<<

Discriminatingbetween Tropical Vegetation FormationsUsing Reconstructed High- Resolution -AScatterometer Data

Pery J. Hardln and Davld G. Long

Abstract most land imaging applications require spatial resolutions scatterometersare radars designed finer than the 50 or 25 km appropriate for large oce€rnex- to measurc nearsurface wind speed and direction over the panses.However, this paper will presentresults of a discrim- ination experiment designed to determine whether SASS . This was the primary mission for the Seasat-A Scat- imagery reconstructedto a S-km resolution possessesade- terometer (sass), which acquired 14.6-Ghz data from so-krn resolution cells during its three-month mission in 1978. How- quate information for discriminating between several types of ever, an image reconstuction technique utilizing overlap in tropical vegetation in central South America. As shown in Figure 1, the study area covers most of the central and south- resolution cells from successivesatellite orbits can improve that spatial resolution to 5 km over land. An experiment central Brazilian states and tenitories, portions of Paraguay's conducted on reconstructed imagery of central South Amer- Gran Chaco, eastern Bolivia, and the Guiana Highlands of Venezuela, Guyana, and Suriname. While the experiment fo- ica illustrates the potential of this imagery for bopical vege- tation analysis. In a supervised discrimination expefiment, cuses on the existing vegetation cover of this region, the goal of the paper is to encourageothers to consider reconstructed radar backscatter coefficients from the reconstructed imagery imagery as a data source for any medium-scale were found adequotely diverse to distinguish between humid land imaging applications where backscatter at microwave forest, subhumid woodland, climatic savanna, and edaphic frequencies is an important geophysical parameter. savanna with a high degree of confidence, The potential for Although the data used in this experiment originated deriving geophysical information from reconstructed scattero- meter imagery of 's land surface is discussed. with Seasat-A,several additional spaceborne are either currently active or in various planning stages.Like lntroduction Seasat-A,these instruments will acquire backscatter data Satellite scatterometersare active microwave radar instru- over land. The EuropeanSpace Agency (usa) successfully ments originally designed to measure the radar backscatter of launched its ERS-1satellite into a quasi-polarmission-adjust- the 'ssurface under all-weatherconditions. The first able orbit in the summer of 1991. The instrument payload spacebornescatterometer, designated S-t93, flew as a part of included the Active Microwave Instrument (aut) which was the Skylab missions in 1973. Acquiring both oceanand land capableof operating in a wind scatterometermode (5.3 Ghz) data at a frequencyof t3.g Ghz, the 5-193 experimentsdem- for the production of wind vector products. ESAis also plan- onstrated scatterometerswere feasible instruments for space- ning additional follow-on missions. (nscar), borne platforms (Sobti et al., 7975). Between |une and The NASAScatterometer designed for acquisi- Octoberof 1978, the Seasat-ASatellite Scatterometer(SASS) tion of data at 74 Ghz, is planned as part of the payload was able to obtain nearly continuous global coverageat a aboard the 1995 Advanced Earth Orbiting Satellite (ADEos). spatial unit cell resolution of SOkm until a catastrophic As discussedby Naderi et al.(1.992),NSCAT is heavily based short-circuit in the satellite'selectrical systemterminated on SASSwith significant improvements to enhance backscat- subsequent data acquisition. Although Seasat-Aoperated for ter measurement accuracy at a spatial resolution of 25 km. In only 99 days, the wind vector information derived from SAss order to support the (EoS) mission, a data conclusively showed accurate wind velocity measure- 14.6 Ghz scatterometer (EoS Scatterometer)based on the ments could be made from scatterometerson polar orbiting NscAT design has been selected for flight in the late 1990s. platforms (Jonesef al., 7982). The next section of the paper presents a very brief his- Without exception, the primary mission for all satellite tory of scatterometerimaging of tana surfaces.The subse- scatterometershas been the determination of ocean surface quent section provides some background information wind direction and velocity. This is not surprising,because concerningthe ground processingconcept for high resolution Photogrammetric Engineering & , P.J.Hardin is with the Departmentof Geography,690 H Vol. 60, No. 12, December1994, pp. 7453-7462. SWKT, Brigham Young University, Ptovo, UT 84602. D.G. Long is with the Department of Electrical and Computer oo99-77 72 | 94| 6012-7453$3. 00/0 Engineering,448 CG, Brigham Young University, Provo, UT @ 1994 American Society for Photogrammetry 84602. and RemoteSensing

