density interpretation. Study area 1 is located on the urban fringe of the metropolitan area and consists chiefly of new single-family residential housing, some small lakes and reservoirs, public land, and open space in the form of recreation (golf courses) and agricultural land awaiting urban develop-

SAR IMAGERY FOR URBAN ANALYSIS

449

SCALE

/ [ ]

SAR CLASS.

0

ACTUALCL~S.

0

CORRECTRADAR CLASSIFICATION

I

2 I

m

NEW RESIDENTIALAREAS NOTED WITH RADAR

KILOMETERS MILES -~-l N ~

LOOK DIRECTION INCORRECTRADAR CLASSIFICATION

FIGURE 8. Comparison of large scale (1:41,000) optical SAR classification with actual land cover for study area 1. Key: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; (17) recreation.

ment. The optical interpretation of this study area produced an overall accuracy of 93.9%. It was not possible to detect land-cover devoted to extractive, public, or utilities activites, but these cover types comprised only 3% of the total land cover. Ninety-five percent of the residential land cover was correctly identified. The density interpretation produced a lower overall accuracy of 89.5% including a 12%

error in the residential category. Figures 8 and 9 provide an indication of the accuracies, sources of error, and ability of the SAR imagery to accurately detect new residential construction areas. Study area 3 is an older area of urban development, interior to the current urban fringe and includes a mix of more varied land cover activities. Approximately 14% of the area is industrial/commercial activ-

450

F.M. HENDERSON

SCALE

'~

SAR CLASS

0

ACTUAL CLASS

0

CORRECT RADAR CLASSIFICATION

2

I

1

mm

NEW RESIDENTIAL AREAS NOTED WITH RADAR

R(LOMEIERS MILES ~

LOOK DIRECTION

INCORRECT RADAR CLASSIFICATION FMH:maa

FIGURE 9. Comparison of large scale (1:41,000) density SAR classification with actual land cover for study area 1. Key: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; (17) recreation.

ity in blocks and strips while some 70% is residential--a mix of older, single family homes and apartments (Figs. 10 and ll). The overall optical interpretation accuracy was 84% but dropped to 76.9% for the density interpretation. Although residential identification remained high in both cases there was considerable confu-

sion with other categories, particularly commercial land cover. The central business district, downtown commercial/industrial core, transportation hub, interior multifamily and single family housing, and city parks comprise study area 4. The complex nature of this land-cover mix presented diverse

SAR IMAGERY FOR URBAN ANALYSIS

451

CI/R

R/F ~

O/P

~li i R/R R/C

~'./F

R/C RIR

R,/F FIr

~O/FR/R

O/P

R/T*~

RIP

t RIP

R/P"~

SkR CLASS.

"

Cl/C

~L'R/o

0

1

I

I

SCALE

2 KILOMETERS I

LOOK

I

FMH:maa FIGURE 10. Comparison of large scale (1:41,000) optical SAR classification with actual land cover for study area 3. Key: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; (17) recreation.

interpretation problems as many of the land-cover categories were not distinct on the SAR imagery. This fact is reflected in the lower accuracies achieved--77.4% for the optical and 69.9% for the density analysis. As was the case in previous study areas the major problem was the inability to distinguish small areal units devoted to transportation, public land cover, and

small parcels of open space devoted to parks, cemeteries, and idle land. A second factor was the similarity in appearance on the imagery between commercial/service and commercial/industrial land cover versus other land-cover types, especially the older row housing and multistory residential buildings (Figs. 12 and 13). In spite of these problems the overall accu-

452

F. M. HENDERSON

SCALE SAR

CLASS

N

0

I ACTUAL

CLASS

0

CORRECT CLASSIFICATION

i

?

1 ]

KILOMETERS

LOOK

I MILES

D~RECTION

iNCORRECT CLASSIFICATION FMH:maa

FIGURE 11. Comparison of large scale (1:41,000) density SAR classification with actual land cover for study area 3. Key: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (VO water; (F) recreation.

racy of the optical method (77.4%) indicates that SAR imagery can produce acceptable results for the major land-cover types. Seventy-three percent of the residential, 83% of the commercial, and 100% of the commercial/industrial land cover were correctly identified. Study area 5 includes Denver's airport and the surrounding area. This study area

presented the most problems in interpretation and produced the greatest error (70.8% accuracy for the optical and 62.5% accuracy for the density interpretation). The major cause of interpretation error was that areas adjacent to the airfield did not present distinct, identifiable tones and textures on the SAR imagery. Comparison of the aerial photography

