Introduction

The city of Angkor exploited waters flowing down from the Kulen, a Khmer expertise which grew out of a earlier tradition of water management: control in periods of flood and conservation during the dry season. Vestiges of this tradition are detectable on images generated from the NASA/JPL SIR-C data, acquired on September 30, 1994 during the 15th orbit of the space shuttle Endeavour.

Radar data from three scattering mechanisms form curvilinear patterns which are described here for eleven mound sites in the Angkor region. The sites average 200 metres in diameter and are located at an elevation between 11 and 23 metres above sea level. The curvilinear patterns do not necessarily enclose the mound but are in sectors, which may relate to slope and drainage. It is suggested that these patterns are manmade moats and dikes, remnants of pre-Angkorean (pre-9th c. AD) Khmer water management.

Archaeological applications of the NASA/JPL radar have principally been conducted in arid regions. These have detected earlier manmade linear features (e.g. roads at Ubar and along the Silk Road) , or climatic changes affecting habitation (e.g. desiccated subsurface water channels in the present day SafSaf Egyptian desert). Subsurface detection relies on an absence of ground level moisture and vegetation, allowing shorter and longer wavelengths (both C-band [6cm] and L-band [23cm]) to sense variations in the soil's dielectric constant which may be characteristic of former river beds. At Angkor, stratigraphic interpretation is not completely precluded, as the radar backscatter presents hydrological patterns of different chronological periods. However, the greatest potential is the discrimination of surface variations resulting from the radar's sensitivity to moisture and vegetation.

Initial examination of the Angkor scene was of a colour composite (Fig.1; LHHLHVCHV [rgb]; Fig.2 map). The LHV band (green) dominates Mt.Kulen, highlighting relief on interior parts. The mid-floodplain may be separated into three sectors from west to east: Puok, Central Angkor, and Hariharalaya. The central swath contains one 'corridor' of bright return spotted with small tanks leading down to the North Baray. Other rectilinear features, either water control devices or temple enclosures are also visible.

Further study of the SAR data from Angkor included examination of distinct features on images produced from the many possible polarization representations of the data (e.g. total power, hh, vv or hv cross-section, synthesised cross-sections for arbitrary transmit and received polarizations, hh-vv phase difference, hh-vv correlation coefficient, etc.) It seemed pertinent to attempt to link particular scattering mechanisms to various man-made and natural features at Angkor.

However, developmental generalizations were questionable given the large number of input parameters: manmade archaeological features (e.g.prehistoric mounds, barays, linear dikes, roads, temple enceintes), natural elements(e.g. rivers, streams, forest), and agricultural (e.g. rice paddy, a range of other crops, forest fellings) to support. The available literature which offer explanations of particular polarimetric signatures are mathematically based, sometimes for example, yielding combinations of three scattering matrices. These, however, are difficult to relate to physical scattering models, and in turn to the landscape of Angkor.

The classification, decomposition and modelling of polarimetric SAR data has received much attention in recent literature. Many complex models of polarimetric radar backscatter have a larger number of ground-truth input parameters than the radar measurement output parameters.For example, in a forested area input requirements might include trunk dielectric constants, ground roughness and dielectric constant, branch size and angular distribution, measurements of tree heights and diameters, and tree density. Given availability of such information, these models are predictive, capable of solving a 'forward' search for particular polarimetric signatures. However, it is difficult, if not impossible, to invert these models to provide a unique initial interpretation. This dilemma, seen at Angkor, has been a major problem in analyzing polarimetric SAR data such as that produced by the NASA/JPL AIRSAR and SIR-C systems.

Scattering mechanism model

The model used in this paper classifies polarimetric SAR observations in relation to three scattering mechanisms. ¹ These are even or double-bounce scatter from a pair of orthogonal surfaces with different dielectric constants, volume (canopy) scatter from a cloud of randomly oriented dipoles, and Bragg scatter from a moderately rough surface. In the L-band and C-band images generated using this model these mechanisms were displayed in the red, green and blue channels respectively. While the model has had various applications, including tropical rain forest areas, it has not been applied in Southeast Asia, and not in relation to archaeological investigation of the terrain.

