In 2012, the Khmer Archaeological Lidar Consortium (KALC) acquired lidar data over 370 km2 of the Greater Angkor Region. These data revealed details of the urban form in the CCC and high-resolution (mm vertical resolution) topographical information about the region (Evans, et al., 2013). Combining this lidar imagery and other forms of remote sensing, including satellite imagery, researchers from the Greater Angkor Project have identified over 25,000 archaeological features in the Greater Angkor Region, including 1,000 temples, 9,000 reservoirs, and 9,000 mounded areas used for occupation. In addition, the temples and occupation mounds have been ground verified, and the reservoirs are very clearly identifiable in the remote sensing imagery. Figures 2 and 3 illustrate these features in the context of the Greater Angkor Region.
The basic unit of agricultural production in the AMA was the community temple, which many scholars have argued were centers of activity for medieval Khmer farming communities (Hall, 2011; Lustig & Lustig, 2019; L. A. Sedov, 1963; L.A. Sedov, 1967). These temples consist of an elevated rectilinear area of civic-ceremonial buildings, often surrounded by a moat, and associated residential areas and hydraulic farming infrastructure (Evans, et al., 2007). Researchers have been able to associate reservoirs and embankments with specific temples based on orientation and proximity (Klassen, 2018; Klassen, et al., 2021). In total, 46% of reservoirs (99% of reservoir water) and 92% of moats (28% of moat water) have been associated with specific temples in the AMA (see Table 3).
Recent advancements in statistical methods, including machine learning, have allowed researchers to date community temples and their associated moats and reservoirs with an average absolute error of 49–66 years (Klassen et al., 2018). These data were then combined with other archaeological data, association algorithms, art historical analysis, and radiocarbon dates to predict dates for features associated with temple communities in the AMA and archaeological features in the CCC. In total, researchers have been able to predict dates for over 20,000 of the 25,000 archaeological features that have been mapped in the Greater Angkor Region (Klassen et al., 2021). When combined with insights from inscriptions (Lustig & Lustig, 2019), this has led to a reasonably well-understood history of the agricultural landscape over time (Klassen & Evans, 2020). We carried forward these chronological assessments of community temples in this analysis. We acknowledge that the dating is subject to revision as more data are collected; however, we believe the level of chronological control we have achieved is sufficient for the present study.
For this analysis, we exclude features from the CCC and focus on agricultural temple communities from the surrounding AMA. To determine which reservoirs, moats, and temples are in the AMA, we used the Clip tool in ArcGIS 10.7.1 to remove features within CCC zones as defined by Klassen et al. (2021) from consideration. For total feature counts in the AMA and CCC, see Table 2.
Estimating Temple Platform, Reservoir, and Moat Volumes
The digital terrain model (DTM) derived from the lidar data allows us to estimate the volume of water in reservoirs and moats and the volume of earth moved to create temple platforms and reservoir embankments for those features with lidar coverage. We calculated the depth and height of these features using the Zonal Statistics tool in ArcGIS 10.7.1 to the nearest cm. Then, we used the Range function to calculate the maximum value from the DTM raster minus the minimum value of the DTM raster inside of each moat, reservoir, mound, and temple polygon. We argue that the maximum and minimum values represent the original form of the features before they experienced erosion over the last 1000 years.
Once we calculated the depths of features within the lidar coverage, we used ordinary least-squares regression to measure the direction and strength of any relationship between feature depth and surface area. Temples with stone architecture (with heights > 2 m) were excluded from the regressions (Table 1). We then applied this equation to estimate depths (and thus volumes) based on the surface area of features within the Greater Angkor Region but not included in the 370 km2 zone of lidar coverage (Table 2).
Table.1 Depth regressions for features in the CCC and the AMA Table.2 Features in the CCC and the AMA Estimating Labor Inputs
Labor inputs are based on the estimated populations of each temple community from Klassen et al. (2021). These estimates are based on the method developed by Hanson and Ortman (2017), which uses settlement scaling theory to translate the areas of more than 9,000 occupational mounds to estimates of their resident populations. Importantly, this method assumes that larger occupational mounds are more densely inhabited on average. These models were calibrated using ethnographic information concerning the typical population size of Khmer farming communities. We set the minimum population size of an occupational mound to one household of five persons. Each mound was then associated with the closest temple, and estimated populations of those mounds were summed to estimate that temple community’s population. For temples that no longer had surviving occupation mounds, we assigned an estimate of 497 people per community based on our indexing of mound areas to ethnographic information concerning the typical population size of Khmer farming communities (Delvert, 1961). Since an unknown fraction of these mounds has been destroyed for subsequent (post-Angkor) agricultural purposes and development, the preserved occupational mounds reflected in Table 3 provide a minimal estimate of food-producing temple community populations.
