Tải bản đầy đủ - 0 (trang)
5 AN EXAMPLE: PREDICTIVE MODELLING IN ACTION

5 AN EXAMPLE: PREDICTIVE MODELLING IN ACTION

Tải bản đầy đủ - 0trang

158



SPATIAL TECHNOLOGY AND ARCHAEOLOGY



Figure 8.4 Examples of the environmental variables selected as inputs for the final predictive model: distance to water

(left) and elevation (right). From Westcott and Kuiper (2000:63–65). Reproduced with permission.



exercise carried out by Westcott and Kuiper (2000). This study seeks to create an inductive model of

prehistoric site distributions in a large, unsurveyed coastal area of the Upper Chesapeake Bay. The problem

facing the archaeologists centred upon a study zone known as the Aberdeen Proving Ground (APG)—a 30,

400-hectare area that had been owned and utilised by the army for much of the 20th century. Due to a

combination of restrictions, unexploded ordnance and marshy conditions, only 1% of the overall area had

witnessed any form of structured archaeological survey, with only 46 prehistoric sites known from the

entire zone. Hence the need for a predictive model to facilitate the effective management of the

archaeological resource.

Ideally the formulation of the decision rule for the predictive model for the APG would be based upon the

results of a controlled, intensive survey within the actual APG zone. However, for the reasons outlined

above this was not possible. As a result the researchers decided to use what they termed ‘available data’ as a

proxy. This comprised careful study of a database of 572 prehistoric sites located in areas of the Upper

Chesapeake Bay that most closely resembled the environment of the APG. In an important deductive step

these sites were broken down into two distinct classes: shell-middens and lithic scatters, which were

assumed to have been located with respect to different, although overlapping, sets of variables. For each of

these sites a range of locational environmental factors (e.g. proximity to water, soil, soil drainage,

topographic setting, slope, aspect etc.) were recorded along with a comparative ‘background’ sample of 500

random locations taken from within the APG.

To decide upon which variables to use in the construction of the predictive model an inductive procedure

was followed based upon the generation of a series of frequency tables for the site and random background

locations with respect to each of the environmental variables and combinations of variables. This enabled a

number of variables to be eliminated—for example in the case of soil drainage the percentage of site

locations occurring on well drained soils was not significantly different to the percentage of well-drained

soils occurring in the background data.

In addition, aspect was inconsistently recorded on the site datasheets and seen to be closely correlated

with the particular shoreline (east or west). As a result of this exercise the following set of four core

environmental predictors were identified (Figure 8.4):

1. distance to water (fresh, brackish);

2. water type (e.g. river, creek, swamp);

3. elevation range;

4. topographic setting (e.g. floodplain, terrace, hilltop, bluff).



LOCATION MODELS AND PREDICTION



159



Although logistic and linear regression techniques were considered, it was acknowledged that the ‘available

data’ approach taken did not meet the statistical assumptions of these models. Instead, the weightings for

particular combinations of classes possible between the four predictors were determined by comparing site

locations from the proxy data set to the various combinations (see Table 8.2),

A high potential designation was then assigned to areas where sites occurred in any given unique

combination of the four variables over 20% of the time; medium potential for sites occurring between 6.25

and 20% of the time; and low potential for the remainder. Construction of the predictive model then

involved combining the four variables and allocating each cell its appropriate potential classification

dependent upon the unique combination of variables. The resultant shell-midden predictive model is shown

in Figure 8.5.

The efficacy of the final model (shown in Figure 8.5) was initially evaluated by comparing the known 46

prehistoric site locations within the APG to it. Looking to the shell-middens of the 13 known in the study

area, 12 fell within the high potential zone and 1 in the low potential. A more formal assessment of the

overall performance was undertaken using Kvamme’s simple gain statistic:

Table 8.2 The frequency table generated for the shell midden sites. From Westcott and Kuiper (2000:67). Reproduced

with permission.