SAR IMAGERY FOR URBAN ANALYSIS

453

I/Cl CI/k R/R

CI/R

~,~R/P

R/R

~-R/P

R/P~

C/R I/CI C/R

C/P

T/T

O/P

/

R/P.~

5

ClC

R/R RIR

O/P"

F/F

SCALE SAR CLASS

0

1

I

I

ACTUAL CLASS

o

g

D

CORRECT CLASSIFICATION

2

I ~

KILOMETERS

LOOK

I .,tes

O,RECT,ON

INCORRECT CLASSIFICATION FMH:maa

FIGURE 12. Comparison of large scale (1:41,000) optical SAR classification with actual land cover for study area 4. Key: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; (17) recreation.

with the SAR imagery and SAR landcover maps indicated three factors were responsible. First was the inability to define the boundary of the airfield. Much of the grass field and open space around the runway was interpreted as transportation while the United States Geological Survey map classified it as open space or recreation. Second, commercial/industrial ac-

tivity adjacent to the air terminal and transportation land cover were all similar in appearance on the SAR imagery and could not be differentiated. Third, a large agricultural field adjacent to the runway was classified as commercial/industrial on the SAR imagery owing to instances of bright spectral response surrounded by a dark gray area. This gave the impression

454

F. M. HENDERSON

SCALE SAR CLASS

0

i

I

I

ACIUAL CLASS

0

N

--7

CORRECT CLASSIFICATION

2

I ]

KILOMEIERS

LOOK

] MILES

DIRECTION

INCORRECT CLASSIFICATION FMH maa

FIGURE 13. Comparison of large scale (1:41,000) density SAR classification with actual land cover for study area 4. Key: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; (F) recreation.

of a group of buildings and grounds. In fact, the spectral rehu'n was caused by a series of low cultivated ridges oriented parallel to the flightline (perpendicular to look direction) and is a characteristic that has been discussed by Waite et al. (1980). Study area 6 was an urban fringe area similar to study area 1. As might be expected the detection accuracy was quite

high for both interpretation methods, but the optical method (92.1%) was again more precise than the density (88.8%). The chief source of error was confusion of residential with open space land cover. The average accuracy for all five study areas was 83.6% using the black-and-white images/optical interpretation and 77.5% for the level-sliced, color-coded/density

SAR IMAGERY FOR URBAN ANALYSIS TABLE 2 Optical Interpretation Summary of (1:41,000) Study Areas a A1 STtrDY AREA

0

LAND COVER

AFGO C/CS CI/I E P R T U W Totals

C

A3

455

Omission/CommissionError in Acres for Five large-Scale A4

O C

0 C

A6

O C

O C

TOTALS

1,348 1,696 192 808 572 1,528 76 0 1,512 60 1,664 1,168 132 352 88 0 40 12 5,624 5,624 34,284

8 424 NA NA NA NA 28 0 180 0 192 0 NA NA 24 0 0 8 432 432

120 200 NA NA 440 260 NA NA 176 0 284 616 32 0 24 0 NA NA 1,076 1,076

0 192 808 0 356 NA NA 296 60 792 412 0 0 NA NA NA NA 1,636 1,636

780 NA NA 120 700 48 0 832 0 16 132 100 352 40 0 0 0 1,964 1,964

56 292 NA NA 12 212 NA NA 28 0 380 8 NA NA NA NA 40 4 516 516

7,052

6,724

7,224

6,724

6,560

Site Total Acres

356

A5

808

aKey: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (17) recreation; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; NA = not applicable; o = error of omission; c = error of commission.

interpretation.

A

summary

of errors

of

omission and commission in terms of areal units for each category and each study area and the combined totals is provided in the confusion matrices of Tables 2 and 3. Tables 4 and 5 provide the same in/ormarion in percent figures. A summary of the accuracies attained for both methods in each of the six study areas is provided in Table 6. In each instance, the optical interpretation is superior to the density interpretation. Using a binomial test the probability of this being a chance occurence was found to be less than .05, providing strong if not definitive support for the optical method. In summary, newly constructed singlefamily residential areas and housing de-

velopments on the urban fringe were readily visible on the large-scale imagery as were most older residential areas in the interior of the city, although to a lesser extent. The distinct appearance of new fringe residential areas on the SAR image should prove advantageous in monitoring the extent, direction, and pattern of urban expansion. Other Level II land-cover categories could not be separated precisely according to the United States Geological Survey classification system. It is apparent from the omission figures that extractive, public, utilities, and transportation land cover could not be accurately identified from the SAR imagery. However, these parcels are usually very small in size (less than 3 hectares) and comprise

