A predictive model of archaeological potential: An example from northwestern Belize

Applied Geography 29 (2009) 542–555

Contents lists available at ScienceDirect

Applied Geography journal homepage: www.elsevier.com/locate/apgeog

A predictive model of archaeological potential: An example from northwestern Belize Sallie Vaughn a, *, Tom Crawford b a b

Department of Geography, East Carolina University, NC 27858, USA Brewster A-234, Department of Geography, East Carolina University, Greenville, NC 27858, USA

a b s t r a c t Keywords: Binary Logistic Regression Remote sensing Landscape archaeology Maya Belize

Binary Logistic Regression is used to identify areas of high archaeological potential in a portion of northwestern Belize. The predictive modeling process involves remotely sensed imagery, Geographic Information System (GIS) data and techniques, and multivariate statistical approaches. Predictive variables represent both the pre-historic current landscape of the ancient Maya and the present day physical landscape. An optimal predictive model obtained using logistic regression includes one variable derived from a Landsat image representing contemporary vegetation patterns associated with Maya settlement and two variables derived from a digital elevation model (DEM) and an analog hydrography map representing resource endowments relevant to the ancient Maya. The predictive model identifies several areas of high archaeological probability as well as areas that are unlikely to contain any archaeological remains. Results can be used to inform future field surveys in a more cost efficient manner. Prior research has utilized remote sensing and GIS approaches for Maya site identification in the southern lowlands region of the Mexican Yucatan peninsula and the northern lowlands of Peten, Guatemala. This research represents the first landscape archaeological approach using satellite imagery for the Maya region in northwestern Belize. Ó 2009 Published by Elsevier Ltd.

Introduction Researchers discover archaeological sites through systematic field survey of the landscape – a costly, time consuming, and labor intensive process. Geospatial techniques that integrate Geographic Information Systems (GIS), remote sensing, and predictive modeling can enhance and expedite the process of site discovery by identifying areas of high archaeological potential for subsequent field investigation. By maximizing efforts in high probability areas, time, energy, and money generally used for surveying areas with little or no archaeological wealth can be reallocated to mapping and excavation efforts in higher probability areas. This project uses GIS in conjunction with remotely sensed imagery, paper map data, and Binary Logistic Regression to predict the probability of ancient Maya archaeological site presence in northwestern Belize.

Study area and background The ancient Maya were a complex civilization occupying present day southeastern Mexico, Belize, Guatemala, and portions of Honduras and El Salvador (Fig. 1). The Maya region of Mesoamerica is broken into two basic culture areas based on

* Corresponding author at permanent address: 2915 Chapel Hill Rd., Durham, NC 27707, USA. Tel.: þ1 252 412 4474; fax: þ1 919 461 1215. E-mail addresses: [email protected] (S. Vaughn), [email protected] (T. Crawford). 0143-6228/$ – see front matter Ó 2009 Published by Elsevier Ltd. doi:10.1016/j.apgeog.2009.01.001

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

543

Fig. 1. The Ancient Maya world, its sub-regions, and well-known sites.

topography: the highlands and the lowlands. Northern and Southern sub-regions within the lowlands are differentiated on the basis of Maya cultural elements in those areas (Coe, 2005). Maya civilization began in the lowlands near the present day borders of northern Guatemala, Mexico, and Belize and then spread to other regions in Mesoamerica (Coe, 2005; Demarest, 2004). Hallmarks associated with Maya civilization such as hieroglyphic writing, the calendrical system, massive stone cities, and the corbel arch originate from this centrally located portion of the lowlands (Coe, 2005: 24). A portion of this region in present day northwestern Belize, also known as the central Maya lowlands, serves as the focus of study for this research. The Maya are known for their large stone cities such as Chichen Itza, Copa´n, and Tikal. By 300 BC., they had established large, complex trade networks, a hierarchical social system, and built numerous settlements of varying sizes throughout the region (Coe, 2005; Demarest, 2004). Using evidence from the complex social, economic, political, and religious hierarchies, a ranking system can be applied to the settlements of the Maya. Many researchers recognize a four-tiered ranking system (Adams & Jones, 1981; Folan, Faust, Lutz, & Gunn, 2000). Other ranking systems reveal as few as three and as many as nine tiers within the settlement hierarchy (Adams & Jones, 1981). Evidence of a hierarchical settlement ranking system may be observed today by the number and function of structures, cultural remains such as pottery and stone tools, food preparation items such as grinding stones, and other cultural remnants such as large carved stone pillars known as stelae, and painted stucco masks that adorned buildings. Higher ranking settlements were sizable, had a large number of multi-purpose structures, and functioned as religious, economic, social, and political centers. Smaller, lower-ranking settlements, had fewer structures, most of which were dedicated to residential purposes, and typically contained resource specific artifacts indicating the particular economic activity of the settlement, such as stone tool production, pottery making, or agriculture (Coe, 2005; Demarest, 2004). A common feature of Maya settlements, regardless of size, is an emphasis on enclosed interior courtyards or plazas surrounded on all four sides by structures which may serve as residential structures or civic and ceremonial structures (Adams & Jones, 1981). The courtyard floor was often covered with locally made limestone plaster (Demarest, 2004; Saturno, Sever, Irwin, Howell, & Garrison, 2007). Typically, the four structures surrounding the courtyard were oriented towards the cardinal directions with a specific emphasis on the structure built on the eastern side of the complex (Demarest, 2004; Ricketson, 1933). Quite often, the eastern structure in a courtyard complex is the tallest and most ornate due to an associate between the east, sunrise, and re-birth. These structures likely served a religious function (Adams & Jones, 1981; Demarest, 2004).

544

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

The Maya culture progressed, becoming increasingly complex as populations grew. In the central lowlands, populations peaked between 600 and 900 AD., during what is known as the Late Classic period. This pinnacle in population was accompanied by an increase in construction of small residential settlements and the expansion of large civic centers. The expansion of settlements across the landscape is still visible today in the form of mound complexes, accumulations of cut stone and plaster, monumental cities, and concentrations of artifacts just below the surface of the soil. Due to their visibility across the landscape, the majority of well-known and well-studied archaeological sites date to the Late Classic period. This research, and the predictive model generated from it, also focuses on the Late Classic period. Between 800 and 925 AD., depending on the specific location, the Maya civilization is said to have collapsed. Most, scholars agree that the Maya civilization experienced a dramatic reduction in population, social reorganization, and settlement abandonment (Coe, 2005). Interpretations of this collapse are highly varied and debated but usually allude to human ecological factors or environmental stressors such as drought, overexploitation of natural resources, soil runoff, or volcanic eruption (Coe, 2005; Demarest, 2004; Demarest, Rice, & Rice, 2004; Dunning, Jones, Beach, & Luzzadder-Beach, 2003). Through continued discovery and investigation of the settlements and physical materials left behind by the Maya, researchers can further examine and ultimately explain their cultural evolution and eventual collapse.

