Royal Society Publishing

Predicting pre-Columbian anthropogenic soils in Amazonia

C. H. McMichael , M. W. Palace , M. B. Bush , B. Braswell , S. Hagen , E. G. Neves , M. R. Silman , E. K. Tamanaha , C. Czarnecki

Abstract

The extent and intensity of pre-Columbian impacts on lowland Amazonia have remained uncertain and controversial. Various indicators can be used to gauge the impact of pre-Columbian societies, but the formation of nutrient-enriched terra preta soils has been widely accepted as an indication of long-term settlement and site fidelity. Using known and newly discovered terra preta sites and maximum entropy algorithms (Maxent), we determined the influence of regional environmental conditions on the likelihood that terra pretas would have been formed at any given location in lowland Amazonia. Terra pretas were most frequently found in central and eastern Amazonia along the lower courses of the major Amazonian rivers. Terrain, hydrologic and soil characteristics were more important predictors of terra preta distributions than climatic conditions. Our modelling efforts indicated that terra pretas are likely to be found throughout ca 154 063 km2 or 3.2% of the forest. We also predict that terra preta formation was limited in most of western Amazonia. Model results suggested that the distribution of terra preta was highly predictable based on environmental parameters. We provided targets for future archaeological surveys under the vast forest canopy and also highlighted how few of the long-term forest inventory sites in Amazonia are able to capture the effects of historical disturbance.

1. Introduction

The perception that the entirety of the Amazon Basin was a virgin forest or ‘counterfeit paradise’ before European arrival to the Americas has been refuted, as recent research has documented evidence of earthwork formation, complex societies [13] and modified soils [4,5]. However, the overall extent to which lowland Amazonia was actively transformed into a modified landscape or cultural parkland by native people remains unknown and debated [610]. A newly emerging view is that pre-Columbian impacts across Amazonia were heterogeneous, both in space and degree [9,10]. However, in the areas that were occupied, the question remained: to what extent were the legacies of pre-Columbian societies evident in modern forest structure, dynamics and biodiversity patterns [6,9,11]?

Legacies of prehistoric influence in Amazonia are visible from aerial surveys and satellite imagery along the southern periphery of the basin, particularly in the Llanos de Moxos of Bolivia (e.g. [1,1214]), the geoglyph-rich region of Acre, Brazil and northern Bolivia [3,1518], and the Upper Xingu in Brazil [2,19] (figure 1). Whether similar impacts occur deep within the interior forests and across most of the basin remains unknown [9,10].

Figure 1.

Locations of previously published terra preta and terra preta-free sites, major archaeological sites (references = M [20], X [2], S [21], C [22], Llanos de Moxos [12], geoglyphs [3,18]), and river systems within Amazonia.

Potentially, one of the most telling vestiges of pre-Columbian activity in the interior forests is Amazonian black earths, known as terra pretas. These anthropogenic soil types were usually formed between 500 and 2500 years ago and contain significantly higher nutrient levels than typical Amazonian soils [4,20,2328]. Terra pretas indicate sedentary pre-Columbian settlements and are the most widely reported archaeological feature in Amazonia. They occur at most of the major archaeological sites in the interior forests including the Central Amazon Project [4,23], Santarem [28], Marajo [29] and the Upper Xingu [19,30] (figure 1). Terra preta formation has been estimated to occur in as much as 10% of the forests [6,7], but a prediction derived from empirical data and modelling is lacking.

Basin-wide soil surveys show that terra preta distributions are non-random (figure 1). Terra pretas occur throughout Amazonia, but are most commonly found in the central and eastern forests [26,31], and in riverine settings, particularly along bluffs overlooking the rivers [32]. However, the noted patterns may result from increased sampling effort in these locations. Less is known of terra preta occurrence in the westernmost and interfluvial forests, though most soil surveys in these areas contain no evidence of terra pretas [10,3338] (figure 1).

As the potential for terra preta formation exists throughout Amazonia, we hypothesized that regional environmental conditions and local sociocultural practices influenced how people lived and that, in turn, affected their likelihood of creating terra preta. Thus, we modelled the potential distribution of terra pretas based on presence locations and a suite of climatic, geological, terrain and hydrological conditions across Amazonia. From our model, we determined the regional environmental factors that most strongly influence the likelihood of terra preta formation and estimated the spatial extent and total area of terra preta occurrence across Amazonia. We validated our model with additional terra preta presence and absence locations, and compared the output of our model with locations of previous ecological and archaeological research in Amazonia.