PE&RS amount of confusion. As a result of studying incidence angle dependenceand diurnal variation, Birrer et o1.(1982) found the Amazon rain forest a homogeneoustarget which was azi- muthally isotropic, insensitiveto polarization, and which ap- peared to changeonly slightly from morning to afternoon. Bracalenteand Sweet (t98+) confirmed these earlier findings, and suggestedimportant corrections to improve the estima- tion of backscatterfrom SASSdata. Theseearlier studies have all found support from Ken- nett and Li's (1989b)examination of 33 broad global land- cover types in a exhaustive effort to locate suitable scatterometercalibration sites.In conducting that effort, Ken- nett and Li (198gb)found a certain amount of overlap exist- ing among the severalland-cover categories, and substantial variability existing within the categories.Furthermore, while the early demise of Seasat-Aprecluded a conclusive exami- nation of seasonalchange in global land-coverbackscatter, the three-month data set provided evidence that areas of greatestseasonality (e.g., arctic, deciduous forests)would ex- hibit the greatestvariance in backscatter throughout the year. The conversewould also likely hold true-backscatter coeffi- Figure1. The studyarea. cients in areas of little seasonality such as the central Ama- zon rain forest would probably deviate little through the months. In conclusion, becausethe Amazon rain forest was large, appearedazimuthally isotropic, was insensitive to po- scatterometry and the production of backscatter coefficients larization, presented little diurnal change, and could be ex- used in the investigationof tropical land cover types. Be- pected to change less throughout the year than other large causethe reconstructionprocess is describedintricately else- mid-latitude areas,it was deemedthe most suitable target for where by Long et 01.(19s3), this section will limited be to future calibration of scatterometersin support of their ocean conceptsrequired to appreciatethe sectionsfollowing. The imaging mission. fourth section will report the methodology used to capture and compare backscatter signatures of several tropical land- cover classesfrom the high resolution data. The goal of this A Conceptfor High Resolution Scatterometry investigationwas to assessthe potential for utilizing recon- TheSASS Deslgn structed Ku-band scatterometry to discriminate between Spacebornescatterometers transmit microwave pulses to the broad tropical vegetationclasses. The results the of discrimi- ocean surfaceand measurebackscattered power received at nation will then experiments be reported.Finally, conclu- the inshument, allowing estimation of the normalized radar sions will be drawn, and implications for future research cross section (d) of the surface.In simple terms, the radar will be suggested. cross section can be consideredto be target albedo at radar frequencies. From each illuminated location on the Earth, the ScatterometersandLand Applications total power receivedby a radar is the sum of the power Although primarily designedfor oceanicwind studies,SaSS backscatteredby the target, noise from the frequency-specific data were also collected over global regions of ice and land natural emissivity of the Earth-atmospheresystem, and noise during the ill-fated three-monthflight of Seasat-A.Recently, from the instrument itself. Once noise is subtracted from the Kennett and Li (1g8Sa)were successfulin mapping global total received power, o" can then be calculatedusing the ba- SASSbackscatter with a 111-km resolution (at the equator) sic radar equation (seeFung and Ulaby, 1983). from the original SASS50-km resolution data. From this map- A simplified diagram of the Sass antenna illumination ping, a rough, imperfect,but obvious agreementexisted be- pattern is illustrated in Figure 2. As the satellite revolves in tween global land cover and ta.6-Ghz backscatter.This was orbit, two setsof antennasare used to illuminate 475-km particularly evident in desertand equatorialforest regions swaths on both sides of the satellite track. Moment by mo- where the resemblanceof backscattermaps to published ment, each antenna produces an instantaneous footprint land-coverand geologicmaps was striking. which is several hundred kilometres long but only a few kil- While the potential of sRss data for broad land-cover ometreswide. The 475-km swath swept by these antennasis mapping was apparent,most land imaging with scatterome- then further resolved into smaller observation cells. A sass- ters has historically focusedonly on identifying land-cover classscatterometer achieves along-track resolution by a com- targetssuitably large,homogeneous, diurnally invariant, and bination of nanow antenna pattern and timing of transmit seasonallyunchanging to calibratefuture spacebornescatter- pulses,Cross track resolution is obtained by Doppler filter- ometers.In exploratory work using Skylab 5-L93 scatterome- ing, the nanow beam pattern, and the antenna illumination ter data from severalportions of the Earth, Sobti et aL (tszs) geometry (seeLong ef o/., 1988). found that, while 13.9-Ghzradar has very little surfacepene- The antennas on each side of the instrument are ar- tration, backscatterat this frequencyis very sensitiveto veg- ranged at two different azimuth angles. As the satellite or- etation cover, surfacewater, soil moisture, and physiography, bits. a resolution element on the surfaceof the Earth is These authors also concluded that targets defined a priori ac- observed first by the forward looking antenna and then by cording to physiographicand land-usedescriptions rather the aft antennaa minute or two later. This producestwo sets than expectedmicrowave responsesuffered from a certain of co-located,nearly simultaneousobservation pairs from

L4g monly accepted model for the incidence angle dependence SUBSATELLITE of o' is TRACK + 10 log,o o'(0) : A + B(0 - 4O') or alternatively f(0) : d^l19)l where A and B are independent of the incidence angle 0 and dependenton the target character,and where B is dependent orr0. It should be noted that A is the value of o'where d equals40'. The constant a is requisite for the exponential ex- pression of the logarithmic relationshipsand correspondsto the value 1.01/10. A given o{ measurementcell for t}re rb cross-trackcell and the kd antennabeam, i.e., o1 (i) will partially or com- pletely cover a number of the smaller resolution elements.In [he noise-freecase, the total power oo receivedby the scatter- ometer for this d cell will be the sum of the echo power re- turned from each theoretical high resolution element.This can be written

c2.k o2.k

or: E 2 h@,a;k)o"(c,a), u:ur* "t:"r,r where the weighting function h(c,a;k) is a function of the ra- ANTENNA3 dar parameters(e.g., antenna gain, slant range)computed_for lft+ each resolution element for the particular antenna beam k. a7skm ----+f- ago These parametersare also assumedconstant over any single high retolution element.Ignoring uncertaintiesin calibration Figure2. A simplifieddiagram of the sAssantenna illumi- and spacecraftattitude, h(c,a;k) is a known constantand this nationpattern. The four antennasare arrangedso the previous equation can be written in terms of A and B as same locationon the Earthis observedfirst by the for- few ward lookingantenna and then by the aft antennaa ,r: f i n(",o;t)o"(.pllB(c,o;k)lB(.p) minuteslater. Data from the near-nadirswath, consisting of threeadditional doppler cells, were not usedin this Multiple of oo which overlap the desired study. - cell in the high-resolution"r;;;; grid can be used to estimateA and B. Two imagesof high-resolutioncells are produced in the process,one each foi A and B. The estimation problem for a iingle cell in the high-resolution grid can be summarize1 by azimuth angles.The result of this arrangement two different the following equations.Given that F@rk,a)) and ft(c,o,'k)are o" measurementsmade of each is a set of vectors-multiple constantsfor a single oo measurementcell, assuminga rea-- by 50-km along-track/cross-track cell over an irregular 50-km sonableapproximalion that Oo(c,o): 0*, and being provided of thesemultiple measurements sampling grid. The purpose with multiple overlapping low-resolution SASSmeasurements and direction using a geo- is to derive surfacewind speed o(i), the problem is to estimateA and B where physical model relating oceansurface roughness to wind ve- locity (seeUlaby ef a,1.,1982; 1sB6). flb,a) : oar'"t[B@r(c,o))]a''"t,*i* or : h{c,a;k)drb,a). Prccessingbr HighResolutlon ""8,,7, To understandthe processfor obtaining higher resolution from the scatterometerimagery, consider a grid of high reso- The interested reader can consult Long ef al. (tssz;1993) for lution elementson the Earth's surface.Each element of this a complete discussion of the numerical approaches to the ac- grid has a cross-trackdimension of s" km and an along-track tual calculation of the A and B for each element, discussions dimension of s, km relative to the nadir track' Imperfectly concerning noise, and equations describing variance in the imposed on this grid is a secondirregular grid of the o" estimates. -"^"rnrr-"tts. Eich cell of this secondgrid has correspond- ing dimensionsgreater than the first (i.e., it is lower resolu- Applylngthe HighResolutlon Concept to SASSData tion). The goal of high resolution processingis to estimate The key to achieving high resolution by signal processing is parameterJof the smaller cells hom the o' measurementsof overlap in the SASSoo measurements. However, the SASSo' ihe larger cells. measurements made by a single antenna on a single beam ThL observedo' of a given resolution element depends during one orbit have no overlap, and the forward and aft only a small amount. on (t) the target charactersuch as roughness,conduclivity' - beam measurements are overlapped etc.; and (z) tne incidence angle and azimuth angle of the ob- Furthermore, when this lack of overlap is combined with the servation.As discussedby Birrer et aI. (7982),when it can problem of small gaps in the surface coverage for a single or- be assumedthere is no azimuthal modulation of d, a com- 6it, the result is insuffrcient data to attempt the resolution