Study area and background The study area is located almost wholly on lands owned by the Programme for Belize (PfB) in the southern Maya lowlands (Fig. 2). The PfB owns and manages over 250,000 acres of jungle, swamp, and savannah in the northwestern section of Belize. The main objective of the PfB is to preserve the natural landscape for ecological and archaeological research. A 100,000 acre section of their property, including a one kilometer buffer zone, was selected for analysis due to its relatively pristine status with low levels of present day human developments. Three areas, two permanently inundated swamps and one cattle pasture, were omitted from the study area due to their exceptional vegetational characteristics with respect to the rest of the study area. Geologically, the study area is underlain by a rolling karstic plain bisected by three steep escarpments and several streams. Numerous physical features associated with a karst environment such as sink holes, caves, and steep hillocks are present. Limestone is not conducive to soil generation or the formation of streams, rivers, ponds, valleys and other fluvial features associated with many non-karst landscapes that experience significant human settlement. Tectonic activity has impacted the study area resulting in three major southwest to northeast trending escarpments. Each escarpment represents a change in elevation of about 60 meters, and the only permanent streams and rivers in the area run along the bases of these escarpments. Topography, soil characteristics, mineral content, and vegetation both above and below an escarpment are significantly different (Dunning et al., 2003; Ford & Fedick, 1988). Large rock outcrops, which are common in areas of high relief such as escarpments, are generally more weathered and chemically eroded than areas of little topographic relief (Dunning et al., 2003). Typically, the steep, hilly slopes of escarpments have thin soils only 10–15 cm deep. Areas with less topographic variation allow the accumulation of soils up to half a meter deep (Ford & Fedick, 1988). In turn, the differences in soil characteristics encourage the growth of different species and density of vegetation (Dunning et al., 2003). These variations were essential to the Maya who took advantage of the many microenvironments of the region (Demarest, 2004; Scarborough & Valdez, 2003). Maya habitation has been documented within the study area as early as 1200 B.C. with highest populations occurring during the Late Classic period (600–900 A.D.) (Coe, 2005; Demarest, 2004). Over 100 ancient Maya settlements known to be inhabited during the Late Classic period have been identified by prior surveys. However, the exact physical locations of only 69 of these settlements have been recorded and were available at the time of this study. Locational information about these sites was gathered by the lead author from numerous journal articles, field reports, and dissertations. Information regarding the physical size and likely sociopolitical function was also recorded where available. Over a total of four months, the lead author participated in archaeological excavation and survey within the study area. In addition to this field work, the lead author also visited numerous sites of varying sizes within the study area and the surrounding Central Maya Lowlands. Insight gained during these experiences relating to the sociopolitical and economic functions of Maya sites was used to enhance published site ranking systems and validate the application of such systems to the study area. For the purposes of this study, the term ‘‘site’’ minimally refers to any structure, or group of structures, constructed by the Maya for year-round residential purposes. The term does not encompass field huts, resource extraction sites, and isolated artifact finds. Of the 69 archaeological sites included in this study, 54 are classified as these small villages or hamlets. These sites are very numerous and likely consisted of extended families who participated in agriculture and the production of household items such as baskets, pottery, and stone tools. At their largest, sites consist of hundreds of economic, religious, vocational, social, and/or political structures in addition to residences. In total, three archaeological sites of this magnitude have been discovered within the study area. It is likely that these sites served as regional centers of commerce, religion, and government. Due to their relative size and magnitude, these sites are less frequent in the sociopolitical and physical landscapes. The remaining 12 sites were assigned a rank between that of hamlet and regional center. The physical locations of these archaeological sites and their associated environmental characteristics provide the basis for predictive modeling. Sociopolitical characteristics provide a framework for ranking these sites according to Christaller’s marketing principles and central place theory.

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

545

Fig. 2. Study Area, northwestern Belize.

Walter Christaller introduced the concept of central place theory to explain the hierarchical spatial ordering of settlements in southern Germany (Baskin, 1966). He believed the spatial distribution of settlements to be a function of local and regional economics and politics. Specifically, settlement size and market function dictate the spatial and hierarchical relationships between settlements in a region. Using these concepts of market size and the exchange of goods, four categories of settlement have been identified for this portion of the Maya region (Garza & Kurjack, 1980) (Table 1). Typically, the largest and most magnificent of archaeological sites are the only ones earmarked for study and preservation. Smaller, seemingly less significant sites, have often been regarded as unimportant or not worthy of investigation, let alone preservation. While many sites in the Maya region have been set aside as archaeological parks, a focus on individual islands of preservation does not allow a more complete regional view of the archaeological record. From the point of view of a single site, the regional and social context of the ancient landscape is ignored. This top-down, individual site-based perspective of Maya settlement pattern and sociopolitical organization creates bias within the study of their civilization. A shift from the site-based point of view to a more regional landscape perspective is necessary to better understand ancient Maya settlement patterns which may provide clues to Maya sociopolitical organization. Regional predictive models of ancient Maya archaeological site presence can help researchers more fully understand the ancient Maya landscape and sites of various sizes; not just the largest and most frequently documented.

546

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

Table 1 Site ranking system within the Maya region. Rank

Type

Function

Size

Frequency

1 2 3 4

Civic Center Regional Center Local Center Village, hamlet, town

Residential, social, administrative, religious, and commercial Residential, social, and commercial Residential and commercial Residential