Previous applications of the model have included a tropical rain forest area, Belize; a boreal forest site, Alaska; an arid semi-desert site in Wyoming; a mixed agriculture site, The Netherlands; a mixed urban/pine forest site, Germany. The most relevant of these to Angkor is the Belize study, although the results were based on AIRSAR P-band as well as C-Band and L-band, and compared backscatter over a range of incidence angles between 21 and 58 degrees (the Angkor SIR/C data is at an angle of around 40 degrees). In the Belize study, an area measuring 12.3 x 12.6 km was classified into types of landcover. The results of the study showed that the model could be used to the first order to determine the dominant scattering mechanisms which give rise to observed backscatter in polarimetric SAR data. This in turn allowed assessment of the fit of the model for the different land cover types to the surface, volume and double bounce components. These proved useful in differentiating between different surface cover types, as well as monitoring changes in surface cover. The following figure (Figure 1, A. Freeman) illustrates the three scattering mechanims for the Belize study.

The same potential is seen in applying the model to Angkor where classification of landcover is fundamental in understanding terrain and hydrological preferences of prehistoric and historic settlements and water management structures. Decomposition of the radar signatures or measurements into different scattering mechanisms allows the observer to differentiate between different land forms. For example, terra firme in the form of raised dikes and mounds can be separated from areas prone to flooding. Rice paddies can be differentiated from areas overgrown with denser vegetation. Circular 'moats' and linear-form canals can be detected by the effect of water on the radar signature, when present beneath a layer of vegetation or with no vegetation cover.

Figure 3 Application of the model to a three-frequency, multi-polarization AIRSAR data set over the Gallon Jug area (A.Freeman)

Puok-Mokak mounds

In the study of early water management in the region of Angkor, the richest distribution of sites is found to the west of the historical urban area, north of the present town of Puok. These extend further north, some twenty-five kilometres to, and past the town of Mokak. The Puok and Mokak mound sites are located along the limits of an alluvial fan spreading southwest from the Kulen massif. Within what appears to be an ancient river bed forming the Puok-Mokak 'corridor', these slightly raised mounds - now just off the edge of the alluvial fan - were isolated through a combination of processes. These included gradual erosion and downcutting of water along the edge of the alluvial fan (north to south and northeast to southwest), continued deposition of sediment, and aeolian action.

The Puok-Mokak 'fan' is separated in its southern portions by remnant streams flowing northeast to southwest. These diffuse the margins of the terrace in contrast to the clear perimeter of the northern bulge. This pattern is repeated to the east of the Roluos River, but flipped vertically, so that the more clearly delimited portions are to the south. Chau Srei Vibol rests on the western edge of one of these. As with the Puok area to the west, this combination of canopy and more inundated drainage areas presents a hospitable zone for prehistoric occupation. One of three dikes containing Angkor period kilns has been reported in this area, southeast of Chau Srei Vibol. ² It is possible that the Indratataka takes advantage of another of these interfaces. However,the later building of the East Baray, as well as the diversion of the Stung Siem Reap have so altered the hydrology of the Hariharalaya area that reconstruction of the original setting is problematic.³

Both prehistoric and Angkorean habitation relied on the movement of water across the floodplain. The Puok-Mokak mound distribution took advantage of two rivers: the ancient course of the Stung Siem Reap, and the Stung Trahling (Ta-net) rising north of the western portion of the Kulen Massif. This primarily feeds into the Stung Sangke emptying into the northwest corner of the Tonle Sap. Tributaries flowing south lightly downcut the edges of Kulen's old alluvium, separating vegetated zones from barer or inundated areas .

Isolated patches, today village mounds, are generally located west and south of the main spread of the alluvium. Only one, Lovea, is seen as moated on aerial photographs and SPOT images. On scenes generated from the radar data, other mounds appear moated. These sites benefited from the slope of the land which brought water to the mound perimeter. Manmade moats in some cases exploited this collection process, acting as diversionary structures during times of inundation.