Table.3 Reservoirs and moats that were grouped with temples in the Angkor Metropolitan Area (AMA) We recognize that farmers would have also contributed to other ritual, political, and economic activities. However, we argue that the amount of effort each farmer put toward agricultural production was independent of scale, so comparing outputs by scale capture the effects of scale irrespective of the other activities of these individuals.
Estimating Land Inputs
The Greater Angkor Region is often defined as the 3000 km2 regional watershed. This area included large undeveloped patches in the North between Angkor and the Kulen mountains. Of the 3000 km2, Figs. 2 and 3 illustrate that approximately 1500 km2 of the land in the AMA was developed by community temples (Evans, 2007; Pottier, 1999), and over 1000 km2 of Angkorian period ricefields have been mapped over that 1500 km2 area (Hawken, 2012). This finding suggests it is reasonable to estimate the associated land in production for each temple community based on distances to adjacent temples during the period the temple was constructed and in subsequent periods as additional temple communities filled in the landscape. The land included in this analysis includes space for occupation, ricefields, and reservoirs. There does not appear to be a stark contrast between agriculturally productive and non-productive spaces as occupation mounds often included gardens (Castillo, et al., 2020), reservoirs were a source of fish protein, and inscriptions provide provisions to allow pigs to forage in ricefields (Jacob, 1993), all of which is included in our calculations of productivity.
Construction periods for agricultural temples are from Klassen et al., (2018), as discussed above. We have used these dates to assign temples to each of the six periods, ranging from 37 to 136 years. However, we acknowledge that there is likely to be some error in the period assignment of individual temples due to the absolute average error of 49–66 years in the dating method itself.
To calculate the amount of land available to each temple community, we estimated the transverse dimension of each community’s territory as the mean of the distance from each of the 921 dated temples in the dataset to its three nearest neighbors, based on their center-point locations. We squared this distance to estimate the total hectares of land within this territory. This method assumes temple community territories were generally rectilinear in outline. Still, it does not presume any specific geometry or even that the associated territory was centered on the temple. This method also avoids some of the boundary issues associated with alternative methods of defining site catchments, such as buffers or Thiessen polygons. It allows temple territories to vary continuously based on distances to neighboring temples.
Historical records indicate that land-associated community temples, and presumably their associated infrastructure, were often purchased and combined with pre-existing temples or used toward the foundation of new temple communities (Jacob, 1993; Lustig & Lustig, 2019). Although we cannot map these purchases, approximately 10% of new temples constructed during each period are within 250 m of another temple, based on the coordinates of their center points. Because temples and their associated moats were typically 40–80 m across, this implies a gap of less than about 150 m, which seems too close to reflect adjacent independent communities. We, therefore, consider such cases to represent the purchase of an existing temple by a new owner. To take this into account, we aggregated the data (including populations, territories, reservoir volumes, and total productivities) for all occupied temples within 250 m of one another during a given period, assigned these data to the most recently constructed temple in that group, and recalculated the community center location as the average of the contributing temple locations.
While we do not have comprehensive datasets that can be used to definitively model the terminus of occupation across our study region, emerging evidence suggests community temples and their associated infrastructure remained in use from the time of their dedication through the end of our study period in the fourteenth century CE, which is often referred to as the “collapse” of the Angkor. For example, in an extensive survey across the Greater Angkor Region, researchers have identified post-Angkorian ceramic scatters on occupation and temple mounds that indicate continued occupation through the post-Angkorian Period (Brotherson, 2019), while surveys within Siem Reap town that suggest the temples were used until at least the seventeenth century CE (Vitou, 2012). Similarly, Heng and Stark excavated three temple communities (Wat Athvea Area, Wat Chedei Area, Wat Prei Phdau Area) and found that each site had a long and complex occupation history with ceramics that date from the foundation of the temples until the 15th–twentieth centuries CE (Heng & Stark, 2020). Ultimately, any fallibilities with the assumption that all temple communities are occupied after their foundation to the end of the study period are not expected to impact our results because proportionate changes in the parameters of the model would not change any observed scaling patterns. This means that if only a percentage of the temple communities or occupation mounds are active during a given period, the scaling relationships should remain the same as there are no adjustments in parameters that co-vary with the total area of occupation mounds.
We estimated temple territories separately for each period based on all temples that had been constructed by that time unless a temple’s land and infrastructure met our criteria for an inferred purchase. Thus, for Period 1, territories are based only on temples constructed during that period, but Period 2 territories are based on temples constructed during Period 1 or 2, Period 3 territories are based on temples constructed during Periods 1–3, and so forth. The distribution of territories for all temples occupied during each period is presented in Fig. 5D. This chart shows that median temple territories fell from more than 500 hectares during Period 1 to about 200 hectares by Period 3 and a mere 140 hectares by Period 6. Also, the cumulative distributions show that new temple territories became less variable in size over time. Both are strong indications of increased packing of farming communities on the landscape, suggesting agricultural intensification.