Large

Few

Very small

Numerous

Other researchers have begun to adopt a regional landscape perspective using the term ‘‘landscape archaeology’’ (Anschuetz, Wilshusen, & Scheik, 2001; Ashmore & Knapp, 1999; Sever & Irwin, 2003). Landscape archaeology focuses on the relationships between humans and the natural and built environments. Remote sensing data and techniques are especially germane to a landscape archaeology approach towards site prediction due to the broader synoptic spatial scale associated with typical imagery footprints that are not possible via intensive ground-based surveys (Saturno et al., 2007). Such data and techniques allow researchers to discover features that may go unnoticed to researchers on the ground from a vantage point provided by airborne and spaceborne platforms. Prior remote sensing research of the ancient Maya has used a variety of sensors to investigate sites in Peten region – a region located directly west of the present study area across an international border in neighboring Guatemala. Investigators have used Landsat Thematic Mapper (TM), Landsat Enhanced Thematic Mapper (ETM), IKONOS, and Quickbird satellite imagery along with STAR-3i and AIRSAR radar airborne imagery to successfully identify settlement sites, roadways, canals, water reservoirs, and bajos – the seasonally inundated swamps whose use by the ancient Maya for agriculture has been a topic of hot debate (Saturno et al., 2007; Sever, 1998; Sever & Irwin, 2003). Recently, high resolution imagery originating from IKONOS satellite imagery has been used to map vegetation signatures in the Peten region (Saturno et al., 2007). Regions with particularly high vegetation signatures were visited by survey crews to determine if there is a link between vegetation signatures and ancient Maya settlements. It was determined that the vegetation signatures detected by high resolution IKONOS imagery is an excellent indicator of archaeological settlements. Researchers believe that high concentrations of limestone plaster and limestone blocks common to ancient Maya structures and courtyards have affected the vegetational patterns in areas once inhabited by the Maya (Saturno et al., 2007). Vegetational characteristics due to ancient landscape modifications can be mapped to guide researchers in locating previously unrecorded archaeological sites. Unfortunately, high resolution imagery such as IKONOS imagery is not readily available for the entire Maya region and the cost is prohibitive to most archaeologists. Other research has used Landsat ETM and airborne IFSAR sensors to investigate settlement pattern relationships to environmental resources in a region surrounding Holmul, Guatemala (Estrada-Belli & Koch, 2007). A study by Pope and Dahlin (1989) in the lowlands region of the Mexican Yucatan peninsula, located north of our study area, used radar imagery and Landsat Thematic Mapper (TM) to examine large canals associated with ancient Maya wetland agriculture. These examples, and the broader adoption of remote sensing within the archaeology community (Wiseman & El-Baz, 2007), point to the utility of a landscape archaeology paradigm. While other research has investigated the ancient Maya in northwestern Belize (Ford & Fedick, 1988; Green, 1973; Scarborough & Valdez, 2003), our research represents the first remote sensing oriented study of settlement patterns in this region. Also, this research represents an effort to uncover a relationship between vegetational patterns captured with freely available, low resolution satellite imagery. Data and methods Fifty of the 69 archaeological sites with known geographic locations along with 50 random locations were chosen to supply a set of 100 point locations within the study area. These points representing both known archaeological sites and locations of unknown archaeological state were established to serve as the dependent variable (STATE). Predictive variables used in this research focus on the relationship between the ancient Maya people and the environmental resource endowments they were likely to utilize in order to establish and maintain viable settlements. Present day topographic and landscape characteristics such as soils, slope, and distance to water are assumed to be similar to ancient conditions and serve as predictive variables representing environmental endowments. Additionally, remote sensing vegetation indices were used as signals of land cover conditions in the late 20th century which may be indicative of ancient landscape modifications and help to predict settlement locations. Binary Logistic Regression (BLR) identifies predictive variables that correlate with site presence or absence. Once predictive correlations have been defined, they can be projected into areas with no existing site presence/absence data to indicate areas of archaeological potential. Dependent variable The dependent variable (STATE) is a dichotomous variable recording site presence or absence. Of the 69 known archaeological sites dating to the Late Classic period with locational information, a subset of 50 was randomly chosen. The remainder of known sites (n ¼ 19) will be used to test the predictive model. Then, 50 random locations were generated which have an unknown archaeological site status. It is assumed that these locations do not contain archaeological sites. The fifty known

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

547

Table 2 Variable names and descriptions. Variable name

Scale

Description

STATE (dependent) EAST SOUTH DIST_WATER DIST_AG_LAND AG_LAND GREEN LAND_SYSTEM NDVI SLOPE WETNESS

Nominal Nominal Interval Ratio Ratio Ratio Interval Nominal Interval Ratio Interval

Binary variable where 1 ¼ archaeological site present and 0 ¼ archaeological site absent Binary variable where 1 ¼ a slope which faces northeast, due east, or southeast and 0 represents all other directions Difference in degrees from due south; ranges from 0 to 180 Distance in meters to a water feature Distance in meters to a flat area of at least 90 square meters Sum, in hectares, of flat area within 450 meter radius of a location Tasseled Cap greenness index standardized such that all values fall between 0 and 100 Three categories of land systems based on usefulness for agriculture; ranges from 0 to 2 Normalized Difference Vegetation Index standardized such that all values fall between 0 and 100 Integer values for slope in degrees; ranges from 0 to 30 Tasseled Cap wetness index standardized such that all values fall between 0 and 100

sites (STATE ¼ 1) and 50 randomly generated sites (STATE ¼ 0) are combined and used as training points within the logistic regression method. For each of the 100 training points, measurements of the candidate predictive variables are acquired for inclusion in the database. Predictive variables Variables were derived from remotely sensed imagery including elevation data from the Shuttle Radar Topography Mission (SRTM) and Landsat TM5 imagery. A cloud-free Landsat TM5 image from 27 December 1989 was acquired representing a period of transition from the wet to dry season. Other predictive variables were created from paper maps digitized by the author. Variable summaries and descriptions are provided in Table 2. Thirty meter resolution SRTM digital elevation model (DEM) covering the study area was unavailable for this research due to data access restrictions. A 90 meter DEM of the study area is available, but not used in this analysis. Derivatives of the 30 meter resolution DEM such as hillshade, aspect, and slope were available at a 30 meter resolution and acquired from the Earth Resources Observation Systems (EROS). These elevation products were used to construct several predictive variables for the model. Two aspect variables were created to determine if the Maya preferred to build in areas that face certain directions (EAST, SOUTH). Like many present day cultures, the east represented re-birth and resurrection for the Maya, and they tended to build religious structures on the eastern side of their plazas and courtyards (Demarest, 2004; Ricketson, 1933). Other cultures are known to prefer certain aspects for settlement construction due to the pursuit or avoidance of midday sunlight (Christopherson & Entz, 2001). Slope, measured in degrees, is used as predictive variable (SLOPE) and as the basis for two other variables (DIST_AG_LAND and AG_LAND). Areas with little (less than one degree) or zero slope which hold water during the rainy season may have been avoided for habitation. Conversely, it is possible that extremely steep slopes may have been avoided due to the difficulty of settlement construction and habitation (Ford & Fedick, 1992). Raw slope values are then used to identify areas with slopes between one and five degrees which have enough relief to be well-drained during rainy periods but are not so steep that soil runs off with the rainwater. Previous studies have identified regions with slope between one and five degrees as preferred by the Maya for agricultural fields and residences for farmers (Ford & Fedick, 1992). However, these areas were not preferred at the complete exclusion of regions exceeding five degrees in slope (Ford & Fedick, 1992; Kunen & Hughbanks, 2003). In fact, during the Late Classic period, population pressures may have forced the Maya to utilize regions which would otherwise be unsuitable for agriculture (Walling, 2006). The model produced here assumes that areas with slope between one and five degrees would have been perceived as suitable locations for agricultural fields and residences for farmers. Therefore, this model here emphasizes the desirability of these agriculturally preferred areas. Next, distance in meters to areas with a slope between one and five degrees was calculated (DIST_AG_LAND) with the hypothesis that shorter distances to such areas would be preferred. Additionally, the area in hectares (AG_LAND) of these preferred zones was calculated. Being an agricultural society, it is hypothesized that the Maya would have valued large amounts of arable land within a close proximity to their settlements (Ford & Fedick, 1992). Four topographic maps at a scale of 1:50,000 constructed by the Ministry of Defence, United Kingdom in 1993 were scanned and georeferenced to provide data for additional variables. Streams, rivers, and springs were digitized from these topographic maps of the study area to generate a data layer representing water features. It is assumed that the present day configuration of water features in the study area is similar to that of the Late Classic period. Distance to permanent water features was calculated with the expectation that proximity to water sources would be advantageous for site locations (DIST_WATER). Well-drained, thin soils on ground of moderate relief would be ideal for the construction of stone buildings while thicker soils would have been sought out for agricultural purposes (Ford & Fedick, 1992; Kunen & Hughbanks, 2003; Turner & Harrison, 1983). A 1:100,000 scale map of land systems generated by the Natural Resources Institute in 1991 was used to generate categories of potential for agricultural use and settlement. This map indicates soil families, general topography (such as undulating plain, swamp, low karst, steep slope, etc.), and predominant vegetation types (such as savanna, marsh, mangrove, forest). This map was scanned, georeferenced, digitized, and used to construct a land system variable with 3 categories