The southern portion of the Puok valley is of particular interest, as the ancient course of the Stung Siem Reap flowed into this area. Clear evidence of this can be seen on the SIR-C images. There is a roughness in the radar return, with contrasting areas of canopy and smoother, inundated ricefields - many of which study by Claude Jacques of the SPOT images have shown to bear remnants of ancient bunds. The northeast to southwest outflow of water here in former times is also demonstrated by the u-shaped dike located midway between the mound sites of Phum Reul and Lovea. Lovea, however, is also - and perhaps more closely - associated with a more homogeneous portion of alluvium to the north. Phum Trei Nhor, and Phum Nokor Pheas are also located just off this more northerly spread. The Phum Sret area [processed as the Roluos scene] also contains these variations in alluvial formations.

Angkor and Northeast Thailand

All of the eleven sites where curvilinear patterns were studied are mounds; some are inhabited; all are rural. Three are on the Puok scene: Phum Reul, Lovea, and Phum Trei Nhor. Two are on the Hariharalaya scene: Phum Stung and O Dek. Six are on the Mokak scene: Mokak, Sala Khum Ta Saom, Phum Pongro, Phum Chruoy Chakrey, Phum Tonle Sa, and Phum Nokor Pheas.

Comparison may also made with early forms of water management in North East Thailand, specifically mounds surrounded by earthworks and moats. The patterns visible on images generated from the SIR-C radar data are similar in form to moats around sites in Northeast Thailand. With some exceptions, the moats of the Angkor sites are vestigial, not typically apparent on the ground, and not visible on optical imagery.

As suggested by the inclusion of nine out of eleven sites in this region, the Puok and Mokak scenes have proved the richest for the present study. The sites are not large but similar in size to many of the mounds of North East Thailand. Among the moated sites in Buri Ram province on the Isan scene generated from the SIR-C data, Ban Sai Rayong (14.58n x 103.01e) and Ban Talung Kao (14.39n x 103.02e) are some 300 metres in diametre. Muang Fang 14.49n x 103.00e) is larger, about 355 metres east-west and 466 metres north-south. It is also surrounded by three and possibly four earthworks. However, there is not necessarily a correlation between mound size and number of earthworks. For example, the mound of Ban Takhong, Buri Ram province (15.13n x 103.20e) , also with three earthworks, is 250 metres. That said, multiple earthworks are not generally associated the floodplain sites of North East Thailand, and to date, it is floodplain sites which have been investigated in the Angkor region. ⁴

Description of sites

Mokak sites

Mokak (13.39n x 103.42e) is a problematic site, for it appears on aerial photographs to have water management structures on is northeast side, whereas the majority of sites show a stronger indication of water control on the downslope side, the southwest. However, at Mokak, there is also an Angkor period tank, and a clear northeast to southwest inflow of water. This extends beyond the area processed for the Mokak scene.

Sala Khum Ta Saom (13.39n x 103.41) presents a very circular appearance with a continuous moat and possibly an outer earthwork. An area 500 metres north of Phum Ka Ro Lum also shows traces of a moat. It is one kilometre west of Ta Saom, on the opposite bank of the stream running northeast to southwest.

Phum Tonle Sa (13.38n x 103.43e) and Phum Pongro (2) (13.37n x 103.43e) are four kilometres south east of Mokak. On 1954 French aerial photographs Phum Tonle Sa presents clear evidence of water pooling on the south of the mound. An additional barrage and small tank are located to the northeast. At Pongro, the outer perimeter of a band of ricefields encircling the mound suggests that remnants of an earthwork remain. Phum Romiet, 2.5 kilometres northeast of Phum Pongro, also shows evidence of a moat, but was not included in the data processed for the Mokak scene.

Phum Chruoy Chakrey (13.36n x 103.42e) and a site 500 metres north east of Phum Nokor Pheas (13.36n x 103.43e) parallel the main north-south road (Rte. 671) at the same latitude, on the west and east respectively. Nokor Pheas is interesting as it appears to be uninhabited, although 1:50,000 maps do note a temple. It forms a 'bridge' across the south end of a northeast-southwest inundated strip, possibly a former streambed. In this context, it - and all the sites on the Mokak and Puok scenes except Chakrey - conform to the pattern of being slightly isolated from the main spread of alluvium. Chakrey's hydrological relationship is to Mokak and Ta Saom to the north.

Puok sites

Phum Trei Nhor (13.34n x 103.43e) is 3 kilometres further south along the road. As with Nokor Pheas and Phum Chakrey, there is a parallel site on the west of the road, Phum Thipdei, although it was not included in the present study.