1455 improvement schemewith data from a single SRSSpass. It is common knowledge that rainfall can attenuate back- However, due to precessionin the spacecraftorbit, a point scatterat 14.6-Ghzfrequencies. In a preliminary analysisof on the Earth's surface was observed from the same beam Amazon region SASSdata, Birrer ef al. (LS82) utilized COIS with slightly different azimuth and incidence anglesabout imagery to remove all the SASSmeasurements which showed once every three days. Data were also acquired by the instru- significant cover. The procedurewas quickly deemed ment on both ascendingand descendingorbital nodes. The unnecessaryand abandoned,as "few measurementsshould resulting set of criss-crossingmeasurement swaths in a single have been corrupted" due to the dry seasonacquisition. orbit combined from several orbits provides ample overlap to Our own empirical evidence supports the conclusions of facilitate accurateestimation of A and B over much of the Birrer ef aL.(7982). Seasat-A was launched on 27 fune 1978 Earth. and operatedsuccessfully until 10 October 1978. SASSmeas- The combination of multiple-pass colocated sAss data urements of the Amazon Basin were made twice daily. Dur- into a single set was pioneered by Bracalenteand Sweet ing the first part of the mission, acquisitionswere made (1984)and utilized by Kennett and Li (1989a;1989b), al- between 0500 and 0630 and between 1600 and 1830. During though neither group of researchersattempted to processthe the last part of the mission, acquisition occurred from 0900 data for improved resolution. Instead, the goal of combining to 1200 and from 2100 to 2400. Given the late forenoon and several orbits of data into a single set of o" measurementswas afternoonnature of convectiveprecipitation, we would have to calculate various statistics for sensor calibration purposes. expected the largest difference in measured op from rainfall In order to obtain a high-resolutionreconstructed image contamination to be evident when comparing the nighttime of backscattercoefficients suitable for land cover analysis, and afternoon SASSdata measurements of the same area. We the target should satisfy several criteria. found no such evidence in any area examined. e Thedependence of o. on the time of dayshould be smallor We would also have expected any rainfall contamination correctable. to be most apparent in the tropical rain forest areas. Statisti- o The dependence of d on tle azimuth angle should be small cally, this rainfall would lower the 4 coefEcientsand pro- or known. In the experimentdescribed later, we have as- duce high variances in the 4 coefficients for the rain forest sumed there is no azimuthal modulation in o". classes.However, the opposite was found to be true. As de- o The A and B coefhcients in the equation which summarize scribed in the results. the forest classeshave the smallest the incidence angle dependenceof d (i.e., 10log,oe@) : A variances and highest A means of any vegetation group stud- (0 - + B a0')) should remain constantfor the data acquisi- ied. Thus, it appearsthat, even if rainfall contaminatesthe d tion period. In practical terms, this implies that the land- measurements,the effect on the A imagery is minor. scapeshould not changeduring the acquisitionperiod. . Heavy continuousrainfall should not persistfor the data ac- To understandhow only minor contamination from rain quisiiionperiod. is detectable in an equatorial region renown for its roin for- est, it is important to remember that the reconstruction is Meeting these conditions is essential for generating a mean- primarily a weighted averaging function utilizing several ingful high-resolution image with minimum variation within measurements.Because multiple passesof data are averaged, Iand-cover classes and maximum statistical distance between rainfall would need to persist for dozensof days during the the classes. Two more constraints must be added which are Seasat-Amission over the same high-resolution pixel being peculiar to the image reconsbucticn itself. reconstructed.In addition, as detailed above, Seasat-A'sorbit . The topography should be relatively flat. For reasons de- changed and data were collected at different solar times. For scribedbelow, we have assumedthe topographyeffects.are any significant contamination to occur, the rain would be re- negligible. quired to fall at and late afternoon during the first part . Over the range of incidence angles used for the reconstruc- of the mission, and then switch to late morning and late : tion, the linear model 10log,oo'(d) A + B (0 - 40') ade- night until the mission concluded. These precedingtwo con- quately describesthe incidence angle dependenceof d for ditions are unlikely during the driest season year the target. of the where convectiveprecipitation is a sporadic episode,spa- While some of these issues have been explored by other tially localized, and fundamentally a high-sun event on any researchers, the exact nature of these dependencies at 14.6 specified day. It is a common misconceptionthat rain falls Ghz is only poorly understood for most land-cover types. daily in the Amazon Basin. For example, a weather station Birrer ef al. (1,982)found that a small diurnal change did oc- in Manaus, Brazil, located at a latitude of 3"S in the heart of cur in the Amazon rain forest, with sunrise measurements the Amazon Basin, records a yearly normal rainfall of 200 approximately f dB higher than measurements acquired at cm. Only 16 cm (B percent) oi tfrii falls between 1 July and different times of the day. During the 99 days of the Seasat 31 September. Other stations within the study area show mission, no change in the diurnal relations was noted. similar dry season patterns during the lifetime of Seasat-A. Changes in global backscatter over the 99-day image ac- Trying to eliminate topography's effect is somewhat quisition period were assessedby Kennett and Li (1989a) by problematic because,as mentioned before,the interaction be- (r) dividing the globe into a grid of 1'by 1' mapping cells, tween azimuth angle, incidence angle,and topographic fac- (2) calculating a 99-day mean d for each of the cells, (3) cal- tors (e.g.,elevation, slope, relative reliefl are very poorly culating mean oo values for each cell during successive two- understood,especially at several-kilomeberesolutions and week periods, and (4) mapping the deviation from the 99-day L4.6-Ghz frequencies. While elevations in the study Elreaex- mean for each two-week segment of the mission life. Based tend from nearly sea level to 900 metres in the Venezuelan on these maps, the period spanning 13 July to 30 fuly, Ken- Highlands, the primary topographic problem in the recon- nett and Li (1989) found some extremely small scattered sa- struction is terrain with severelocal relative relief over large vanna areas in central South America with d values above portions of the 50-km by 50-km SASSd measurements. the 99-day mean by 1.5 dB. Given the change in the savanna As mentioned previously, A and B coefficient imagesare which occurs as the dry season progresses, this is not sur- estimatedin the reconstructionprocess. While A is the inci- prising. dence angle normalized oo,the B coefEcientdescribes the in-