2. Material and methods

(a) Soil surveys

Typically, archaeological researchers have not published the exact georeferenced coordinates of terra preta locations for a number of reasons that include the relatively recent innovation of GPS to identify locations accurately and protection of sensitive sites. However, WinklerPrins & Aldrich [39] compiled the locations of over 500 terra preta sites, which were classified as either high, medium or low confidence, according to the predicted accuracy of the geographical coordinates. In total, 454 sites fell within lowland Amazonia (figure 1). We combined these data points with 463 georeferenced terra preta sites uncovered by E.G.N. and E.K.T. in archaeological field surveys from 2008 to 2012 to obtain a total of 917 known terra preta locations (figure 1).

Cooper et al. [34] published a comprehensive database of soil surveys conducted in Brazil, with most data points originating from the field campaign of the RADAMBrasil Project (Radar in Amazonia) [33]. In total, 1791 of these sites lie within Amazonia (figure 1). No site in the database contained notes or descriptions regarding any pottery sherds or soils with the black coloration or anomalously high nutrient content of terra pretas. Other Amazonian soil surveys [10,37,40] also report no evidence of terra preta. The locations of terra preta-free sites from all of these surveys were combined to form a database of terra preta absences (n = 1962). The combined terra preta and terra preta-free dataset contained 2879 sites.

(b) Environmental data

Environmental data was acquired or derived for the 4 876 193 km2 of Amazonian forests (figure 1), and if necessary, resampled to 90 m spatial resolution. All geospatial data organization and analyses was conducted using open-source python code that incorporated Geospatial Data Abstraction Library (gdal) and osgeo libraries. We compiled 45 environmental layers for possible inclusion in the model, including bioclimatic (n = 19), soil (n = 16), terrain (n = 8), hydrological (n = 2) and geological (n = 1) characteristics (electronic supplementary material, table S1; descriptions below). To reduce autocorrelation between variables, we performed cross-correlation analyses on the bioclimatic, terrain and soils variable groups [41] (see the electronic supplementary material, figure S1). Additional organization and data extraction were performed with ArcGIS v. 10.0.

A total of 22 variables were chosen for inclusion in the model, including nine climatic, six edaphic, four topographical, two hydrological and one geological variable (see the electronic supplementary material, table S1; described below). The WorldClim Bioclimatic variables, which include precipitation and temperature parameters across the globe at a 1 km resolution [42], were used as climatic predictors. To account for autocorrelation between variables, we reduced the 19 variables down to nine using cross-correlation matrices [41] (see the electronic supplementary material, figure S1 and table S1). We used the Harmonized World Soil Database v. 1.2 [43] for edaphic characteristics across Amazonia. As topsoil and subsoil characteristics were highly correlated with each other, we used primarily subsoil characteristics in the predictive model (n = 6) (see the electronic supplementary material, figure S1 and table S1). The topographic data were derived from the Shuttle Radar Topography Mission (SRTM) 90 m resolution elevation data (data available from the US Geological Survey (USGS), http://eros.usgs.gov). Elevation data were included in the model and were also used to derive metrics of terrain roughness and slope. The largest intercell difference of a central pixel and the adjacent eight pixels is defined as roughness [44] and was used to identify river bluffs. For each pixel, we also calculated the distance to the nearest terrain changes, including: (i) bluffs greater than 25 m, (ii) bluffs greater than 50 m, (iii) slopes greater than 7°, and (iv) slopes greater than 15°. These metrics were added to account for the lack of precision of some of the terra preta locations, and the fact that people may have formed terra pretas near the bluffs as opposed to directly atop them. Roughness and slope were highly correlated with each other, so only roughness metrics were used in the model (see the electronic supplementary material, figure S1 and table S1). The hydrological variables stem from the HydroSHEDS dataset (available at http://hydrosheds.cr.usgs.gov), which identifies the river networks of Amazonia [45]. ‘Upcell’ values are features of the HydroSHEDS dataset and represent the maximum flow accumulation at any location in the river network. We used rivers with upcell values of greater than 15 000 to delineate perennial rivers, similar to previous studies [46] (figure 1). We calculated the distance to the nearest riverbank (km) and size of the nearest river (upcell value) for each pixel across Amazonia (see the electronic supplementary material, table S1) to use as predictors of terra preta. The geological provinces polygon raster was obtained from the USGS (http://energy.cr.usgs.gov).