548

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

characterizing settlement and agricultural potential (LAND_SYSTEM). Areas that are steep (slopes in excess of 5 percent), have very poorly drained, or extremely thin soils were assigned a LAND_SYSTEM value of 0 indicating a low suitability for both agriculture and the construction of buildings. It is assumed that these areas would have been prohibitive of both agriculture and settlement due to high relief and soil moisture content during the rainy season. Areas of medium relief with moderately drained soils were deemed suitable for both agriculture and settlement were assigned a value of 1. Regions with well-drained soils of moderate relief, which would have been ideal for agriculture and settlement, were deemed highly suitable for agriculture, and were assigned a value of 2. It is assumed that the Maya would have settled in areas near agriculturally suitable areas (LAND_SYSTEM ¼ 2) which are also appropriate for the construction of buildings (Ford & Fedick, 1992). The ancient Maya removed and displaced tons of limestone bedrock to construct the majority of their buildings including temples, homes, markets, and ceremonial structures (Coe, 2005). After the collapse of their civilization, the Maya largely abandoned their settlements and the structures within them. Hundreds of years later, these unkempt structures are exceptionally suitable habitats for vigorous vegetation (Dunning et al., 2003; Furley, 1975). Both seeds and soils collect in the cracks between the stones of crumbling structures. Plazas, which are frequently surrounded by stone buildings on all sides, provide artificial depressions ideal for the collection of organic materials and rainwater. These depressions allow the rapid germination and re-growth of forest vegetation as compared to areas without landscape modifications. The result is a marked increase in vegetational density and health associated with ancient structures. Three indices that capture these vegetational characteristics were derived from the Landsat TM5 image: Normalized Difference Vegetation Index (NDVI), Tasseled Cap greenness, and wetness (Jensen, 2006). Both NDVI and greenness measure vegetational characteristics while wetness measures the amount of water present in the canopy and underlying soils. It is important to note that these indices do not quantify the ancient Maya landscape or vegetation levels during their era. What they do record is the overall health, density, and wetness of the canopy as a result of ancient landscape modifications. Previous research has defined a relationship between vegetation signatures detected with high resolution IKONOS sensors and archaeological settlements (Saturno et al., 2007). Three common vegetational indices were chosen in an attempt to identify a similar relationship between present day vegetational characteristics detected at a 30 meter resolution and archaeological settlements. Descriptive statistics for both the dependent and predictive variables are provided in Table 3.

Binary Logistic Regression Binary Logistic Regression (BLR) is one of several statistical techniques used in predictive modeling that identifies relationships between candidate independent (predictive) variables and the dependent variable (Wrigley, 1976). Unlike discriminant analysis which is another common predictive modeling technique, BLR has relaxed statistical requirements and assumptions. BLR does not assume normality of the variables, but does require non-collinearity between independent variables (Easterbrook, 1999). Variables that exhibit collinearity may not be used simultaneously to estimate a predictive model. Despite the relaxed requirements of BLR, this statistical method has been shown to be more robust than many other methods (Kvamme, 1990; Press & Wilson, 1978). A Binary Logistic Regression takes the following form:

ProbabilityðyÞ ¼ 1= 1 þ Exp  b0 þ b1 X1 þ b2 X2 .bp Xp






Where Exp is the base of the natural log system, e, raised to the contents of the best linear combination of predictive variables. In addition to a constant term, b0, each b1 to bp represents the effect of the associated predictive variable. The probability of site presence (y) ranges from 0 to 1 and may be interpreted as a ratio. For example, a probability value of zero indicates no possibility of archaeological site presence. Locations with a probability value of 0.8 are twice as likely to contain a site as those with a probability value of 0.4. Values for each of the 10 predictive variables at the 100 point locations were obtained. Differences in the relationships of specific predictive variables at the known sites (STATE ¼ 1) and relationships at the points with an unknown archaeological state (STATE ¼ 0) provides the basis for the predictive model. Regression models were estimated using the SPSS statistical

Table 3 Descriptive statistics. Variable name

N

Minimum

Maximum

Mean

Standard Deviation

STATE EAST SOUTH DIST_WATER DIST_AG_LAND AG_LAND GREEN LAND_SYS NDVI SLOPE WETNESS

100 100 100 100 100 100 100 100 100 100 100

0 0 1 0 0 13.59 43.45 0 79.07 0 73.21

1 1 169 182.51 5351.89 52.65 78.81 2 85.23 14.78 90.21

0.50 0.33 72.50 17.30 1045.18 38.66 58.79 1.90 82.39 3.66 84.60

0.54 0.47 47.46 30.27 998.90 9.12 7.20 0.39 0.98 2.76 2.34

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

549

Table 4 Model variable coefficients, Wald Statistics, and odds ratios.