PE&RS cidence angle dependenceof oo.In areaswhere terrain is ;;;; ;d,"the act,tal measurement incidence angle may be rfiJfrtlv d'iife.ent when viewed from different azimuth angles' fnls witt result in errors in the,4 and B estimates'However' that B within the study areais very small (< o'rs)' iit""oti"e r""."tting .4 errors from topograP-hVwould be minimal a.rd uppeur"asvariance in a:ry cilculated statistical moments' Piecedent,our own evidence,and computational con- strainisluia"d oo" selection of the linear model rolog'o c(0) to = A * E fO - 40") in preferenceto higher order models a"t.riUt tiie incidence angle dependence of o' for the central 5o"if, a-".ican study a"e". Bi..". et aI' (7982) stated that itt" rit"pt" Iinear fit described the incidence an-gledePend- of'o" extremely well in the Amazon rain forest' Our own".r"" e.aohical analvsis of o' versus incidence angle (d) dem- ."rt lt"i that the linear relationship is particularly valid .""t tf," narrow incidence angle range between 23oand 55'' As describedbelow, we utilized only the SASSmeasurements imagewith a *it[i" tft"t extremeiy narrow incidence angle r-ange.for the Figure4. The reconstructedA coefficient Note the finer reconstruction.As a practical matter, models of higher order re-solutionelement size of O'04" by 0'04". beyold visiblealong the courseof the AmazonRiver when *o"fJ ."q"lte estimaiion of more parameterimages detail the A and B imagesproduced by the simple linear model' comparedwith Figure3' The computation'alcbmplexity and resourcesrequired to es- - timate ,{^and B are already foimidable the.cost of recon- higher order coeificient imageswould likely prove rtr""ti"g and horizontally polarized meas- profiifritl""."n'urthermore.given our,experierrce, it is likely of making both vertically the radar bickscatter, in these images only verti- lhat coefficient imagesof higher order would become pro- urements"of polarized data were used. Additionally, orbits for gressivelydegraded in qualitY' cally whitli the spacecraft attitude determination was in error (see as CoverDiscilmination ExPedment Bracalenteand Sweet, 1984) were excluded' Furthermore, A above,only d measurementsin the incidence an- Fij"t"t 3 and 4 are rt coefffcientimages which illustrate the mentioned -55" from 23oto with noise below a certain-prede- reJolution improvement of the new reconstruction method gle range level were used as data for the reconstruction' *h"" to central South America. Figure 3 is a 1'/2"by ierminJd A simple experiment was devised to test the utility of 1/2' resoiution"pptiud image (similar to those generatedby-Kennett - imagery for discriminating between broad ii-trsasa)) illuitrating the inherenispatial resolution of the reconstructed ves,etationclasses witf,in-the study area of central South itr"""a snsi rad"i. Figrrre 4 ii a reconstructed A image *ith 1. Using the 1:5,000,000-sCaleVegetation Map of resolution elemenl size of 0.04"by 0.04ogenerated using the Afierica. Americo (UNESCO, 1S80), large homogeneous rectan- iinr desoibed in detaiiby Long el o'l' (1es3)' The South reeions foi ra different vegetation formations were de- B imag"e"Goritfrmcontains very little information and is not shown' eular the limitationi of the map source material, ifrree"-o"tfts of data (1 )uly through 10 October 1978) were [i-it"a.h""tizing inexact natur"eof cartographic classification, unavoidable ;;;J1" g;""rate both ih"t"i-"g"t. wttile sAss was capable the seneralizations, and the difflculties in drawing exact bounda- iies for vegetationclasses which likely blend in transition zones, the"rectangular areas were delimited only forthe larger vegetationhap polygons. Thirty-two rectangular '**-t*"3 ."gIo.rt iere delineateh,iotaling over 740,000km'' The srfiallestsolitary region (35oo krn1 consisted.ofdegraded tropical forest formitions, while the largestsingle region (81,200km') consistedof Brazilian, Paraguayan,and Bolivian - bho"o. The averageregion size was 23,736 km2 this corre- sponds to approximately 925 high resolution pixels' Once theseregions were defined, the high resolution A and B im- ages for"central South America were constructed, and lixel uilrrur were then extracted for each of them and saved for further analysis'In the discrimination tests describedbelow' only one haif of the pixels for each of the 18 formations weie used for training. The remaining half were reserved for testing the discrimination'

DlscdmlnatlonMethodologY To establish a baseline estimate of discrimination ability' was an initial analysisto classify the 1B vegetationfor- Figure3. The A coefficientimage with a resolutionele there in the test set uiing quadratic discriminant functions mint size of 1-/2"by 1'/2" illustratingthe inherentspatial mations from the traininglei classes'If we designate the A resolutionof the sASsradar. senerated ind B coefficients of a single unassigned pixel (.fl as the vec-

L457 PE&RS TleLe1. Cotrustott Mnrntxron rxe 1&GnoupSupeRvtseo Drscnrurlmot Expenruerur.FonunrroN Nnues WEnE Ourreo FRoMTHE TlsLe ro Slve Spnce.gur THElDENTrw or Elcx ls GrvENrr{ TleLe 3. Actual Class Climatic Edaphic Forest Woodland Savanna Savanna Predicted Class 10 77 12 13 15 77 Forest 7 726 149 89 34 41730000000980000 2 90 7720 52 775 621102000000547000 J 5 0 92 74 00000000000000 77 767 77 4745 0574000006030000 J 6 35 4 0 8741000000134300 6 0 22 4 35 03377077000030600