Phum Reul (2) is initially difficult to distinguish on images generated from the radar data. The mound, like that of Trei Nhor, is nestled close to the edge of the alluvium. However, the backscatter from the mound is merged with return from three to four villages to the south and south west of the mound. The most southerly of these may have an inundated area on its south western perimeter. Thus what initially appears as a contiguous mass is in fact several different areas. Further verification is needed for the diametre of Reul: measured as 200 metres, ground check in December 1996 suggests the mound is larger.

The Lovea mound is recorded as being the largest of the sites with a diameter of 350 metres. Further ground survey, however, may reduce this. For example, the present monastery, on the east side of the site, may not have formed part of the original mound. On the other hand, Lovea remains unique in the visible remnants of two earthworks, which may be a reflection of an importance accorded to its size. Other interesting aspects of the site are the u-shaped bund or dike midway between Phum Reul and Lovea, the canal running north from the site, the east-west dike north of the site, and the large rectangular baray to the south.

Hariharalaya sites

The Hariharalaya scene is the darkest of the three, and the most difficult to read in any polarization combination or in the images generated from the scattering model. The scattering model data for Stung and Roluos Chas was interesting nonetheless as both sites were visited by the author several times, between December 1992 and April 1995.

An exterior earthwork is faintly visible at Phum Stung on aerial photographs [most visibly on the 1945 Williams-Hunt Collection, somewhat on the French 1954 1:40,000 cover]. Moat remnants were confirmed on the ground on the east and west sides of the mound. With these known locations, it is possible to brighten the Hariharalaya scene to the extent where they can be seen as green strips flanking the darker area of the moats. Phum Stung is largely uninhabited, with parts given over to upland crops. A stone tool was recovered from a corn field, and a few pottery sherds.

Ground survey at Roluos Chas confirmed the presence of a raised mound, flanked on the east by a stream. The site is quite densely inhabited, and despite a conviction that this should be a village of longterm occupation, no finds were recovered. In addition, the radar return was dark, and enhancement failed to show up the type of curvilinear pattern present at the other sites. Thus the site was not included in the final tabulations.

O Dek (O Spean Dek or Kaek, 13.22n x 103.57e) is an ambiguous site. It appears quite clearly on aerial photographs as a mound nestled in a curve of a small stream ['o']. However, it is only possible to assess it with the radar data by considerable brightening of the enhanced L-band image generated from the scattering model. The mound then becomes visible, although ground survey (not yet possible due to security problems) may confirm a smaller mound diametre than that recorded (300 metres).

Study of curvilinear patterns: method

Although the patterns are visible on the LHHLHVCHV composite, points forming the curvilinear patterns at the Angkor mounds were selected from enhanced L-band scattering images [ScatLenh.tif] to focus on the double bounce, canopy and odd scattering mechanisms (red, green, blue on the rgb images). Points were selected where they formed patterns around the generally higher volume return of the central mound, and tabulated by geographical sector: NE, NW, SE, SW. They were then averaged to give six measures for each site, expressed in decibels (dB): C-vol, L-vol, C-odd, L-odd, C-dbl, L-dbl. A number of sites have an additional exterior 'ring', or fragmented curves, probably remnant earthworks.

A total of 20 to 28 pixels were measured for each site. In most cases the pattern pixels formed a single row, although there were some parts of the curves that widened to two pixels. Given a resolution of approximately 25 metres per pixel for the processed SIR-C data, this width was in keeping with the measurable width of the visible moat at Lovea, and also the average of thirty metres for the moats of North East Thai sites.⁵

Table 1. Average return (in decibels) of curvilinear patterns for volume, odd and double bounce at C-band and L-band