(c) Models and analysis

We modelled terra pretas based solely on the presence data for several reasons. First, the sampling strategies of the presence-only dataset were targeted and localized, whereas the terra preta-free datasets were based on landscape-scale stratified sampling designs. Second, the accuracy of locations in the RADAM dataset, which comprised the majority of points in the terra preta-free dataset, may be questionable as the points were collected prior to modern GPS technology. Also, RADAM and other regional Brazilian soil surveys do not contain evidence of terra pretas, although it is not explicit as to whether they actually searched for them. However, because of the extensive chemical analyses performed on these soil samples, terra pretas most likely would have been identified through the presence of abundant artefacts, charcoal or anomalous nutrient levels.

Maxent v. 3.3.3 (http://www.cs.princeton.edu/~schapire/maxent/) was used to model terra preta distributions, as it commonly outperformed other modelling techniques that use presence-only or presence–absence data, especially when dealing with small sample sizes or non-random sampling [4749]. The predictive model was produced using only the ‘high confidence’ sites from the WinklerPrins and Aldrich [39] database (n = 201), sites surveyed by Paz et al. [50] (n = 8), and the sites surveyed by E.G.N. and E.K.T. (n = 463; total n = 701). We performed 10-fold cross-validation of the model, which was advantageous because all the data points were used for model training and also for performance testing. We evaluated area under the curve (AUC) statistics, which indicated the predictive capacity of a model, and how much better (or worse) a given model performed compared with a random model, where AUC = 0.5. Response curves for each environmental variable, a table of permutation importance for each of the environmental parameters, and jackknife tests for the training and testing datasets were all used to assess the relative importance of the predictor variables.

To analyse model uncertainty, we then extracted the predicted values for each of the terra preta and non-terra preta sites (n = 2879) (figure 1). A probability threshold value for predicting terra preta presence and absence was generated based on maximizing the percentage of sites correctly classified [51]. This method allowed us to estimate false positive and false negative rates of the model. Using this threshold, we were also able to predict whether each pixel across Amazonia probably contains terra preta or not (i.e. predicted presence and absence). From this, we calculated the extent of forests that probably contained terra preta as the area of the pixels predicted as present. All statistical analyses were performed with the ‘PresenceAbsence’ [51] and ‘stats’ [52] packages for R using RStudio v. 0.96.331 [53].

3. Results

The outputs of the maximum entropy model included probabilities of finding terra preta for each pixel across Amazonia (figure 2). The probabilities ranged from 6.2 × 10−9 to 0.96, with most of Amazonia at the low end of that range. The average AUC value obtained from the 10-fold cross-validation of the model was 0.957, with a standard deviation of ±0.006, which was much higher than that of a random prediction (0.5), and indicated that the model fitted the data well and had a high predictive capacity.

Figure 2.

Probabilities of terra preta occurrence based on predictive models. Legend of archaeological sites is the same as in figure 1. The black line indicates a potential cultural boundary where the probability of terra preta formation decreases and disappears and is replaced by alternative subsistence strategies in southwestern Amazonia (see text for details).

Terrain, geological and hydrological characteristics were more important in shaping terra preta distributions than were climatic ones (table 1 and figures 35). Elevation was the most important factor in the model, and the potential distribution was constrained to areas primarily between 0 and 100 m elevation (figure 3). The second most important predictor was geological province, and terra pretas are most likely to be found in the Solimões Basin, Ucayali Basin and Brazilian Shield regions (figure 3). As expected, riverine settings were also important in the model. Our model suggested that terra pretas are also most probable within 10 km of a river, and near bluffs more than 25 m (figure 3). The type of river does not seem to be as important, as areas of high terra preta probability fall along black-water (including the Rio Negro), white-water (including the Amazon) and clear-water rivers (including the Tapajos) [26,55].

View this table:
Table 1.

Permutation importance (%), AUC of model with variable excluded (AUC minus predictor), and AUC of model with only individual predictors included (AUC only predictor) for each environmental layer used in Maxent models.

Figure 3.

Response curves for the geological and geographical variables (see Material and methods) included in the predictive model. Because many of these variables are correlated with each other (see the electronic supplementary material, figure S4), these response curves are based on a Maxent model using only the corresponding variable [54].

Figure 4.

Response curves for the soil variables [43] included in the predictive model. Because many of these variables are correlated with each other (see the electronic supplementary material, figure S4), these response curves are based on a Maxent model using only the corresponding variable [54]. S CEC, subsoil cation exchange content.

Figure 5.