EAST GREEN CLOSE_AG_LAND Constant

B

Wald

Sig.

Odds

0.7548 0.0664 0.0005 4.1265

2.7704 4.2629 3.1135 4.7457

0.0960 0.0390 0.0776 0.0294

2.1271 1.0686 0.9995 0.0161

package which contains several methods for predictive variable entry to and removal from model estimations. The Backward Wald removal method was chosen which operates by eliminating the independent variable of least significance to the model at each iteration. Wald statistics needed be significant at the 0.1 level for the associated variable to remain in the model. Variable removal stops when each of the remaining variables is significant to the overall predictive model. The remaining variables and their coefficients are used to generate the predictive model. Results Exploratory analysis revealed that the variables GREEN, NDVI, and WET exhibited collinearity. Therefore, only one of these variables, GREEN, was used in subsequent models. GREEN was chosen because it was consistently more significant to the model than either NDVI or WET. The SLOPE variable exhibits collinearity with the two variables that were derived from it: AG_LAND and DIST_AG_LAND. To insure that models were not estimated using collinear variables, several models with different combinations of variables were estimated. Ultimately, a single model containing no collinear variables emerged. The predictive model resulting from this process included variables EAST, GREEN, and a variable representing the interaction between AG_LAND and DIST_AG_LAND called CLOSE_AG_LAND. The interaction between AG_LAND and DIST_AG_LAND quantifies the amount of agriculturally suitable land within a close proximity to a specific location. It is assumed that large areas of agriculturally suitable land would have been desirable, especially close to settlements. More parsimonious models were considered, including those which used AG_LAND and DIST_AG_LAND individually, but the adopted model provided the best statistical fit. Variables, their coefficients, significance, and odds ratios are listed in Table 4. The final predictive model is estimated as:

y ¼ 1=ð1 þ ðExp ð  4:1265 þ ðEAST 0:7548Þ þ ðCLOSE AG LAND  0:0005Þ þ ðGREEN 0:0664ÞÞÞÞ: Coefficients for predictive variables EAST and GREEN are both positive indicating that east facing locations with high amounts of vegetation are correlated with a higher probability of site presence. For example, a one unit increase in GREEN increases the odds of archaeological site presence by 1.0686. Therefore, it would appear that areas with high vegetational greenness are somehow associated with ancient Maya settlement. As Saturno et al. (2007) hypothesizes, this relationship is likely due to the high levels of plaster and cut limestone blocks found within Maya settlements. Saturno and others theorize that chemical amendments to the soil associated with decaying ancient settlement results in a detectable increase in present day vegetational health. A tendency for settling on slopes which face a certain aspect has been noted among many pre-historic societies (Christopherson & Entz, 2001). The predictive model presented here suggests such a tendency for the Maya to settle on slopes which face east. A preference for religious and ceremonial structures to be located on the eastern side of a courtyard is a wellknown characteristic of Maya settlement (Demarest, 2004: 201; Ricketson, 1933). Typically, these structures are religious or ceremonial in nature and have been reported to assist in tracking celestial movements (Ricketson, 1933). It is also possible that east facing slopes provided an unobstructed viewpoint for religious leaders and others to observe the night sky. For the variable CLOSE_AG_LAND, high values indicate large distances to large areas of arable land. The assumption is that the Maya would have preferred to live close to agriculturally viable land. However, the opposite situation is observed in this model. Based on the negative coefficient assigned to this variable, it seems that the Maya preferred smaller, and more easily accessible, arable land. This finding is somewhat contradictory to the assumptions of the research presented here. Several scenarios are possible which may help explain the negative relationship of this variable to ancient settlement. It is possible that the definition of arable or agriculturally suitable land used here is inaccurate. Several researchers have identified thriving Late Classic agricultural communities in areas which do not have a slope between one and five degrees (Walling, 2006). Additionally, many of the larger regional and civic centers did not participate in the production of their own food. Therefore, the proximity to arable land is not a factor in settlement patterns. Ideally, a larger initial number of known archaeological sites of each of the four ranks could be modeled separately. However, the foundation data for such an investigation does not exist. A more likely scenario alludes to the incredible population growth and resource exploitation that occurred during the Late Classic period. During this period, settlements within the study area experienced large population growth and settlement expansion. The escalating population required an ever-increasing quantity of resources such as food, water, and fabric. As populations reached apogee, every acre of arable land was used to feed and clothe the citizens. Land once considered too steep, too rocky, or containing poor soils would have been used. Several researchers have encountered terracing practices which would have leveled the ground surface and retained soils needed for growing crops (Walling, 2006). Others suggest

550

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

that the Maya drained swampy areas to create agricultural fields where once there were none. Therefore, a variable which quantifies the amount of arable land and the distance to that land is problematic and raises more questions than it answers. However, this variable is significant to the predictive model and is retained and used to generate predictive surfaces within the study area. According to the best predictive model chosen here, archaeological sites are likely to be located in areas which face east, have high vegetation greenness values, and do not have large quantities of arable land within a close distance. This model incorporates two characteristics of the Maya landscape (EAST and DIST_AG_LAND) and one of the present day landscape (GREEN). Likely, slopes which today face east, also faced east during the Late Classic period of Maya civilization. It is also quite probable that areas which exhibit between one and five percent slope today are comparable to those of the Maya era. The question of whether or not these regions would be considered agriculturally preferable by the Maya is up for debate. The third variable, GREEN, does not quantify the landscape of the Maya during the Late Classic. Instead, it characterizes a relationship between present day vegetational characteristics and the landscape modifications implemented by the Maya hundreds of years ago. Using the Spatial Analyst extension of ArcGIS 9.1, the Binary Logistic Regression formula and estimated parameter values were used to create a predictive surface. Using map algebra, the BLR formula computes a probability value for each grid cell located within the study area. The result is a continuous surface representation of archaeological site probability for the study area. The resulting predictive surface contains values between zero and one which indicate the probability of finding an archaeological site (Fig. 3). The probability surface exhibits a complete range of archaeological site potential across the study area. Several large areas of both extremely low and extremely high probability are present (Fig. 4). Areas which have extremely low archaeological site probabilities which correlate well with known karstic sinkhole plains. These areas quite often lack any topographic relief and, due to local soil and bedrock characteristics, cannot retain water. These conditions are not conducive to the growth of vegetation and therefore have low greenness values that influence the predictive model. These areas likely would have been avoided for habitation by the Maya. Other areas contain large regions of high archaeological potential. This is especially evident in the northeastern section of the study area. This region predominantly faces east and has high vegetational greenness values. According to the model estimated here, this region has the potential to contain archaeological settlements. The remainder of the study area has a probability value which varies from grid cell to grid cell. This is likely due to the variable topography associated with the bistected karstic plain of the region. Of the three predictive variables used to estimate this model, it is important to understand the relative importance of each variable to the model. To do so, the variable coefficients are standardized using the following formula:

b0 ¼ ðbx  sx  RÞ=sy Where b is the unstandardized coefficient for predictive variable x, sx is the standard deviation of the predictive variable x, sy is the standard deviation of the dependent variable, and R is the square root of Nagelkerke R2 statistic. Once calculated, the absolute values of each coefficient reveals its relative importance to the model. Variables and their standardized coefficients are listed by order of importance in Table 5. GREEN, followed by CLOSE_AG_LAND and EAST, has the most influence on archaeological site probability. This result emphasizes the expectation that ancient Maya stone structures such as temples, houses, courtyards, etc. affect present day vegetation patterns that can be used to predict site presence. The overall importance of GREEN to the predictive model corroborates the findings of Saturno et al. (2007) which indicate a strong relationship between vegetational signatures and archaeological settlements. Additionally, areas with low greenness values tend to be associated with sink hole plains which are typically avoided by past- and present day people for habitation.

Fig. 3. Archaeological probability surface (higher probabilities are indicated by darker shades).

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

551

Fig. 4. Areas of high and low archaeological probability.

Model validation Ideally, a model such as this would be field tested to assess its accuracy and predictive ability. This type of field validation is not feasible for most projects, including this one. Field validation would require several months of intensive archaeological survey requiring a field camp, numerous personnel, supplies, expensive computer and GPS systems, as well as generous funding and research permits from the Belizean government. However, quantitative assessment methods revealed the model to have moderate predictive ability. Sixty-six percent of the original 50 known sites are located in areas predicted to have high archaeological probability. Sixty percent of the 50 randomly selected sites were classified as high archaeological probability, indicating an overall predictive accuracy of both known and random locations of 63 percent (Table 6). Model accuracy can be assessed by comparing levels of correct classification with levels obtained purely by chance (Spicer, 2005). Even if the predictive variables have no relationship to archaeological site presence and absence, correct classification would still occur some percentage of the time. This percentage is referred to as chance accuracy. Because the dependent variable is binary (STATE ¼ 0 or STATE ¼ 1), the chance accuracy is 50 percent. The predictive model is considered effective if it correctly predicts archaeological site presence or absence at a rate better than chance (50 perfect). The criterion that the model must be 25% more successful at correctly classifying presence and absence than chance alone was implemented. Proportional chance accuracy is used to assess the predictive ability of the model over a chance classification.

Proportional chance accuracy ¼

  ðnumPresence=NÞ2 þðnumAbsence=NÞ2 125:

Proportional chance accuracy is calculated as the sum of number of sites with a value of presence and absence divided by the total number of sites, squared and multiplied by 125. The required proportional chance accuracy is therefore 62.5%. Accuracy of the predictive model at 63% exceeds the required proportional chance accuracy (Table 7). Therefore, this model can be considered a moderately effective predictor of archaeological site presence in northwestern Belize.

Kvamme’s Gain statistic Kvamme’s Gain statistic is another validation technique (Kvamme, 1990) that uses the predicted site probabilities at a set of known locations that were not included in the original model estimation. In this example, 19 known archaeological sites were withheld from the original set of training points. The original 100 training points and previously withheld 19 archaeological sites are then placed over the probability surface to examine the site classification accuracy of the model and Gain. Table 5 Model variable standardized coefficients.

GREEN CLOSE_AG_LAND EAST

Unstandardized Coefficient

Standardized Coefficient

0.0664 0.0005 0.7548

0.1218 0.1045 0.0909

552

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

Table 6 Model classification. Predicted

State

0 1

Percentage

Absence

Presence

Correct

30 17

20 33

60 66 63

Overall Percentage

The percentage of high probability land area and the percentage of known archaeological sites which fall within those regions serve as the basis for Kvamme’s Gain statistic:

Gain ¼ 1  ð%of area with high probability=%of sites with high probabilityÞ: This statistic ranges from 1 to þ1 and measures the predictive ability of a probability surface. Large positive values indicate that a predictive model is effective at predicting archaeological sites. A Gain of zero indicates that the model is no more successful at predicting archaeological sites than guessing. Ideally, the percentage of the study area classified as high probability is small, and the percentage of sites which fall within that area is large, resulting in a Gain statistic close to one. For this project, 42.82% of the study area and 57.89% of the withheld archaeological sites (11 of 19) fall within the high probability category which generates a moderate gain statistic of 0.26. Therefore, the archaeological predictive model has a positive predictability with an improvement over chance of about 26%. The validation results suggest that this predictive model is effective at identifying areas of high archaeological probability. Importantly, this model has identified several regions of the study area which are very unlikely to contain archaeological sites. Additionally, field surveys guided by this model are predicted to encounter archaeological sites about 25% of the time more often than the conventional, uninformed random survey. This model was generated using freely available satellite imagery data and published map sources. The statistical techniques and remote sensing indices are also fairly common and relatively simple to implement within a GIS. A large majority of the predictive work can be completed before any field work is necessary. The result is a model which provides a guide for future archaeological surveys in northwester Belize. Implementing a model such as this has the potential to streamline field projects and reduce the amount of time and effort required to locate new archaeological sites. Specific areas can be identified for intensive field surveys and other areas can be omitted completely. Also, the predictive model can be utilized in the field by means of GPS technologies. Field crews could make use of the predictive surface within the GPS to target high probability areas and change the itinerary of survey work using informed deductions about the local landscape. Anyone using this model should be more successful at finding archaeological sites than if performing a blind, uninformed survey. Therefore, this model and associated probability surface can be used to reduce the amount of time and money required to enhance the knowledge of the ancient Maya archaeological record. Discussion Once the archaeological potential model is mapped, several patterns become apparent. High and low site probability values tend to cluster in certain areas throughout the region. Several areas of low archaeological potential are known karstic sinkhole plains. Other areas with similarly low archaeological potential, vegetational greenness, and slope characteristics may also be karstic sinkhole plains which have yet to be surveyed. A clear connection between present day vegetational greenness and ancient Maya settlements is suggested by this research. The cracked and decaying structures built hundreds of years ago today support healthy stands of vegetation. Other researchers have suggested that concentrations of limestone block and plaster associated with Maya structures have altered the soil chemistry resulting in more robust vegetation (Saturno et al., 2007). There also appears to be a connection between the aspect of a slope and the probability that an archaeological site is located there. Whether this is due to the avoidance of the

