Abstract
Predictive models are a component of GIS-based statistical approaches, which hold an important place in archaeological research due to advances in relevant theories and tools. Predictive models, developed through the statistical processing and analysis of environmental variables that influence site location, aid in understanding the cultural and natural landscape of the study area and contribute to the development of plans for improved cultural heritage management. This article, employing a statistical-analytical approach and data from archaeological surveys, aims to present a predictive model for a part of the eastern Kurdistan region where archaeological surveys have not yet been conducted. Prediction modelling was performed using the MaxEnt machine learning method, with eleven factors as natural variables and presence data (areas) required for modelling. The model area was divided into two experimental sections (Bijar and Dehgolan) and a prediction area (Qorveh), as the prediction model for Qorveh city was based on the natural variables and presence data from Bijar and Dehgolan cities. Finally, the prediction map was divided into four classes: very high, high, medium, and low suitability areas. The very high suitability area, which comprises 10% of the total model area, contains 59% of the Iron Age sites in eastern Kurdistan. It was found that vegetation cover, land use, and distance from rivers were among the most influential factors in the model. Also, the initial data in Qorveh indicate that 62% of the sites are located in an area comprising 8% with very high desirability, supporting the accuracy of the prediction. The AUC statistic is 0.836, and the finding value for the model has been calculated as 0.82, which indicates a prediction model with an approach value close to 1.
Keywords: Archaeological Prediction Model, GIS, MaxEnt, Eastern Kurdistan, Iron Age.
Introduction
The use of GIS and statistical modelling to map possible locations of archaeological sites has increased over the past decades. (Stefan & Sîrbu, 2010; Niknami et al., 2007; Alirezaei et al., 2019). Currently, archaeological prediction models are a powerful tool for preventing natural and human damage to historical and cultural resources (Danese et al., 2014), and for increasing the efficiency of archaeological field activities and cultural heritage management (Howard et al., 2016; Balla et al., 2014). Statistical modelling, as a perspective for identifying suitable areas for selecting prehistoric settlements, has been widely used by geographers and archaeologists (Sharafi et al., 2016; Verhagen & Dragut, 2012; Kaimaris, 2018). This method can be effectively considered a form of archaeological exploration. This paper aims to develop a concept for creating a prediction model using the MaxEnt method in the Eastern part of Kurdistan, to evaluate similar studies in archaeology, and to determine settings and suggest ways to optimise such approaches. Simultaneously, using a statistical approach based on GIS, the prediction model, and archaeological survey data from the eastern part of Kurdistan, it identifies the most favourable locations for the formation of Iron Age settlements. Accordingly, using environmental and archaeological data from Bijar and Dehgolan counties, a prediction model for Qorveh county has been developed using the MaxEnt method. Due to the lack of access to archaeological data for Qorveh County, the study area has been divided into two sections: experimental and prediction. This approach allows for predictions in the Qorveh region based on data from the experimental area (Bijar and Dehgolan counties), despite the absence of the required data. This method can thus be proposed as a study approach, with related recommendations, for another similar research.
Research Method: This research employed field and library (descriptive-analytical) methods, utilising the Geographic Information System (GIS) for the preparation and interpretation of GIS maps to analyse the settlement habitats of 96 Iron Age sites in eastern Kurdistan (Bijar, Qorveh, Dehgolan). The MaxEnt model was used to predict the distribution of Iron Age sites in eastern Kurdistan.
Data
The present study used a digital elevation model with a spatial resolution of 28 metres. Any change in these data will result in changes in climate, livelihoods, and other factors (Khosrowzadeh & Habibi, 2015: 109). The digital elevation model is used to extract new information such as slope, slope direction, and land curvature. This information is relatively common and significant, and is generally used in predictive models in archaeology. Land curvature data have also been used, which are defined as the rate of slope change (Whitworth, 2011: 469). The prediction model in this paper will be implemented using the principle of maximum entropy (MaxEnt). Such predictive modelling in archaeology requires two types of input data: environmental data (environmental variables that have a direct or indirect effect on the location of historical sites based on archaeological studies) and data related to archaeological sites, also known as presence data. The study area covers the political geography of Bijar, Qorveh, and Dehgolan counties. This study aims to make the most accurate prediction of Iron Age archaeological sites in Qorveh county using presence data (sites) from surrounding areas such as Bijar and Dehgolan counties. Given the similar climate and landscape in the eastern part of Iranian Kurdistan, the prediction of site formation locations in Qorveh County will be presented based on the MaxEnt prediction model.
Discussion
The final result of the prediction model for the eastern part of Kurdistan was based on the frequency ratio (FR) of the land cover and land use variables, which were among the most influential factors in the model. Their impact coefficients were estimated to be 24.3 and 32.6, respectively. Based on the classification of the forecast map, the low-desirability region covers the largest area within the forecast range, comprising 72% of the total. In contrast, better results can be observed due to the reduction in the area of regions in the high-desirability group. These regions, categorised as very high and high-desirability groups, comprise 10.5 percent and 7.5 percent of the total area, respectively. In contrast, it includes the largest number of areas, accounting for 89 percent of the total. These areas comprise 57 and 29 areas, respectively. While the areas with the highest potential are highlighted, they significantly reduce the area available for archaeological investigation.
Conclusion
MaxEnt modelling requires the use of presence data (areas). For this purpose, the prediction model is defined to include two categories of areas. First, the experimental area contains presence data (areas) as well as environmental factors and variables to configure the prediction model. Second: Prediction area; this section and perspective include the city of Qorveh in the eastern part of Kurdistan. The experimental area for this modelling contains 96 presence data points (areas) with an Iron Age chronology. Of these, 25% were used as test data and 75% as training data. Modelling with the MaxEnt method identifies the most influential variables by examining each one. Factors such as vegetation cover, land use, distance from the village, and distance from water sources are among the most influential variables on the model results. In this case, MaxEnt statistical analysis of other variables is also presented. Examination of the altitude variable shows that it has the greatest impact on sites within the altitude range of 1378–1400. The greatest impact of rivers on the sites occurred at a distance of 1000 metres. From 1000 to 3000 metres, the impact was least, but from 3000 to 5000 metres, the influence on site location became significant again. It has been found that at a distance of 2000 metres from villages, the likelihood of site formation is greater than at greater distances. The prediction map is divided into four groups based on the specified threshold value mentioned earlier: very high, high, medium, and low suitability areas. According to this division, the very high suitability area covers 10.5% of the total model area, and 59% of the sites (occurrence data) are located within this area. The high, medium, and low suitability areas include 30%, 6.5%, and 4.5% of the sites within these areas, respectively. Given that the very high and low suitability prediction areas are small but contain the largest percentage of sites, the model can be considered predictive.