Response curves for the climatic variables [42] included in the predictive model. Because many of these variables are correlated with each other (see Material and methods and the electronic supplementary material, figure S4), these response curves are based on a Maxent model using only the corresponding variable [54].

Model results also indicated that the probability of finding terra preta is almost zero at the major archaeological sites in Bolivia, including the Llanos de Moxos and the geoglyph-rich region (figure 2). Similarly, the model predicted that terra pretas are very unlikely to be found in the forests in close proximity to the western Amazonian earthworks. An absence of terra preta was also predicted across most of western Amazonia and in the upper catchment areas of southeastern Amazonia.

When considering the predicted probabilities of all known terra preta presence and absence points (figure 1), a probability threshold of 0.25 was able to correctly classify the highest percentage of known sites. Using the 0.25 threshold, 89% of all cases were predicted correctly, with a false negative rate of 6.5% and a false positive rate of 4.7%. When this probability threshold was applied across the basin, ca 154 063 km2 or 3.2% of the forest was predicted as having conditions that made terra preta formation likely.

4. Discussion

(a) Spatio-temporal heterogeneity of pre-Columbian peoples

Terra preta formation was most probable in eastern and central Amazonia, and in association with the lower courses of major rivers (figures 2 and 6). The older and highly weathered cratonic soils of eastern Amazonia were particularly nutrient-poor [37,56], and terra preta formation increased fertility in soils and afforded repeatable harvests in nutrient-poor areas at higher population densities. The larger acreage terra pretas, discovered to date, also lay near the Amazon River channel and its major tributaries, but smaller terra pretas occurred on the floodplains and in upper catchment areas [5,26,32,39,55,5759]. This distribution may indicate a settlement preference for larger rivers, but may also be related to flood conditions, which are longer but not as deep in the lower courses. Furthermore, the eastern sections of the Amazon River are less turbid and have slower flow and meander rates compared with the western sections [60]. This stability may have also favoured larger and more permanent settlements in eastern Amazonia, particularly in regions such as Santarem, where some of the oldest records of human occupation in Amazonia have been identified [21,25]. The first European account of travel down the Amazon River in 1541–1542, written by Gaspar de Carvajal, described flourishing and fruitful societies in the easternmost sections of river, although at least some aspects of these accounts are of dubious reliability, i.e. the description of a nation of warrior women—the Amazons [61].

Figure 6.

(a) Terra preta probabilities in relation to locations of long-term forest inventories, large scale biosphere-atmosphere experiment in Amazonia (LBA) C-flux towers, geoglyphs, bamboo-dominated forests and major archaeological sites. The legend of archaeological sites is shown in figure 1. The location and spatial extent of panels (b) and (c) are shown as the red and black rectangles respectively. (b) Regional view of areas with forest inventory data and an increased probability of terra pretas. (c) Regional-scale view of the potential cultural boundary between ancient cultures that formed terra preta, and those that used alternative subsistence practices.

The lack of terra pretas in western Amazonia may be because the Andean-derived soils of western Amazonia did not require nutrient enrichment, or because the human footprint in western Amazonia is for the most part smaller than that in the east. These possibilities are not mutually exclusive, although multiple lines of palaeoecological evidence suggest heterogeneous and localized low-impact human activity in much of western Amazonia [10]. Also, Carvajal's account describes the western river sections as desolate, with the expedition starving to the point of eating shoe leather [61]. That is not to say people were absent from western Amazonia; most evidence indicates that with the exception of the earthwork-rich areas in southwestern Amazonia, people were sparsely scattered through the landscape, clearing small tracts of land and growing crops on localized scales.

Cultural differences undoubtedly contributed to the distribution of terra preta formation, and these boundaries may have been just as important as environmental conditions (figure 2). The Llanos de Moxos regions of Bolivia is located in seasonally inundated forests at the forest–savannah transition, and instead of terra preta formation, large societies sustained themselves by using techniques such as fish weirs and raised-field agriculture (e.g. [12]). In the nearby geoglyph-rich regions of Acre, Brazil and northern Bolivia, our model predicts an absence of terra preta (figures 2 and 3). These predictions are concordant with archaeological surveys which report an absence of terra pretas within the region [18]. The forests surrounding both sites were also predicted to be terra preta-free. Thus, the geoglyph builders may have been more culturally similar to people of the Beni. Interestingly, the geoglyph-rich region at least partially overlaps with expansive bamboo-dominated forests in Amazonia (figure 2) [62] that have persisted in the landscape for thousands of years [36]. A direct connection between these features has not yet been demonstrated with empirical data.