Woodland 707916a27398 08919 7900962963684 80009030 2507 116 2 000000072 900020355 58 335 74 1100004701 100000000 27 194 103 920308072072 Climatic 77 0 0 77 0 3 0 43 0133 28 7220 7 92 96 34 473 372 227 Savanna 720000000000 18 280 o 771 85 77 158 88 130000000000 300 4047627277 Edaphic 0 0 0 0 0 0 20 0 31 0 0 7 0 r74 20 46 21 747 $avanna 15 000000002000013 100 77 16 0 0 0 0 0 0 0 0 0 0 28 0 47 2 2s 42 65 36 T7 0 0 8 0 0 0 0 0 13 0 301 48 27 723 85 83 346 138 18 0 0 0 0 0 0 0 6 3 2 782 0 104 t57 27 270 230 277

tor xr, and the individual vegetation group mean vectors, map were never mixed with members of the grassland cate- group covariance matrices, and group covariance matdx de- gory. Conversely, some of members in the woodland category terminants as p, (, and d, respectively,the equation for the might logically be grouped with scrub formations without in- quadratic discriminant function distanceQ between unlabe- surmountable interpretation problems. led pixel 7 and vegetation group h is given by In the ultimate discriminant analysis, the goal was to de- termine whether sufficient statistical distance in the ,4 and B - + Qp: k,- Pr)'(;'(xi tr,) ln(dh) coefficient data existed between the final supervised land- Like other classifiersof this genre,once the function calcula- cover groupings for successfulland-cover mapping. As in all tions have been made for all groups,the pixel is assignedto the previous analyses, quadratic discriminant functions were the group which generatesthe smallest distance. calculated for each of the final groups using the training data After this baselineclassification was established,the set alone. These functions were then used to classify each original 18 categorieswere continually regroupedinto a pixel from the testing set. much smaller set until the classification accuracy produced The agreementbetween each pixel's predicted and ac- by the quadratic discriminant functions reached an accepta- tual group was summarized in confusion matrix form for ac- ble level. However, potential groupings which would require curacy assessment.The quality of the supervised tenuous interpretationswere not allowed. For example,the classification was assessedby calculating both an overall floristic formations designated as forest types on the UNESCO Cohen's (1960)kappa (r) measureand a simple percentage.

TneLE2. Corurusror.rMnrnrx PnoouceD FRoM rHE Foun€Roup Supenvrseo Findingsand Discussion Dtscnruntatror,rExpEnruerur. SupenlsedDlscdmlnatlon: Membenhip and Accuncy Actual Class Attempts to classify the original 18 categories without any regroupingproduced a classificationaccuracy of 57 percent Climatic Edaphic (6 : From (Table 1) and graphi- Predicted Class Woodland Savanna Savinna 0.52). the confusion matrix cal analvsis. it was clear that substantial backscatter similar- Forest 7SO3 108 0 225 ity exist-edwithin groups roughly considered forest, 98o/o 3o/o Oo/o 5% woodland, and savanna.However, there was apparently good Woodland 95 3599 717 151 discrimination between these broad groups. Logically re- lo/o 860/0 4o/o 3o/o grouping the 18 formations into four groups improved the classificationaccuracy to 81 percent (x = 0.74). In this case Climatic Savanna I 207 1064 935 the four categoriesconsisted of humid forest, tropical wood- 17o/o 5o/o 4oo/o 20o/o lands, climatic savannas,and edaphic savannas(Table 2). Edaphic Savanna 23 285 1456 3274 Table 3 summarizes the A and B information for each of the 17o/o 7o/o 55o/o 77o/o original 18 cover categories as subdivided into these four cat- egories.Combining the edaphic and climatic savannagroups I Otal 8030 4193 2637 4585 into a single savanna category generated a marked improve- lOOo/o TOOo/o 1,OOo/o lOOo/o ment (94 percent, r : 0.91). Table 4 presentsthe confusion Tmle 3. Sur"runnvINFoRMAT|oN roR rnE 18 VEcrnflor.rFonultroNs Saupleo FRoM THE Vreeretror,r Map or SourNAuentcn. Txe NuMeeRstn Pnneuxests nrren Elcx FoRMATtoNNAME MAy BE Useoro Cnoss-RereRENcETHE Gnoup ro rne AppnopnnteRow aruo CoLUMN oF Tnele 1. Class Mean Stdv. Mean Stdv. Vegetation Formation n A A B B Forest Extremelymoist ombrophilousforest (t) 234 -7.79 0.193 -o.r24 0.002 Moist evergreenseasonal forest (2) 2too -7.54 0.203 -o.122 0.003 Ombrophilous submontane forest (3) 253 -7.50 1.006 -o.121 0.002 Very moist ombrophilous forest (4) 4974 -7.67 0.203 -0.118 0.o02 Degraded forest formations (5) 121 -7.54 0.400 -o.125 o.oo4 Tropical evergreenseasonal lowland forest (6) 300 -7.84 0.589 -o.t23 o.003 Woodland Degradedwoodland formations (7) 499 -8.59 0.373 -o.t24 0.002 Chaco(8) 2592 -9.46 0.494 -0.106 0.003 Caatinga (9) 933 -9.30 0.696 -0.118 o.007 Degradedcaatinga woodland formations (1o) 168 -9.82 o.472 -o.116 0.003 Climatic Savanna Campo cerradoN (11) 1961 -10.04 0.681 -o.122 0.004 Degradedsubhumid campo cerradoand crops (12) 342 - 11.18 0.491 -0.131 0.002 Degradedlowland campo cerrado(13) 334 - 10.65 0.563 0.119 0.002 Edaphic Savanna Campossuio and limpo (14) 1015 -10.14 1.641 -o.t27 0.004 Grasslandwith palms (rs) 327 -70.77 7.126 -o.727 o.003 Pantanal{16) 1058 - 10.33 0.938 -o.722 o.002 Cultivatedcrops [t7) 7273 - 10.43 0.903 -o.r25 0.003 Campo cerradoS (18) 918 -10.94 o.734 -o.tzl o.004