	C-vol dB	L-vol dB	C-odd dB	L-odd dB	C-dbl dB	L-dbl dB	dia [m]	elevation m
P1 reul	-6.0	-9.9	-16.1	-19.3	-16.6	-17.0	200	18
P2 lovea	-7.0	-14.9	-24.6	-23.7	-20.3	-12.8	300	11
P3 trei nhor	-4.8	-9.8	-11.5	-16.0	-12.3	-11.6	200	19
M4 mokak	-7.5	-16.6	-14.9	-21.7	-11.4	-18.3	160	23
M5 ta saom	-5.4	-15.4	-17.3	-16.6	-15.3	-10.5	200	23
M6 tonle sa	-6.6	-15.3	-19.2	-17	-15.6	-11	180	22
M7 chakrey	-7.0	-13.9	-15.6	-23.7	-11.6	-14.6	170	20
M8 pongro	-7.3	-15.0	-16.4	-23.8	-9.4	-13.8	180	20
M9 nokor pheas	-5.4	-13.5	-23.5	-14.7	-19.1	-14.5	200	20
H10 o dek	-7.2	-16.6	-17.7	-13.2	-19.9	-18.9	200	15
H11 stung	-6.1	-12.2	-12.8	-13.1	-18.8	-18.1	200	12
AVERAGE	-6.4	-13.9	-17.2	-18.4	-15.5	-14.6	199.1	18.5

The measurements were taken from files generated by A. Freeman. The data was transformed from a 1-254 scale to decibels with Excel, using the formula: 10*LOG10((255 -C27)*0.0039212 + 0.0001). The decibel measurements were then averaged per sector and as a whole. Data from Lovea and Ta Saom are included as examples.⁶

It was not expected that every site would be the same, for the presumption is that, like North East Thailand, that each site is unique, and that any manmade alterations such as moats are in accordance both with the contours of the mound that they encircle, the available water in the vicinity of the mound, and the slope of the terrain. The greatest value of the returns, is in fact in their variation, providing a remarkably sensitive indicator of what to look for on the ground to detect the presence of moats.⁷

Volume return

The C-vol return was generally brighter than the L-vol, although there was variation in both. The standard deviation for the C-band volume was less than that for the L-band volume [+0.91 versus +2.38]. A brighter C in both would strengthen the case for low growing vegetation over water. A higher return for C-vol than L-vol may suggest low growing vegetation over moist or inundated areas , which might explain why the moats are not easily seen during ground survey.

Sites of particular interest to check in this regard would be Ta Saom and O Dek, although they are difficult to access. Lovea, more accessible, could be checked on the southwest sector where the C-vol and L-vol difference was quite high [-6.3 versus -22.4] , and then compared to other sectors, such as the southeast or northwest where the difference is less [-6.5 versus -11.7, and -7.7 versus -12.7].

Odd return

The odd return was not as consistent, at times being nearly the same, and at others the C-band or L-band exceeding the other. The averages for the sites at both bands had the highest standard deviation of the three scattering mechanisms [+4.0 for C-band and +4.2 for L-band). Only at Nokor Pheas was the L-odd much brighter than the C-odd [-14.7 versus -23.5]. In comparing the odd and double bounce return, many C-band returns were similar. Pongro was an exception with a double bounce return of -9.4 but only -18.4 for the C-band odd. This was the case for all four sectors of Pongro, although the difference was the greatest in the south west sector [-13.6 versus -4.7].

Double bounce

As with the volume return, a high or bright reading for the double bounce at C-band but not at L-band may suggest low growing vegeation over moist soil or water. This was the case at Mokak, where a bright C-return and dark L-return was seen on average for all three scattering mechanisms. At Ta Saom the L-band return was less [-15.3] than the C-band [-10.5]. At Tonle Sa the double bounce was -15.6 at C-band and -11.0 at L-band. But at Lovea, the L-Band was much brighter than the C-band [-12.8 versus -20.3]. Recalling the much brighter C-band return for volume in the southwest sector of Lovea, it is of note that the double bounce difference was primarily generated by the data from the southwest sector: the C-band return was only -27.6 while the L-band was -8.3. Mokak and Pongro provided the most consistency, with all or nearly all sectors having brighter C-band than L-band returns for volume and double bounce. In many cases, however, the difference between the two wavelengths for the double bounce was small.