456

F. M. HENDERSON TABLE 3 Density Interpretation Summary of (1:41,000) Study Areas a

STUDY AREA LAND COVER

AFGO

A1

A3

o

o C

28 36

P

152

R

404 NA

1.7

16

W

16

0 0 720 720 Site Total Acres

6,888

TOTALS

6,724

7,056

6,724

6,724

0 228

NA

o ~'

104 260 1,044 32 0 NA NA NA NA 1,552 1,552

56

0

T

o ('

NA 2(1 o 708 24 NA NA NA NA 20 ~} 756 756

144

52

68

A6

40 1,028 NA NA 364 0 1,020 248 272 24 NA NA NA NA 2,124 2,124

108

NA 652

0

A5

836 1,080 NA NA 292 1,244 92 0 932 0 148 80 168 116 52 0 NA NA 2,520 2,520

0 60

284 348

NA

8

E

Totals

o C

238

0 C/CI

A4

C

652 C

Omission~Commission Error in Acres for Five Large-Scale

716

0 732 NA NA S t~ NA

1,476 2,920 180 716 1,Of){) 2,328 204 0 1,696 104

2,540 1,464 472 140 68 0 36 0 7,672 7,672 34,116

aKey: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (F) recreation; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (uI utilities; (W) water; NA = not applicable; o = error of omission; c = error of commission.

a small portion of the total urban land cover. By modifying and combining certain related categories a fairly realistic breakdown of urban land cover was possible. A factor throughout the analysis having significant bearing on the problem of category confusion was the specular return of structures. The orientation of streets and walls of structures relative to the radar look direction is a critical factor in the appearance of low commercial and residential land cover (Bryan, 1979; Bryan, 1981; Waite et al., 1978). Using airborne radar systems Bryan (1979) found that when the angle between the radar look direction and street or structure orientation was less than 10°-15 ° a bright return would result. At other an-

gles a darker return might indicate the same land cover type. Examination of the Denver Seasat SAR data supported these earlier conclusions. This condition was observed when as few as three adjacent houses met the criterion. This phenomenon is certainly a problem to be confronted when automated and semiautomated data analyses of such data are conducted. The density-sliced images of the five large-scale study areas were judged of minimal value compared to the optically interpreted black-and-white images of the raw data--particularly in light of time and costs. Considerable information was lost by the assignment of spectral class ranges and colors to the data (Compare Figs. 5 and 14). Although urban related

457

SAR IMAGERY FOR URBAN ANALYSIS TABLE 4 Optical Interpretation Summary of Large-Scale (1:41,000) Study Areas a STtrDy A/~A

A1

A3

A4

A5

A6

0

0

0

0

0

LAND COVEB

AFGO C/CS CI~ E p R T U W

Omission~CommissionError in Percent for Five

C

0.3 9.6 NA NA NA NA 1~.0 0.0 1~.0 0.0 5.2 0.0 NA NA 1~.0 0.0 0.0 0.1

C

14.4 3.4 NA NA ~.8 4.5 NA NA 88.0 0.0 6.0 ~.4 1~.0 0.0 1~.0 0.0 NA NA

C

49.7 0.0 16.8 13.3 0.0 6.6 NA NA 1~.0 0.9 26.9 9.6 0.0 0.0 NA NA NA NA

C

57.4 14.7 NA NA 8.5 13.2 1~.0 0.0 1~.0 0.0 1.3 2.4 5.8 7.0 1~.0 0.0 0.0 0.0

C

1.8 8.4 NA NA 1~.0 3.2 NA NA 1~.0 0.0 13.0 0.2 NA NA NA NA 7.6 0.1

aKey: (A) agricultural; (C) commercial; (CI) commercial-industrial; (CS) commercial-services; (E) extractive; (17) recreation; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; NA = not applicable; o = error of omission; c = error of commission.

categories (e.g., high return/urban commercial and residential) could be extracted, the density range for each class varied among the study areas. Tonal and texture clues of benefit in interpretation of the image were reduced and the overlap and confusion among land-cover categories increased. However, it is believed that reducing image noise by the incorporation of smoothing and averaging algorithms into the data before density-slicing (e.g., median filtering) may increase the amount and quality of information obtainable (Fasler, 1980). Discussion