matrix resulting from the three group discrimination experi- are so heterogeneouswith variable canopy densities,species ment. compositions,plant spacing,tree heights, and canopy re- As mentioned previously, the first general group with sponse to seasonalfactors. For example, there is substantial the highest A coefEcientswas forest.It consistedof two confusion between the degraded woodland formation and the moist tropical ombrophilous forest types, moist evergreen campo sujo/limpo savanna. It is highly likely that a wood- seasonalforest, tropical evergreenseasonal lowland forest, land degraded through cattle grazing, fire, and agriculture ombrophilous submontaneforest, and associateddegraded would exhibit a great variance of backscatter dependent on forest mosaics. the areal extent of the degradation, its severity, or the succes- The secondgeneral group consistedof four woodland sional stage of any sub-climax regrowth. Thus, it is entirely vegetationformations, associated mosaics, and degraded reasonablethat some regions of degraded woodland would woodland landscapes.The specific woodland types included exhibit similar backscatter characteristics to the shrubby drought deciduous woodland known as chaco, thorn forest grassland of campo su/o or the pure grassland of campo with succulentslcaatinga, related degraded woodland forma- Iimpo. tions/mosaics, and degraded caatingo formations. The savanna group consisted of several separateshrub, Summary scrub, and grasslandcategories. These included two floristic When classifiedusing supervisedmethods, it appearsrecon- formations of campo ceftado, campo sujo/limpo, grassland structed sAss imagery has potential for discriminating be- with palms, and seasonallyflooded $assland or pantanal. tween the broad tropical classesof humid forest, subhumid Two varieties of degraded formations and cultivated crops woodlands, and savanna. There is also some ability to distin- were also members of this category. With some loss of accu- guish between climatic and edaphic savanna formations. racy, the savannascould be further subdivided into edaphic In conducting this experiment, the limitations of using a and climatic savannacategories. In the four-group discrimi- nation experiment,the climatic savannasconsisted of de- graded and non-degraded campo cenado formations whereas the edaphic savannaformations consistedof primarily of Tlalr 4. CorurusrolrMetnrx PnoouceD FRoM THE Trnee€Roup Supenuseo grasslands,wet/flooded grasslands,and associatedadjacent DrscnrurHltrottExpenrvent, southern compo cetado. Actual Class The first confusion matrix (Table 1) indicates the greatest confusion problems between the forest and woodland groups Predicted Class Forest Woodland Grass-Shrubland to be primarily a result of backscatter overlap between de- 7892 88 277 graded woodland formations,caatinga, and tropical ever- 98o/o 2o/o 3o/o green seasonallowland forest.The small amount of confusion remaining in the table exists primarily between Woodland 110 3763 377 the woodland and savanna groups. While there is some back- 7o/o 90o/o 5o/o scatter resemblancebetween most members of the woodland Grass-Shrubland 2A 342 6628 and savanna groups, the two degraded woodland formations 17o/o 87o 9zVo appearto be major culprits in severalinstances. This confu- Total 8030 4193 7222 remote sensingtask of distinguish- sion reflectsthe common TOOo/o TAOo/o TOOo/o ing between woodland and shrubland vegetation when both

PE&RS 1459 map as ground truth for understanding the interaction be- Currently, there is a substantial amount of researchbe- tween 14.6-Ghzscatterometry and tropical vegetationbecame ing conducted to map the tropical rain forests of the world readily app^arent.By its very nature, traditionil vegetation using NOAAAVHRR and eRs-r SARdata (TREES,1991). In mapping qf-large regions at small scales requires generaliza- this endeavor,reconstructed sAss imagery of the world's tion, simplification, discreetboundaries, and thuJhides the global rain forests can provide another important historical complexity of the heterogeneousmixtures that occur in real- data source. Furthermore, when applied to cuuent ERS-1 ity. This mixture was readily detectedin processingthe re- scatterometry,the reconstruction processmay allow us to con^structedim?ge. Becauseof this heterogeneity, oily three inexpensively extend high resolution ERS-1SAR tropical for- or four broad classeswere capable of being distinguiihed us- est canopy measurementsfrom small test sites to larger con- ing an o priori classification scheme, and further s-nbdinirio., tinental regions. of the classeswas difficult. However, the potential for utilizing reconstructed SASS Only a general statement of scattering differbnces be- (or other scatterometer)imagery goes beyond land-cover tween forest,woodland, and savannacaribe offered because. mapping or monitoring. The detection of several vegetation unfortunately, 14.6-Ghzvegetation scattering has been stud- class gradations in the reconstructed imagery indicates the ied very littl-e.It has never been studied at S-km pixel scales strong possibility of obtaining critical quantitative informa- where mixed pixels present interpretation problems.Unlike tion about tropical vegetation community character from re- lower frequencyradars where volume scatGringin vegeta- conshucted scatterometerimagery. It is still not clear tion dominates, Ku-band radar primarily scatteis off tlie sur- whether 74.6-Ghzbackscatter coefficients change in response face_ofheavily leafed canopies.The high density and to canopy density, canopy vigor, another canopy structural random orientation and spacingof the leaves in-the tropical characteristic, or more complex quantitative variables. Using forest produce diffuse scattering which results in bright sig- higher resolution SaR,visible, thermal, and imagery nature at the nauow range of incidence angles used in thJ of carefully selectedverification sites as primary data project-The_ extremely moist, heavily canopied,tropical om- sources,it may be possible to describe backscatter-based brophilous forest serves as an excellbnt example of-this classesl9t by simple name alone, but more precisely in strong backscatteringresponse. On the other hand, substan- terms of biomass,moisture content, transpiration rat-es,leaf tial lossesare imposed from multiple reflectionsin the verti- area, or as a percentageof each speciesconstituting the tar- cally structuredgrass component of savannas.As a result, get. This information is as important to understandlng savannasdominated by grassland such as campo limpo will global geophysical processesas simple nominal mapplng of appear very dark. Park-like savannasand woodlands will vegetation.As an example of this approach, our current re- demonstrate a mixture of the two scattering mechanisms and search involves modeling AVHRRderived parameters(e.g., will have an intermediatebackscatter respdnse proportional NDVI)as functions of reconstructedSASS A and B coefE- to the relative amounts of canopied overs1ory,giasiland, and cients. bare grou-ndwithin the pixel. Areas of flooded grassland, The SASSimage reconstructiontechnique also holds woodland, or forest will appear darker than thelr dry coun- great promise for land imaging of other regions of interest terpa-rtsbecause of 's high dielectric constant. during the late summer months of 1978. These include the world's boreal forests,Greenland, and the Antarctic. The po- tential for imaging glacial ice is particularly fascinating, b-e- FinalComments causescatterometer images provide multiple incidence angle Given the successof these discrimination experiments,fur- obsewations. It may be possible to create several coregisteied ther research-into tropical land scatterometry may bring sub- high-resolution images, each possessinga different range of stantial payoffs, While the goal of this projeit was not io incidence angles,to infer ice age,water content, snow cover, createa general19zB vegetation map ol central South Amer- or other parameters of interest. ica, it may be possibleto do so with SASSdata. In such a Given the potential for scatterometersto obtain useful project, much of the confusion in the purely backscatter- information over land, it may be profitable for the scientific basedclassification could be removed by incorporatingtopo- community to " carefully consider both ruscar and its follow- graphic, climatic, pedologic,and other incillaly data. on EoS-erascatterometers as potential sourcesfor land imag- Inclusion of avnnn data collected during the sametime pe- ing. According to Long ef 01.(1993), minor modifications riod is also possible.If the effect of topography on 14.6-Ghz involving changes to the digital subsystem combined with backscattercan be controlled, it may 5e poslible to extend the ground signal processingalgorithm outlined above could the procedurg rgsingthe global SaSSdata set to all the tropi- potentially produce reconstructedimage resolutionsbetween cal regions of the world in order to estimate and map the- 1 and 3 km. These changeswould make NSCATand the EoS- 1.978extent of the global tropical forestsinto broad catego- era scatterometersdual mission instruments which could ries. With the launch of RnEoSin 1995 and the high resolu- provide ground imaging at 74 Ghz with a spatial resolution tion ground processingof wscar imagery using similar compatible with several other EoS instruments, including the techniqr'-es,the potential may exist to assesschange in the Multifrequency Imaging (unr,m), tropical forest character and extent over the 17-vear interval Multi-angle Imaging Spectro-Radiometer(Utsn), and the at identical spatial resolutionsand frequencies.Unlike the Moderate ResolutionImaging Spectrometer(uoots). SASSdata which covered only three months, NSCATcover- age 9t the equatorial and savinna regions for the entire year would also allow change in backscatterthrough the seajons Rehrcnces to be used as a discriminator for more accurafevegetation Birrer, I. E. M. Bracalente,G. Dome, Sweet,and G. Berthold, mapping. It wou-ld also allow real-time t., f. f. monitoring of global 7982. d Signature of the Amazon Rainforest Obtained from the tropical seasonalchanges without interference from cloud SeasatScatterometet, IEEE Transactions on Geoscienceand Re- cover which afflicts visible and near-infrared instruments. mote Sensing,GE-2O(1):17-77.