Estimating Irrigation Inputs
Agricultural intensification involves producing more rice per hectare per year. One of the ways to do this is to increase irrigation inputs. This can be estimated for temple communities based on the volume of water that could be stored in associated reservoirs and moats. In addition to moats surrounding temples, more than 3,000 reservoirs, which were generally constructed above grade, have been mapped for the Greater Angkor Region. During medieval times, water stored in these features was collected from precipitation, groundwater, local runoff, or state hydraulic infrastructure. Today, these ponds, called trapeang, are still used for crops and vegetables cultivated on their edges, but their use for rice cultivation is more limited (Pottier, 1999, 125–133). During medieval times, these reservoirs likely had several uses, including fish production, drinking water for animals, and as a source of irrigation water to flood rice paddies during the dry season (thus increasing the number of crops).
We summed the estimated volumes of all reservoirs and moats associated with each temple based on their proximity and orientation (Klassen, 2018; Klassen, et al., 2021). This measure can be associated with a temple during its construction period, assuming that reservoirs were built based on conditions at that time, or during the final period, under the assumption that reservoirs reflect the final distribution of temples and their associated lands. This volume can be considered a measure of the irrigation inputs to the land associated with a temple community. It can also be considered a measure of the infrastructure for agricultural production utilized by the associated community. At this stage, it is not possible to determine whether reservoirs were added to temple infrastructure as their territories shrank over time or were part of the initial development of a temple community. A program of OSL dating of reservoir embankments might help to answer this question.
Estimating Agricultural Outputs
It is typical in economics to estimate the output of a social group as a product of multiple inputs (Robinson, 1953). In traditional agriculture, these inputs are land, labor, and technology, which includes farming equipment and inputs such as fertilizer, insecticide, and irrigation water (Hunt, 2000). Technology can be interpreted broadly to include know-how. Under ideal circumstances, we would have independent estimates of total outputs in addition to land, labor, and technology for a subset of temple communities. In the absence of such information, we multiply our measures of farming labor, agricultural land, and irrigation inputs to create a measure of temple agricultural output. We consider this three-factor measure to be more realistic than a single-factor measure, such as land by itself, due to the opportunities for multi-cropping provided by hydraulic infrastructure. If two of the three factors were held constant (land and water), one would expect the third factor (labor) to exhibit decreasing returns, but in a context where all three are free to vary, one would expect their product to be proportional to total production and to potentially exhibit increasing returns.
Because agricultural production can generally be viewed as resulting from the strong interaction of several factors of production (i.e., land, labor, and other inputs), one would expect the probability distribution of a multiplicative production index to be log-normal (Aitchinson & Brown, 1957; Limpert, et al., 2001). Figure 4 presents QQ plots for the distributions of log-transformed production indices across temple communities during each period, assuming cumulative occupation. These plots show that these distributions all conform to the expected range of a log-normal distribution except for a few outliers in the upper tails.
Relationship Between Community Scale and Productivity
As discussed above, rice paddy production in South and Southeast Asia was often organized at the local community scale, with groups of laborers coordinating agricultural activities. Under such conditions, one would expect the group’s productivity to increase proportionately with the interactions between workers per unit time and each worker’s effort. This is typical of non-agricultural socioeconomic outputs (Bettencourt, 2013; Bettencourt, et al., 2007; Glaeser & Gottlieb, 2009), but to our knowledge, this relationship has not been considered in previous studies of agricultural intensification. Furthermore, an interaction can involve exchanging physical things or information from one person to another during production. In addition to saving time, one would expect these interactions to improve each worker’s skill through emulation of best practices and social learning (Boyd, et al., 2011). At Angkor, these varied social interactions would have occurred within the developed network of physical spaces that include temples, rice paddies, ditches, and reservoirs.
A foundational assumption in the study of socioeconomic production embedded in physical space is that the benefits resulting from social interactions are balanced against the costs of the movement necessary for these interactions to occur (Isard, 1956). An approach to human networks known as settlement scaling theory builds from this principle to generate expectations for the effect of population size for interaction rates and outcomes. Importantly, agents do not seek to maximize outputs or minimize costs in this framework but simply seek a balance between movement costs and interaction benefits, a condition known as a spatial equilibrium. Under these conditions, the number of interactions per unit time is given by \(I\sim {N}^{\beta }\), where \(N\) is the number of interacting individuals and \(\beta =7/6\) (Bettencourt, 2013, 2014; Lobo, et al., 2020; Ortman & Lobo, 2020; Ortman, et al., 2015). Consequently, interactions increase faster than population and exhibit increasing returns to population scale. If one then considers that there is an average outcome of an interaction, across all types that occur during a production process, then the total agricultural production of a group \(Y\) should increase according to \(Y={Y}_{0}{N}^{\beta }\), where \({Y}_{0}\) is the baseline productivity of a farmer given the available technology and \(\beta \cong 7/6\). This relationship between population and output has been observed archaeologically for civic architecture construction rates in Mesoamerica and South America (Ortman, et al., 2020). Here, we examine whether this relationship also applies to communal wet paddy rice production at Angkor.