Seasat SAR data digitally processed at three different scales were examined in this study. To obtain useful imagery for

analysis, of the entire 100 × 100 km scene (scale--l:500,000) and a medium scale (1:131,000) image it was necessary to employ an averaging algorithm to reduce noise inherent in the data tapes. However, the raw data products were very satisfactory for interpretation of the large scale (1:41,000) imagery. Until smoothing and averaging algorithms can be developed for incorporation into the data prior to level-slicing and color-coding it is believed density-slicing will prove of little value for urban land-cover analysis. Much valuable image texture and tone information is lost when slicing and coding techniques are employed. No improvement in accuracy or level of detail observable was apparent when such data were compared with resuits of the Optical interpretation of the

F. M. HENDERSON

458

TABLE 5 Density Interpretation Summary of Omission/Comission Error itl Percent for Five Large-Scale (1:41,000) Study Areasa STUDYAREA LAND COVER

AFGO C C/CI E P

R T U W

A1

A3

A4

A5

0

0

O

0

C

C

1.0 15.9 100.0 0.0 100.0 0.0 100.0 0.0 100.0 0.0 12.0 1.9 NA NA 100.0 0.0 3.5 0.0

41.6 5.9 NA NA 89.6 0.9 100.0 0.0 100.0 1.6 5.3 57.2 100.0 0.0 NA NA NA NA

c

A6 O

C'

45.5

C

63.7 20.00 NA NA 22.8 22.9 100.0 0,0 100.0 o,o 9.7 1.5 11.0 2.2 100.0 0,0 NA NA

1.7 15.4 11.7 2.1 20.00 NA NA 100.0 o.0 36.7 5.8 60.2 0,4 NA NA NA NA

0.O 19.6 NA NA I(X).0 0.0 NA NA 100.0 0.0 22.2

0.7 NA NA NA NA 3.8 0.0

aKey: (A) agricultural; (C) commercial; (CI) commercial-industrial;(CS) commercial-services;(E) extractive; (17) recreation; (G) cemeteries; (I) industrial; (O) open; (P) public; (R) residential; (T) transportation; (U) utilities; (W) water; NA = not applicable; o = error of omission; c = error of commission.

TABLE 6 Summaryof Large-Scale Study Area Land Cover Interpretation Accuracy STUDY

STUDY

STUDY

STUDY

STUDY

INTERPRETATION

AREA

AREA

AREA

AREA

AREA

METHOD

1

3

4

5

6

AVERAGE

93.9% 89.5%

84.0% 76.9%

77.4% 69.9%

70.8 62.5%

92.1% 88.8%

83.6% 77.5%

OPTICAL

INTERPRETATION DENSITY-SLICED

black-and-white images generated from the data. The merits of each of three scales examined for urban land-cover analysis can be summarized as follows:

b e i d e n t i f i e d for i n c o r p o r a t i o n i n t o g e n eral p l a n n i n g i n v e n t o r i e s of g r o w t h direct i o n a n d l a n d c o v e r c h a n g e . T h e e x t e n t of urban built-up can be defined within a c c e p t a b l e m a p p i n g a c c u r a c i e s of this scale, b u t little d e t a i l is a p p a r e n t .

Small scale (1:500,000) L e v d I l a n d c o v e r classes c a n b e del i m i t e d for s y n o p t i c m a p p i n g of u r b a n areas. A g r i c u l t u r a l l a n d , forest l a n d , a n d r a n g e l a n d a d j a c e n t to u r b a n i z e d areas c a n

M e d i u m scale (1:131,1)00) M o r e p r e c i s e d e l i m i t a t i o n of t h e u r b a n i n f r a s t r u c t u r e is p o s s i b l e at this scale as Level II land-cover category detail can be

SAR IMAGERY FOR URBAN ANALYSIS

459

FIGURE 14. Black-and-white print of density-sliced color subscene, study area 1.

extracted. New residential areas can be delimited as can the commercialservices/industrial core, but small, isolated commercial streets or blocks and the discrimination of older, interior residential areas from adjacent commercial zones (e.g., transition zone) are tenuous. At this scale open space is detectable and its use (e.g., recreation or parks) at times inferred from its size, shape, and spatial location in relation to the urban a r e a - - b u t not consistently. Elements of the transportation network are not always identifiable other than the large airport complex. Large scale (1:41,000) At this scale the most ment of urban growth made. The location and on the urban fringe is contrast between recent

precise measurepatterns can be extent of growth easy due to the residential devel-

opment and surrounding open space. Even small, isolated, low-density housing developments can be detected. The contrast between older interior residential versus commercial activity is generally apparent, but the similarity in gray tone/texture response still remains a problem. However, this indistinction is not unique to radar systems. Buildings in such sections of cities are frequently used for either or both activities and often require ground observation to precisely identify the correct land use. Open space is readily identifiable at this large scale but defining its use is difficult. Classification of open space as to public, institutional, utilities, and extractive land use is imprecise. Transportation elements are visible and more of the transportation network is visible or can be inferred than at any other scale. How-