1480 PE&RS 1993) Bourliere, F., 1983. Tropical Savannas,Vol. 13 of Ecosystemsof the (Received7 August 1992;revised and accepted28 April World, (D.W. Goodill, editor), EIsevier,Amsterdam. Bracalente,E. M., and f. L. Sweet, 7984. Analysis of Notmalized Ra' dar CrossSection (e) Signature of Amazon Rain Forest Using Appendix SeasatScatterometer Data. NASA Technical Memorandum 85779,NASA NTIS, VegetatlonDeflnltlons Tlie definitions below were adapted primarily from Lieth and 1960.A Coefficientof Agreementof Nominal Scales,Edu- Cohen,f., (1989), (1981),and Bourliere (1983)' catio na| and Psychological Measurement' 2O:3746. Werger UNESCO Fung, A. K,, and F. T. Ulaby, 1983. Matter-EnergyInteraction in Caatinga. Semi-deciduous thorn scrub with succulents' It is ih" Mi".o*"ue Region,ChaPter 4 in Manual of Remote Sens- found in northeastern Brazil in areas where the climate rng, Vol 1. (R.N. Cotwell, editor), American Society of Photo- is subhumid to arid. It is bordered on the east by the grammetry, northern variety of campo cerrado and on the east by fones,W. L., L. C. Schroeder,D. H. Boggs,E. M. Bracalente'R. A. the coastal Atlantic forest and related degraded Brown, G. f. Dame,W. J. Pierson,and F. f. Wentz, 1982.The formations. Climax caatinga is ligneous, drought- Seasat-ASatellite Scatterometer:The Geophysical Evaluation of deciduous, thickly branched, narrow leaved, and Remotely SensedWinds over the Ocean.Joutnal of Geophysical physiognomy of Ihe caatinga is -33u. sometimes thorned. The Research,87 t3297 variable. In some areas iimay resemble shrubland Kennett,R. G., and F. K. Li, 1989a.Seasat Over-Land Scatterometer steppe, and in other regions may be a dense scrubland BackscatterCoeffi- Data, Part I: Global Overview ofthe Ku-Band with canopies reaching 10 metres. Cactus may also be IEEE Transactions on Geoscienceand Remote Sensing' cients, found. GE-27(5'):592405' Campo Cenado (N). Medium-tall grassland with several ScatterometerData, Part II: Selec- 1989b. SeasatOver-Land of broad-leaved evergrien trees and shrubs. This Area Land-Target Sites or the Calibration of ipecies tion of Extended and forms a zone Scatterometers.IEEE Transactions on Geoscience formation is found in central Brazil, Spaceborne north an d Remote Sensing,GE-27 (6):7 7 9-7 88. between the Amazonian Forests to the east and and the caatinga along its eastern border. Physiognomy H., and M. A. Werger, 7989' Tropical Rain ForestE"7tyt- Lieth, f. -ond campo cercado (N) varies from a shrub savanna to tems:Biogeographical Ecologicol Studies,Vol. 74b of Eco' of sysfemsof tie World (D'W. Goodall, editor)' Elsevier, woodland savanna of light to dense canopy densities. Amsterdam. Gallery forests and a-reasof grassland are also common. have thick Long, D. G., C. -Y. Chi, and F. K. Li, 1988.The Designof an On- Shrub; and trees average 5 metres in height, toard Digital Doppler Processorfor a SpaceborneScatterometer, bark, twisted branches, and partially deciduous leaves IEEE Transactioii on Geoscienceand Remote Sensing,GE-26: which may be hairy and leathery. Anthropogenic change 869-878. and fire in the campo cerrado (N) make the landscape Long, D. G., and P. J. Hardin, 1993.Vegetation Stu-dies of the Ama- exhemely variable. lon Basin Using Enhanced Resolution SeasatScatterometer Campo Cerxido (S). Medium-tall grassland with several Data, IEEE Traisactions on Geoscienceand Remote Sensing (in ipecies of broad-leaved evergreen trees and shrubs. This review). formation is located in southern Brazil and north-central (N) Long,-hancement D. G., P. I. Hardin, and P. T. Whiting, 1993.Resolution En- Paraguay.It is distinguished from campo cenado by of SpaceborneScatterometer Data' IEEE Transactions its shorter dry seasonand a definite cool season.In on Geoscienceand Remote Sensing, (in press). general, it poisessesa denser understory than the Long, D. G., P. T. Whiting, and P. ]. Hardin, 1992.-HighResolution northern campo ceftado. Primarily grassland with Land/Ice Imaging Using SeasatScatterometer Measurements, sedges,shortel palms are sometimes common. Edaphic 'gz IGARSS Inteinatioial Geoscienceand Remote Sensing Sym' factors are partitularly important to explaining the posium, 7:44O442. various facies of this formation. Naderi, F. M, M. H. Freilich, and D. G' Long, 1992.Spaceborne Ra- Campo sujo. A prairie grassland with scattered shrubs. dar Measurementof Wind Velocity over the Ocean:An Over- Wittriir the itudy area of this paper, they are located in Proceedingsof the view of the NSCAT ScatterometerSystem. northwestern Brazil IEEE,79(6):850-866. Campo limpo. Tall prairie grassland nearly devoid of shrubs Sobti, A., R. K. Moore, and F. T. Ulaby, 1975.Backscatter Response ind tre-es.Within the sludy area, these areas ire mixed Orbit,Tech- at 13.9GHZ for Maior Tettain Types as Seenfrom with compo sr4b in northern Brazil. Report243-4,NASA, Lyndon B. SPaceCenter, nical Johnson Chaco. Drought deciduous lowland (and submontane) Houston,Texas, August 15. woodland. The chaco within the study area is found in Proposd 1991-1992.TREES Technical Series TREES, 7997. Strategy Bolivia, along the plain between the Paraguay and A, No. 1, A of the Commission of the European JoinTProlect and along the Rio Paraguay in Communitiesand the EuropeanSpace Agency, SPI.90.31,foint Parana rivers in Paraguay, ResearchCenter, Ispra, ItalY. southern Brazil, This woodland is typically multi- canopied and deciduous,with cactuslocated in drier and A. K. Fung, 7982.Microwave Remote Ulabv,-Sensing: F. T., R. K. Moore, and edaphic variations Active and Passive, Vo1' 1. Addison-Wesley. a.reas.There are many climatic on a local scale. Microwave Remote Sensing:Active and Passive, Vol' 1986. phytogeographical complex covering over ///. Artech House. Pantanal. A 100,000 km' along the Paraguay River and its tributaries Map of South America. 1:5,000,000'two UNESCO,798O. Vegetation Paraguay.The ponfonal is by the United Nations Educational,Scientific in Brazil, Bolivia, and sheets,publish-ed grassland and Cultural Organization,Paris. primarily a humid savanna of hygrophytic _ whictr is 80 percent flooded during the rainy season.In 1987. VegetationMap of South Americo Explanato-ry No-tes' this grassland is Publishedby the United-Nations Educational,Scientific and areas of higher local relative relief, Cultural Organization, Paris. replacedby shrubs, woodland, and forest.