The value of the scaling coefficient \(\beta\), which determines the effect on the output of population size, can be estimated through ordinary least-squares regression of log-transformed estimates of population and productivity for a group of cases. This is feasible because \(Y={Y}_{0}{N}^{\beta }\) and \(\mathrm{log}Y=\mathrm{log}{Y}_{0}+\beta *\mathrm{log}N\) are equivalent expressions, and the latter is a simple linear function.
Estimating Net Proceeds from Agriculture
One would expect much of the agricultural output of a temple community to have been used to provide subsistence for the people of that community. Inscriptions indicate that a portion of any remaining surplus would also have been transferred to state control in the form of taxation or payment in kind (Lustig, 2009). After that, much of the remaining surplus would have been devoted to construction and maintenance of the community temple and the activities that took place there. We thus consider the total volume of the temple platform(s) in a community as a proxy for the average net proceeds from agriculture in that community, after taking subsistence and transfers into account. These net proceeds were likely only a small percentage of total temple production, and we test this hypothesis as a cross-check on our output measure and as a means of gauging the overall economic fortunes of temple communities. We use the relationship between agricultural outputs and associated temple platform volumes to estimate the average profitability of agriculture for temple communities, with the slope of the best-fit line relating these two log-transformed measures reflects the average percent profit (see “Results,” below).
Identifying Spatial Patterns in Land Use
Given the extensive area of Angkor and the fact that food surpluses ultimately had to reach the non-food-producing population in the urban core, transport costs were a potentially important factor in deciding where and how intensely to engage in agricultural activities within the settlement area. This specific decision can be further conceived as part of a higher-level decision regarding the type of use to which a given plot of land was put (agriculture, housing, a temple, transport or hydraulic infrastructure, etc.). In urban economics, the primary model of land use within cities is the Alonso-Muth-Mills (AMM) model (Alonso, 1964; Brueckner, 1987; O’Sullivan, 2011:125–200), itself an elaboration of Von Thunen’s analysis of agricultural production surrounding a central market (von Thünen, 1966). This model hypothesizes that the increase in transport costs with distance from a central market is counterbalanced by a decrease in land values (land rents in the language of economics).
As with SST, the AMM model builds from the assumption of a balancing of spatial costs and benefits, thus implying that increasing desirability of a location is counteracted by increasing cost (Alonso, 1964; Brueckner, 1987). Under these conditions, moving a residence toward or away from a central work area will change the commuting cost by the change in distance \(\Delta x\) times the transport costs per unit distance \(\varepsilon\). This change will be balanced by a corresponding change in the value of housing per unit area \(\Delta p\) times the area of the house \(a\) (O’Sullivan, 2011:141). This can be expressed as:
$$\Delta p*a=-\Delta x*\varepsilon .$$
(1)
We adapt this relationship to the agricultural landscape of Angkor by viewing \(p\) as the value of agricultural land, which we estimate as to its productivity per hectare, and \(a\) as the area cultivated by an individual farmer.
Notice that the left side of Eq. (1) can also be expressed as \(\Delta a*p\), leading to a system of equations:
$$\frac{\Delta p}{\Delta x}=-\frac{\varepsilon }{a},\mathrm{ and }\frac{\Delta a}{\Delta x}=-\frac{\varepsilon }{p},$$
(2)
where the left sides of both equations represent the slope of the relationship between land productivity and distance, and the relationship between area per farmer and distance, respectively. Since transport costs \(\varepsilon\) are constant, one can solve 2 for \(\varepsilon\) and equate them:
$$-a\left(\Delta p/\Delta x\right)=-p\left(\Delta a/\Delta x\right),$$
(3)
leading to:
$$a=\left[\frac{\Delta a/\Delta x}{\Delta p/\Delta x}\right]*p,$$
(4)
where the term in brackets is itself the slope of the relationship between area per farmer \(a\) and productivity per hectare \(p\). Thus,
$$\left(\Delta a/\Delta p\right)=\left[\frac{\left(\Delta a/\Delta x\right)}{\left(\Delta p/\Delta x\right)}\right].$$
(5)
Equation (5) can be empirically examined using a regression framework. We use the data for temple communities during Period 6, estimating \(a\) as temple territory / temple population, \(p\) as temple productivity / temple territory, and \(x\) as the straight-line distance of that temple from the Bayon temple in the center of Angkor Thom.