Comparison volume, odd, double-bounce return

Interpretation of the results from all eleven moats at the mound sites (Table 1), shows that C-band volume scatter is relatively high (>-7.5 dB) for all sites. In all cases, the volume scatter exceeds the odd or double-bounce. Taken together, in comparison with data from other tropical sites, these suggest the presence of vegetation in the moats. The nature of this vegetation may be assessed through variations in the L-band volume scatter, usually indicative of different vegetation canopy height/density or biomass. The moat at Phum Reul, for example, has:

Cvol = -6 dB Lvol = -9.9 dB

Cdbl = -16.6 dB Ldbl = -19.3 dB

Codd = -16.1 dB Lodd = -17.0 dB

The high percentage of volume scatter at both L-band and C-band suggests woodland or forest (i.e. dense vegetation cover), with no underlying water, as there is comparatively little double-bounce. When Phum Reul is compared to the other moats by comparing the relative levels of double-bounce, volume and odd-bounce scattering in each frequency band it can be seen that the sites vary. For example, Phum Trei Nhor has:

Cvol = -4.8 dB Lvol = -9.8 dB

Cdbl = -12.3 dB Ldbl = -11.6 dB

Codd = -11.5 dB Lodd = -16.0 dB

These figures indicate, relative to Phum Reul, that Trei Nhor has a greater amount of dense vegetation over water in the moat. Both of the Hariharalaya sites, O Dek and Phum Stung, have a relatively high Lodd (~-13 db) and low Lvol (-16.6 dB) and moderate Lvol (-12.2 dB) respectively. Overall, this suggests terra firme , earthworks, with low vegetation cover and moderate vegetation cover.

Some sites have a greater (brighter) Ldbl than (Lvol - 2) and less (darker) Lodd than Lvol or Ldbl, which due to the high Ldbl, indicates strongly the presence of water underneath the vegetation canopy. This is seen at Lovea, Trei Nhor, Mokak, Ta Saom, Tonle Sa, Chakrey, and Pongro.

Of all the sites, based on Lvol scatter, Phum Reul and Trei Nhor have the highest levels of vegetation relative to the other sites. The figures from Nokor Pheas suggest a moderate level of vegetation and perhaps a mixture of terra firme and inundated vegetation.

Conclusion

In summary, the SAR data from SIR-c seems to indicate the presence of water underneath a vegetation canopy, which may be remnants of earlier moats at the sites under study. Also in some caes, the radar data indicates the presence of terra firme, which is likely to take the form of (possible prehistoric) dikes at these sites. These results await verification by field or aerial surveys.

Analysis of the many archaeological features of the Angkor site require a repertoire of techniques which thrive on rather than shy away from the immense quantity of data obtained by SIR-C. ⁸ At the same time, when faced with the huge volume of data, it is difficult to know which variations will provide the most useful insights. Even when single attribute variations are incorporated into correlations, and those into a matrix of correlations, the variables remain too numerous to provide a simple interpretation of their relationship. Another approach to the data is through modelling, such as the example employed here, which fits three scattering mechanisms to polarimetric SAR observations.

The focus is on water management features of the prehistoric landscape where images processed from the SAR data depart unequivocally from optical images. The Angkor mound sites fit this criteria. The natural and manmade features at Angkor which are relevant to the ancient city fully exploit the sensitivity of polarimetric SAR radar, not only to surface and subsurface moisture but all contributors to the biomass. The formation of mounds, diversion of river courses; the siting of temples, moats, dikes, and barays all contributed to the present vegetation and hydrology of Angkor. It has not been possible, however, to generate a developmental or conceptual paradigm to explain fundamental aspects of Angkor such as the conservation and control of water. There is no generally accepted geographical and hydrological history of the city, despite years of excellent scholarship investigating the urban zone and its remains. Yet it is commonly agreed that there exists an intimate relationship between those remains and elements such as vegetation and moisture. In the use of microwave remote sensing to study this relationship, new understanding is brought to the critical transition from village to city through exploitation of water resources to alter the terrain.The succession of royal structures at Angkor are often presented as the height of Khmer culture and innovation. However, in the earlier transition from village to city change was equally if not more radical.

Acknowledgements

Many people and organizations have made this research possible. I would particularly like to express my gratitude to the Jet Propulsion Laboratory, APSARA [Authorite pour la Protection du Site et L'Amémangement de la Région d'Angkor] (Cambodia), the Royal Angkor Foundation (Budapest), the World Monuents Fund (New York), the School of Oriental and African Studies, and the British Academy.