Distance to

water (ft)



Water type Elevation (ft) Topography



Frequency Percentage Cumulative

frequency



Cumulative

percentage



0–500

0–500



Brackish

Brackish



≤20

≤20



75

81



34.7

37.5



75

156



34.7

72.2



0–500

0–500



Brackish

Brackish



>20

>20



14

2



6.5

0.9



170

172



78.7

79.6



0–500

0–500



Fresh

Fresh



≤20

≤20



24

10



11.1

4.6



196

206



90.7

95.4



0–500

>500

>500



Fresh

Brackish

Brackish



>20

≤20

≤20



4

4

2



1.9

1.9

0.9



210

214

216



97.2

99.1

100.0



Terrace/Bluff

Floodplain/

Flat

Terrace/Bluff

Floodplain/

Flat

Terrace/Bluff

Floodplain/

Flat

Terrace/Bluff

Terrace/Bluff

Terrace/Bluff



(8.5)

This is predicated upon the assumption that if the high potential area is small relative to the overall study

area and that if the number of sites found within it is large with respect to the total for the study area then

we have a fairly good model (Kvamme 1988:329). The closer the value of gain is to 1 the better the models

predictive utility. With a calculated value 0.82 for high potential areas and 0.80 for medium predictive areas

the model’s performance was deemed to be good (Westcott and Kuiper 2000:69).

8.6

METHODOLOGICAL ISSUES IN PREDICTIVE MODELLING

Despite the apparent methodological sophistication of methods such as logistic regression, there are a

number of criticisms that can be aimed at the construction of predictive models. Woodman and Woodward



160



SPATIAL TECHNOLOGY AND ARCHAEOLOGY



Figure 8.5 The final predictive model produced for the shell-midden sites. Westcott and Kuiper (2000:68). Reproduced

with permission.



(in press) have usefully reviewed the statistical and methodological basis of predictive modelling in

archaeology and noted several outstanding problems.

They draw attention to the fact that ‘case-control’ studies are normally used in predictive modelling.

Instead of sampling the landscape and then calculating the percentage of sampled points that turn out to be

sites, it is quite usual to gather a sample of sites and a sample of nonsites. This means that all estimates of

the probability of site presence must be relative estimates rather than absolute: it is not possible to use these

studies to predict that a location has x% chance of containing a site and another location a y% chance,

although it may be possible to state that location x is twice as likely as location y.

Woodman and Woodward also criticise the absence of attention in the predictive modelling literature to

the data requirements. Procedures such as logistic regression assume a linear relationship between

dependent and independent variables, an assumption that is rarely tested. Similarly, they point out that very

little attention is given to the nature of the interactions between the independent variables, which can take

the form of correlation, confounding or interaction.

Ebert (2000) is also highly critical of inductive predictive modelling methods, arguing that methods used

to test the accuracy of models by comparing results that have been obtained with one part of a sample

against another (withheld) part of the sample constitute little more than “a grossly inefficient way to

determine if there is inhomogeneity in one’s data” (ibid: 133). He is also critical of the way that maps are

translated directly into variables, and of the accuracy of published models, concluding that:

“Inductive predictive modelling…is not going to get any more accurate than it is right now. It focuses on

the wrong units of analysis, sites rather than systems, and attempts to relate their locations to

‘environmental variables’ which not only are probably not variables at all, but cannot be warranted by any

theoretical argument to be effective predictors of the locations of components of systems across

landscapes.” (ibid: 133)



LOCATION MODELS AND PREDICTION



161



8.7

THE PREDICTION PREDICAMENT: THEORETICAL DIFFERENCES OF

OPINION

“Most archaeological predictive models rest on two fundamental assumptions: first, that the settlement

choices made by ancient peoples were strongly influenced or conditioned by characteristics of the natural

environment; second that the environmental factors that directly influenced these choices are portrayed, at

least indirectly, in modern maps of environmental variation across the area of interest.” (Warren and Asch

2000:6)

As we said earlier in this chapter, the development and application of predictive models of archaeological

site locations has been controversial. We might go so far as to say that predictive modelling of

archaeological site patterns is the most controversial application of GIS within archaeology and although

this volume is not intended to contribute or extend this debate, the reader will need to be aware of the

essential differences of opinion that exist within the discipline (see e.g. Gaffney and van Leusen 1995,

Kvamme 1997, Wheatley 1998, Kuna 2000).