PE&RS Perry I. Hardin ttt|ft" Davrd G. Lone Perry Hardin is an Assistant Professorin the Ge- David Long is an AssistantProfessor in the Elec- T*.IIFrn , ography Department at Brigham Young Univer- Iq$ff trical and Computer Engineering Department at sity and directs the nyu Laboratory for R#.I Brigham Young University. He has a Ph.D. in GeographicInformation Analysis. Holding ^ a TRlil'w Electrical Engineering from the University of - Ph.D. in Geography(Remote Sensing) from the I m r Southern California. He is the principle investi- University of Utah, he is an investigator on several projects gator on several NAsA-sponsoredinterdisciplinary research involving the integration of remote sensing and crs for re- projects in remote sensingand has numerous publications in sourcemanagement. His researchinterests include biogeogra- signal processing and radar scatterometry. His research inter- phy, image processingtheory, climatology, and habitat ests include microwave remote sensing, radar, signal process- modeling. ing, and mesoscaleatmospheric dynamics.

ISPRSCommission lV Proceedingsof the Symposium Mappingand GIS

HeId May 31 - June 3, 1994,Athens, Georgia,USA

Sponsoredby the Center for Remote Sensing and Mapping Science (CRMS) and the American Society for Photogrammetryand Remote Sensing (ASPRS)

Edited by: Roy Welch and Marguerite Remillard

This Proceedings of the Intemational Society for Photograurmetry and Reutote Sensing (ISPRS) Commission IV Syurposium, "Mapping and Geographic Inforuration Systems,"conpletes the transition of Commission IV from a focus on mapping databaseand GIS applications. The 103 technical and poster pap€rs feature in this Proceedings are organized by Working Group, and represent the efforts of authors from more than 28 countries. It is a collection of state-of- art papers ou mapping and GIS applications that belongs on the bookshelves of photogrammetrists, remote sensing scientists aud geographic infonnation systen specialists. WorkingGroup Topic Areas lnclude:

Working Group lV/l Geographic Information System Data and Applications Working Group lV/2 International Mapping frour Space Working Group lV/3 Map and DatabaseRevision Working Group lV/4 DEMs and Digital Orthoiurages for Mapping/GlS Applications Working Group lV/S Extraterrestrial Mapping Working Group lV/6 GIS and Expert Systeursfor Global Euvironurental Databases Working Group llUlV ConceptualAspects of GIS

1994. 710 pp. $70 (softcover);,,4SPRSMenbers $40. Stock # 4633. For moreinformation, see the ASPRSStore.