460

ever, no single type or class of transportation is consistently visible. At this large scale there is a loss of certain spatial association clues inherent at the medium scale, but this problem could possibly be mitigated by mosaicing several large-scale images. In general, some improvement in classification accuracy was possible compared to the medium scale enlargements. At first glance the accuracy of the mediumscale image (87.9%) appears higher than the average for the five large-scale scenes (83.6%) (Table 6). However, the medium-scale image did not include study area 5, the airport scene, where detection accuracy was the lowest. In addition, some small land-cover parcels not visible on the medium scale image were visible on the large-scale scene but were incorrectly identified. Still, the overall merit of the large-scale imagery should be judged in light of costs and information sought. Although the results are not conclusive it is suggested that a better synoptic view and comparable interpretation accuracies can be obtained with the medium-scale image, but for a more exact delimitation of urban growth extent and direction the large-scale images may be preferable. A relatively new urban complex in a semiarid environment was examined in this effort. The question remains as to whether similar results could be expected in a more humid environment, an older urban settlement, one predicated on a different mix of economic activities, or in a smaller or larger metropolitan area. Digitally processed SAR imagery does provide useful information on the urban environment. Medium- and large-scale enlargements provide distinct but complementary urban data. The textural component and the susceptability of radar

F, M. HENDERSON

return to the angular, geometric patterns of man-made structures produce unique signal responses corresponding to urban land-cover types. An effort must be made to precisely understand this relationship and develop more sophisticated preprocessing algorithms. Equally as important, research devoted to the possible synergetic effect of merging the textural component of SAR data with the spectral information available with other sensors (e.g., digitally processed multispectral scanner data) certainly merits serious attention.

Funding for these investigatior~ was provided through NASA / G S F C contract NAS5-26032. The author wishes to thank Stephen W. Wharton for his assistance in imagery generation and application o f algorithms to the raw data tapes. References Anderson, J. R., Hardy, E. E., Roach, J. T., and Witmer, R. E. (1976), A land use and land cover classification system for use with remote sensor data, United States Geological Survey Professional Paper 964. Berry, B. J. L. and Baker, A. M. (1968), Geographical sampling, in Spatial Analysis: A Reader in Statistical Geography, (B. J. L. Berry and D. F. Marble, Eds.) Prentice-Hall, pp. 91-100. Bryan, M. L. (1979), The effect of radar azimuth angle on cultural data, Photogramm. Eng. Remote Sens. 45(8):10971107. Bryan, M. L. (1981), Analysis of two Seasat SAR images of an urban scene, Technical Papers of 47th Annual Meeting of the American Society of Photogrammetry, Washington, D.C., pp. 581-590. Fasler, F. (1980), Texture measurements from Seasat-SAR images for urban land use in-

SAR IMAGERYFOR URBANANALYSIS terpretation. American Society of Photogrammetry Technical Papers, Fall Meeting, Niagara Falls, New York, PS2B1-PS2B10. Inkster, D. R., Lowry, R. T., and Thompson, M. D. (1980), Optimum radar resolution studies for land use and forestry applications. Proceedings of the 14th International Symposium on Remote Sensing o f Environment, Vol. II, Environmental Research Institute of Michigan, Ann Arbor, pp. 865-882. Morain, S. A. and Campbell, J. B. (1974), Radar theory applied to reconnaissance soft surveys. Soil Science Society of America Proceedings, 38:(5) 818-826.

461 Waite, W. P., MacDonald, H. C., Tolman, D. N., Barlow, C. A., and Borengasser, M. (1978), Dual polarized long wavelength radar for discrimination of agricultural land use. Proceedings o f the American Society o f Photogrammetry Fall Technical Meeting, Albuquerque, New Mexico, pp. 595-607. Waite, W. P., MacDonald, H. C., Demarcke, J. S., and Corbell, B. H. (1980), Look direction dependence of radar backscattering cross-section for agricultural fields. Amer/can Society of Photogrammetry Technical Papers, Fall Meeting, Niagara Falls, New York, RS2G1-RS2G14. Received26 April 1982;revised5 luly 1982.