Central to the critique of predictive models has been the accusation that the entire methodology embodies

a theory of archaeology that can be categorised as environmental determinism. The debate about

environmental determinism and predictive modelling has been vigorous. Some advocates of predictive

modelling claim that the utility of models to identify patterns can be isolated from their use as interpretative

tools. As van Leusen puts it:

“Statistics can be used to describe patterning in archaeological datasets in a rigorous manner without

reference to the cause(s) of those patterns, and if the extrapolation of those patterns yields predictions that

are useful in CRM…the method is validated.” (Gaffney and van Leusen 1995:379)

Critics of predictive modelling, as currently practiced, have tended to argue that the description of patterns,

and their modelling for whatever purpose cannot be separated from the process of explanation in this way.

Even if it could be, then the correlation of static archaeological settlement patterns with environmental

variables constitutes a worthless exercise. The argument that inductive modelling can therefore be separated

from deductive, explanatory modelling has been described by Ebert as a product of the “sheer indolence

among those who think we can ‘stop’ at ‘inductive predictive modeling’” (Ebert 2000:130). He goes on to

argue that:

“Predictive modeling will be transformed into a worthwhile adjunct to archaeology and archaeological

thinking only by the formulation of a body of explanatory propositions linking contemporary correlations

with the past. In other words it is productive explanatory thought, and not computers, that can potentially

raise predictive modeling above an anecdotal level.” (ibid)

Other defenders of locational models have objected to the use of the term ‘environmental determinism’ to

describe the approach, arguing that the prevalence of environmental variables in archaeological predictive

models is simply a product of their greater availability over ‘cultural variables’ and so should not be taken

as an indication of theoretical orientation. To quote Kohler:

“Given the subtleties and especially the fluidity of the socio-political environment, is it any wonder that

archaeologists have chosen to concentrate on those relatively stable, distorting factors of the environment for

locational prediction?” (Kohler 1988:21)

Moreover, some (Kvamme 1997:2) have objected on the grounds that the term is an inappropriate—even

offensive—label that associates predictive modellers with late 19th and early 20th century human

geographers such as Rahtzel (1882), whose ideas were later widely discredited.



162



SPATIAL TECHNOLOGY AND ARCHAEOLOGY



8.8

CONCLUSIONS

We have attempted in this chapter to introduce some of the methods central to a large field of research. We

have tried to indicate some of the methods of (mainly inductive) modelling that have been applied to

archaeological situations, devoting greatest attention to the method that is currently most popular, logistic

regression. We have also tried to draw attention to some of the unresolved methodological issues

surrounding the statistical methods and to some of the theoretical concerns that have caused so much debate

in the archaeological GIS literature.

Inevitably, constraints of space have prevented us from exploring any but a small subset of the range of

possible methods for the generation of models and the interested reader will want to follow up many of the

references included here, particularly the papers published in Judge and Sebastian (1988) and, more

recently, Westcott and Brandon (2000).

The perceptive reader will also have understood that predictive modelling is not a field in which we have

any programmatic interest ourselves. While we have attempted to write a balanced and reasonably complete

account of archaeological predictive modelling, it is unlikely that we have been successful in completely

disengaging from an academic debate in which we have been quite active protagonists. Both authors have

published criticisms of predictive modelling (Wheatley 1993, Gillings and Goodrick 1996) and would

subscribe to the position that it is a field with significant unresolved methodological and, more

significantly, theoretical problems.

Some of these we have introduced above, but a theoretical debate such as this is difficult to synthesise

without reducing it to the level of caricature. We therefore hope that archaeologists who want to take a more

active interest in this field will follow up references to the debates and discussions that have been published

in the archaeological literature (notably Kohler 1988, Gaffney et al. 1995, Kvamme 1997, Wheatley 1998,

Church et al. 2000, Ebert 2000) and formulate an independent opinion as to whether Predictive Modelling has

a future in archaeology.



Tài liệu bạn tìm kiếm đã sẵn sàng tải về

5 AN EXAMPLE: PREDICTIVE MODELLING IN ACTION

Tải bản đầy đủ ngay(0 tr)

×