Table 7 Chance classification accuracy and model classification accuracy. Predicted State

Null Model

Percentage Correct

Presence

Absence

State

Presence Absence

0 0

50 50

0 100 50

State

Presence Absence

30 17

20 33

60 66 63

Overall Percentage Predictive Model Overall Percentage

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

553

afternoon sun, or religious preferences for eastern facing slopes, this information can be used by investigators to guide field survey and investigations. The relationship between amounts of arable land and the relative distance from a settlement to that parcel is a bit more problematic. This negative coefficient associated with this variable may indicate that relatively flat arable land was not necessarily a characteristic valued by the Maya. It is known that the Maya were able to farm in areas with steep slopes by terracing the hillside (Kunen & Hughbanks, 2003; Turner & Harrison, 1983; Walling, 2006). This is especially true during the Late Classic period when record population growth resulted in the overexploitation of resources and a movement of people to locations once considered unsuitable or marginal for habitation. However, this relationship is statistically significant and warrants further investigation. A large area of high archaeological potential is located in the northeastern section of the study area. This area exhibits moderate relief, principally faces east, has relatively green vegetation and is well-suited for intensive survey and possible detection of undiscovered archaeological sites. Several other pockets of high archaeological potential are scattered throughout the region and could possibly contain an extremely large and high-ranking archaeological site. Identified pockets of high archaeological potential provide researchers with a locational focus for future ground survey efforts. Christaller’s central place theory (CPT) suggests that large settlements should be few and far between but set at regular intervals. Principles of CPT have been applied in the Belize River Valley region to identify locations that should contain archaeological sites (Marcus, 1973). The mid-sized settlement of El Pilar was successfully discovered close to a location which, according to principles of CPT, should contain a mid-sized settlement (Ford, n.d.). Due to several theoretical limitations of CPT, such as the assumption of an isotropic plane, these principles cannot be used to pinpoint exact locations of supposed settlement. Instead, CPT can be used as a guide to indicate regions which are likely to contain settlements of a particular rank and size. A combination of the principles of CPT and a predictive model which identifies areas of high site probability can further be used to reveal locations that are ripe for archaeological settlement discovery. Once identified, scientific field efforts can be focused in these regions to reduce the costs associated with field survey and increase the likelihood of archaeological site discovery. CPT, in conjunction with the predictive model, can also be used to exclude certain areas from future survey efforts. Within the study area, three settlements have been classified as large civic centers. When principles of CPT concerning the largest settlements are applied across the study area, much of the region is not suitable for a large high-ranking settlement. This is due to the relative proximity of the majority of the study area to known high-ranking civic centers. These sections of the study area that are too close to existing large settlements to contain undiscovered civic centers are grayed out in Fig. 5. The remainder of the study area represents locations which are sufficiently distant from known high-ranking civic centers to contain undiscovered high-ranking settlements. Within these regions where a high-ranking settlement is possible, several large areas of high probability exist. Researchers and archaeologists could use this combined approach of CPT and predictive modeling to refine survey strategies and maximize settlement discovery. Mid-sized settlements would also follow similar central place theory conventions. These settlements, smaller than their higher ranking cousins, would be more numerous and somewhat closer together. Locations classified with high archaeological probability that could contain mid-ranking settlements have been outlined in Fig. 6. Due to the population boom and overexploitation of resources associated with the Late Classic period, the smallest and lowest ranking settlements are likely located throughout the landscape wherever building construction is feasible. Researchers and archaeologists can use CPT and the predictive model estimated here to practice more informed survey tactics. Surveys and excavations can be focused on areas that have been identified as having high archaeological probability.

Fig. 5. Using Christaller’s Central Place Theory, large, high-ranking archaeological sites could be located within the indicated high probability areas.

554

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

Fig. 6. Using Christaller’s Central Place Theory, medium sized rank two and three sites could be located within the indicated high probability areas.

Areas exhibiting low archaeological probabilities can be omitted from surveys reducing the amount of time and money used to explore unproductive areas. Future archaeological surveys informed by the research presented here should be more productive and less costly than surveys which rely on uninformed chance methods. Conclusion Binary Logistic Regression was employed to identify predictive environmental variables for use in modeling the locations of Late Classic ancient Maya settlements. The three variables with the most predictive power are: Tasseled Cap Greenness index, eastern aspect, and a variable that measures the amount and proximity of flat land. The model suggests that archaeological sites are likely to be located in areas which face east, have small plots of easily accessible arable land, and high vegetational greenness. Overall, the predictive model performed fairly well when tested against a set of known sites withheld from the original estimation. A high percentage of these testing sites were, in fact, located in areas determined to have high archaeological potential. Also, nearly 43 percent of the study area is classified as having high archaeological potential indicating the extent to which the region was suited to Maya habitation. However, the only way to truly test a predictive model such as this one is through systematic field testing that remains to be done. The utility of relatively coarse resolution remotely sensed imagery for archaeological site prediction has been established. Previous studies have suggested that higher resolution imagery can detect archaeological sites in a neighboring portion of Guatemala (Estrada-Belli & Koch 2007; Pope & Dahlin, 1989; Saturno et al., 2007). The research presented here suggests that low resolution Landsat imagery may be used where higher resolution imagery is unavailable. Due to the relative ease of obtaining Landsat imagery and digital elevation models (or products derived from them), predictive models that rely on variables such as vegetational health and/or slope can be quickly generated. Areas with high vegetational health which indicate the presence of ancient structures can be identified prior to survey. Such methods are conducive to the field of Cultural Resource Management. Survey teams can focus their efforts within areas suspected to have high archaeological potential and hopefully obtain the most results for a small amount of effort. This is especially important in culturally rich areas such as Central America that are at risk of losing cultural remains due to looting and haphazard development. The archaeological predictive model generated here is a useful tool for those researching the ancient Maya of northwestern Belize. The property holdings of the Programme for Belize represents the largest single tract of land in Belize that has been set aside for conservation. Due to its relatively pristine nature and low risk of development, researchers can utilize this area to gain a better picture of Maya society. All types of settlements are located in northwestern Belize including large civic centers, local cities, and towns, and villages, affording the researcher an incredible opportunity to study the Maya from a regional context. The predictive model presented here should be used to guide future surveys to better understand and preserve the history of the Maya in this region. References Adams, R. E. W., & Jones, Richard C. (1981). Spatial patterns and regional growth among classic Maya cities. American Antiquity, 46(2), 301. Anschuetz, K. F., Wilshusen, R. H., & Scheik, C. L. (2001). An archaeology of landscapes: perspectives and directions. Journal of Archaeological Research, 9(2), 157–211.