We suggest that aside from the areas that lie near ecotonal boundaries, such as Marajo, Xingu and the Llanos de Moxos (figures 1 and 2), our model predicts the areas of lowland interior Amazonia that experienced the most intensive pre-Columbian disturbances. Our reasoning is owing to the fact that terra pretas are found throughout all of the major archaeological sites in the interior of the Amazon Basin, including areas near the Central Amazon Project and Santarem (figures 1 and 2). Our estimates of potential terra preta sites were significantly less than the previous estimates of 10% [6,7] and did not support the assertion that most of Amazonia was a transformed landscape or cultural parkland before European arrival. Instead, these data suggested that pre-Columbian Amazonia was a heterogeneous landscape, with varying degrees of human impacts across the mosaic.

Also, our model probably overestimates the areas where terra pretas will be found at any given time because it predicts the summed distribution taken across all ages. Dates from terra pretas are not always contemporaneous, and range from 3000 to 500 years ago. Most recorded dates are from 1000 to 2000 years ago [22], and the spatial extent of terra pretas probably expanded significantly during this time. Whether terra preta sites were continuously occupied after the original formation remains unknown. Some sites, particularly the smaller ones, may have been used only briefly. Further dating of terra preta sites could determine whether age is associated with size or a particular geographical region. Potentially, with enough dates, the spread of terra preta culture through time could be reconstructed.

Our model contradicts recent assertions that pre-Columbian human disturbance increased Amazonian biodiversity through promoting preferred species [11]. None of the predicted highly impacted areas are as biodiverse as the forests of northwestern Amazonia, which contain the highest tree-species diversity per unit area. Consistent with this view, the western, and particularly the northwestern, forests contain limited, if any, evidence of pre-Columbian human presence [10,63].

(b) Archaeological and ecological implications

Many archaeological sites are discovered accidentally through deforestation [3,15] or as a result of conversations with local people. Our model predicts where ancient impacts should be the most intense, even under the forest canopy, and provides basin-wide target locations for archaeologists who study the rich culture of ancient societies within the vast Amazonian rainforests (figure 6). Distribution models such as the one presented here, could also be applied to other archaeological features, for example geoglyphs, allowing examinations of cultural boundaries and areas of potential cultural overlap (figure 6c).

We do note, however, that because most terra preta and terra preta-free sites lie in Brazil, our most confident predictions are for this country. Elsewhere in Amazonia, the predictions are based on larger interpolations and are less certain. Increased sampling will refine our estimates of terra preta prevalence outside of Brazil.

A huge investment has been made by ecologists towards understanding ecosystem patterns and processes. Most ecological data collection and field surveys, particularly in the Neotropics, have been conducted over recent timescales, i.e. the past 50 years. Although these data have been immensely valuable in deciphering many of the underlying mechanisms that shape forest composition, structure and change, they lack a historical perspective. It has been suggested that the historical contingency, or historical trajectory, can play just as large a role in shaping observed ecological phenomena as modern climatic and abiotic conditions [64,65]. Our model provides a basis for linking historical trajectories of forest change with forest inventory data and ecological surveys conducted in Amazonia over modern timescales.

Most of the previous forest inventory research occurred in areas that most probably do not contain terra pretas, but the long-term study sites near Manaus and Santarem are particularly promising for local-scale evaluations of pre-Columbian disturbance on forest structure (sites C and S in figure 6). Many tree species in Amazonia live several hundred years [66], and forest burning and clearing associated with terra pretas from 2000 to 500 years ago, and in many cases again at the turn of the twentieth century, may have significantly contributed to patterns and dynamics documented in the modern forests. Previous field reports have documented that terra pretas are often associated with lower canopy height [28], denser understory [28], increased frequencies of lianas [67] and are often identified by locals as secondary forests with a predictable suite of species, many of which are domesticated (e.g. [68,69]). Other ‘anthropogenic’ forests that were formed in pre-Columbian periods may be identifiable within the forest today, as some species, including some palms, were heavily promoted species [6,7072].

Data accessibility

All terra preta locations reported in this study will be added to the existing online database of WinklerPrins and Aldrich [39].

Funding statement

This work was funded by NASA Space Archaeology Grant NNX10AM14G, an Iola Hubbard Endowment through the Earth Systems Research Center at the University of New Hampshire, and the NSF Ecology Program (award DEB 0742301).

Acknowledgements

We thank D. Piperno for comments on the manuscript.

  • Received September 20, 2013.
  • Accepted November 28, 2013.

References

View Abstract