S. Vaughn, T. Crawford / Applied Geography 29 (2009) 542–555

555

Ashmore, W., & Knapp, A. B. (Eds.). (1999). Archaeologies of landscape: Contemporary perspectives. Malden, MA: Blackwell Publishers. Baskin, C. W. (1966). Central places in southern Germany. Translated from Die zentralen Orte in Su¨ddeutschland by Christaller. Prentice-Hall, Englewood; Cliffs, New Jersey. Christopherson, G. L., & Entz, T. K. (2001). Field testing ARC GRID-based archaeological probability models in Jordan. Proceedings of the twenty-first annual ESRI user conference, San Diego, California. Coe, M. D. (2005). The Maya. New York: Thames & Hudson. Demarest, A. (2004). Ancient Maya: The rise and fall of a rainforest civilization. Cambridge University Press. Demarest, A. A., Rice, P. M., & Rice, D. S. (Eds.). (2004). The terminal classic in the Maya lowlands: Collapse, transition, and transformation. Boulder: University of Colorado Press. Dunning, N., Jones, J. G., Beach, T., & Luzzadder-Beach, S. (2003). Physiography, habitats, and landscapes of the three rivers region. In V. L. Scarborough, ´n Peninsula. F. Valdez, Jr., & N. Dunning (Eds.), Heterarchy, political economy, and the ancient Maya: The three rivers region of the east-central Yucata Tucson: University of Arizona Press. Easterbrook, R. B. (1999). Predictive probability model for American civil war fortifications using a geographic information system. Master’s Thesis, Virginia Tech. Estrada-Belli, & Koch, M. (2007). Remote sensing and GIS analysis of a Maya city and its landscape: Holmul, Guatemala. In J. Wiseman, & F. El-Baz (Eds.), Remote sensing in archaeology (pp. 263–281). New York: Springer. Folan, W. J., Faust, B., Lutz, W., & Gunn, J. D. (2000). Social and environmental factors in the classic Maya collapse. In W. Lutz, L. Prieto, & W. Sanderson (Eds.), Population, development, and environment on the Yucatan Peninsula: From ancient Maya to 2030. Laxenburg, Austria: IIASA. Research Report RR-00–14. Ford, A. (n.d.). An individual’s search for the ancient Maya. http://www.marc.ucsb.edu/elpilar/director_vision/ancientmaya.html, Accessed 20.09.06. Ford, A., & Fedick, S. (1992). Prehistoric Maya settlement patterns in the Upper Belize River Area: initial results of the Belize River archaeological settlement survey. Journal of Field Archaeology, 19(1), 35. Ford, A., & Fedick, S. L. (1988). Predictive model for ancient Maya settlement: The archaeological resources of the Rio Bravo conservation area, Belize. Riverside: Community Organization Research Institute, University of California. Furley, P. A. (1975). The significance of the Cohune Palm, Orbignya Cohune (Mart.) Dahlgren, on the nature and in the development of the soil profile. Biotropica, 7(1), 32. ´ n. Me´xico, D. F.: Instituto Nacional de Antropologı´a e ´ gico del Estado de Yucata Garza Tarazona, Silvia, & Edward Barna Kurjack Basco (1980). Atlas Arqeuolo Historia. Green, E. L. (1973). Location analysis of prehistoric Maya site in Northern British Honduras. American Antiquity, 38(3), 279. Jensen, J. R. (2006). Remote sensing of the environment: An earth resource perspective. Upper Saddle River, NJ: Prentice Hall. Kunen, J. L., & Hughbanks, P. J. (2003). In: V. L. Scarborough, F. Valdez, Jr., & N. Dunning (Eds.), Heterarchy, political economy, and the ancient Maya: The three rivers region of the east-central Yucatan Peninsula. Tucson: University of Arizona Press. Kvamme, K. (1990). The fundamental principles and practice of predictive archaeological modeling. In A. Voorrips (Ed.), Mathematics and information science in archaeology: A flexible framework. Bonn, Germany: Holos. Marcus, J. (1973). Territorial organization of the lowland classic Maya. Science, 180(4089), 911. Pope, K. O., & Dahlin, B. H. (1989). Ancient Maya wetland agriculture: new insights form ecological and remote sensing research. Journal of Field Archaeology, 16(1), 87–106. Press, S. J., & Wilson, S. (1978). Choosing between logistic regression and discriminant analysis. Journal of the American Statistical Association, 73, 699. Ricketson, O. G. (1933). Excavations at Uaxactun. The Scientific Monthly, 37(1), 72. Saturno, W., Sever, T. L., Irwin, D. E., Howell, B. F., & Garrison, T. G. (2007). Putting us on the map: remote sensing investigation of the ancient Maya landscape. In J. Wiseman, & F. El-Baz (Eds.), Remote sensing in archaeology (pp. 137–160). , New York: Springer. Scarborough, V. L., & Valdez, F., Jr. (2003). The engineered environment and political economy of the three rivers region. In: V. L. Scarborough, F. Valdez, Jr., & N. Dunning (Eds.), Heterarchy, political economy, and the ancient Maya: The three rivers region of the east-central Yucata´n Peninsula. Tucson: University of Arizona Press. Sever, T. L. (1998). Validating prehistoric and current social phenomena upon the landscape of the Peten, Guatemala. In D. Liverman (Ed.), People and pixels: Linking remote sensing and social science (pp. 145–163). Washington, D.C.: National Academy Press. Sever, T. L., & Irwin, D. E. (2003). Landscape archaeology: remote-sensing investigation of the ancient Maya in the Peten rainforest of northern Guatemala. Ancient Mesoamerica, 14, 113–122. Spicer, J. (2005). Making sense of multivariate data analysis. Thousand Oaks, CA: Sage Publications, Inc. Turner, B. L., II, & Harrison, P. D. (Eds.). (1983). Pulltrouser swamp: Ancient Maya habitat, agriculture, and settlement in northern Belize. Austin: University of Texas Press. Walling, S. Chawak But’o’ob, Belize: Prehispanic suburban ceremonialism, demography, and site summation. Paper presented at the 71st Annual Meeting of the Society for American Archaeology, April 26–30, 2006, San Juan, Puerto Rico. Wiseman, J., & El-Baz, F. (Eds.). (2007). Remote sensing in archaeology. New York: Springer. Wrigley, N. (1976). An introduction to the use of logit models in geography. In: Concepts and techniques in modern geography, Vol. 10. London: Institute of British Geographers.