Predicting the geographic distribution of ancient Amazonian archaeological sites with machine learning

Removing predictor variables that were collinear (>0.8) has no effect on the spatial cross-validated AUC.

“We tried removing predictor variables that were collinear (>0.8), but this has no effect on the spatial cross-validated AUC so we left in all variables.”

The random forest classifier consistently outperformed neural networks, extreme gradient boosting tree, support vector classifiers, and lasso logistic regression.

“We first evaluated several common machine learning algorithms including neural networks, extreme gradient boosting tree, support vector classifiers, and lasso logistic regression and found them to be relatively high performing. However, the random forest classifier consistently outperformed the other alternatives.”

The best model at the optimal spatial scale of 1 km (30 arc-second) has an AUC of 0.91 (accuracy 0.88, mean sensitivity 0.83, mean specificity 0.90) under spatial cross-validation.

“Model testing is done with spatial cross-validation, and the best model at the optimal spatial scale of 1 km has an Area Under the Curve of 0.91.”

• Also stated in: § Machine learning · plain

  “Therefore, we revert to the 30 arc-second scale which had the highest of all testing performance with an AUC of 0.91 (accuracy 0.88, mean sensitivity 0.83, mean specificity 0.90) using the spatial cross-validation scheme.”

The predictive model has led to the discovery of 13 new geoglyphs.

“Our predictive model has led to the discovery of 13 new geoglyphs, and it pinpoints specific areas with high probabilities of undiscovered archaeological sites that are currently hidden by rainforests.”

Most of the top predictors are distances to major rivers of different sizes and types, especially black and clear rivers.

“Out of a total 65 geospatial variables, most of the top predictors are distances to major rivers of different sizes and types (especially black and clear rivers, Figs. 2 and 3).”

In both variable-reduction exercises there was no significant change to prediction maps.

“In both exercises there was no significant change to prediction maps (i.e., the models remained essentially the same).”

Reducing the overall number of predictor variables using recursive feature elimination and forward feature selection has limited effect on the AUC.

“Likewise, reducing the overall number of predictor variables using recursive feature elimination in the caret package and the forward feature selection in the CAST package has limited effect on the AUC so we again decided to keep the full set of predictors.”

The predictive model pinpoints specific areas with high probabilities of undiscovered archaeological sites currently hidden by rainforests.

“Our predictive model has led to the discovery of 13 new geoglyphs, and it pinpoints specific areas with high probabilities of undiscovered archaeological sites that are currently hidden by rainforests.”

• Also stated in: § Results · plain

  “Only specific locations exist where there are high probabilities of finding additional earthworks or ADEs that have yet to be discovered (Fig. 5).”

• AssumptionThe authors state this as an assumption: Current non-forested areas have already been discovered, so archaeological sites there are likely already found.

Spatial autocorrelation is strong with these archaeological locations.

“Spatial autocorrelation is strong with these archaeological locations.”

The authors found a consistent decline in spatial cross-validation AUC metrics at coarser resolutions.

“However, we found a consistent decline in spatial cross-validation AUC metrics at coarser resolutions.”

As the authors experimented with different modeling approaches, they found drastic improvements in the generalizability of models with the inclusion of all three site types.

“As we experimented with different modeling approaches, we found drastic improvements in the generalizability of models with the inclusion of all three site types.”

The predictive model generates geographic hotspots for where earthworks and Amazonian Dark Earth (ADE) sites are likely to be found in the future.

“Our predictive model generates geographic hotspots for where we are likely to find earthworks and ADEs in the future.”

• AssumptionRests on an assumption: The geographic distributions of ancient Amazonian archaeological sites are predictable.

The overall model performance is still quite high.

“ADEs and other sites show some geographic overlap making it more difficult for the model to separate the two, although the overall model performance is still quite high.”

The model finds it unlikely that new finds will greatly increase the known extents of earthworks and ADEs across western and northern regions of Amazonia.

“While it is possible that new finds will greatly increase the known extents of earthworks and ADEs across western and northern regions of Amazonia, our model finds this scenario unlikely.”

• AssumptionRests on an assumption: The geographic distributions of ancient Amazonian archaeological sites are predictable.

The model is less likely to confuse earthworks with the other two site types, with only 2% misclassified.

“The model is less likely to confuse earthworks with the other two site types (only 2% misclassified).”

Using model probabilities as a guide, the authors found 13 geoglyphs that were not included in the training data.

“With model probabilities as a guide to specific areas warranting further investigation, we found 13 geoglyphs that were not included in training data.”

• AssumptionRests on an assumption: The geographic distributions of ancient Amazonian archaeological sites are predictable.

In the spatially cross-validated confusion matrix, the model is slightly prone to confuse ADEs with other sites, with a 4% misclassification rate averaged across spatial folds.

“In the spatially cross-validated confusion matrix, the model is slightly prone to confuse ADEs with other sites (4% misclassification rate averaged across spatial folds).”

The authors have now discovered 10 geoglyphs using LiDAR data originally purposed for aboveground biomass models.

“we have now discovered 10 geoglyphs in this exact way”

The numbered areas in Fig. 5 overlap with the listed protected areas.

“In fact, the numbered areas in Fig. 5 overlap with the following protected areas: (1) Parque Estadual Chandless, (2) Reserva Extrativista Ituxí, (3) Reserva Nacional de Vida Silvestre Amazónica Manuripi, (4) Parque Estadual Igarapés do Juruena, (5) Parque Estadual do Matupiri, (6) Parque Nacional do Jaú, (7) Floresta Nacional do Amazonas, and (8) Floresta Nacional de Seracá-Taquera.”

• AssumptionThe authors state this as an assumption: Current non-forested areas have already been discovered, so archaeological sites there are likely already found.

According to the model, earthworks and ADEs may be restricted to less than one-third of Amazonia.

“may be restricted to less than one-third of Amazonia, according to our model.”

• AssumptionRests on an assumption: The geographic distributions of ancient Amazonian archaeological sites are predictable.

Summing up the number of cells using the model predictions shows that other sites dominate more than two thirds of all Amazonia.

“In fact, summing up the number of cells using the model predictions show that other sites dominate more than two thirds of all Amazonia.”

Other non-earthwork/non-ADE sites are expected or predicted to predominate most of Western and Northern Amazonia and essentially everywhere else outside earthwork/ADE areas.

“The limited, albeit impressive, predicted extents of earthworks and ADEs means that other non-ADE/non-earthwork sites are expected to predominate most of Western and Northern Amazonia.”

• Also stated in: § Results · plain

  “Less obvious from the original occurrence data is that other non-earthwork/non-ADE sites are predicted by the model to dominate essentially everywhere else, including most of western and northern Amazonia.”

Seasonality is related to ease of forest clearing and burning given the approximately three-month dry season of the southern rim that contrasts with less seasonality in most of the Amazon Basin.

“Seasonality is related to ease of forest clearing and burning given the approximately three-month dry season of the southern rim that contrasts with less seasonality in most of the Amazon Basin.”

Conventional random cross-validation unrealistically overestimates model performance (AUC > 0.99) because spatially structured data with random cross-validation severely underestimates prediction error.

“Conventional random cross-validation unrealistically overestimates model performance (AUC > 0.99) because spatially structured data with random cross-validation has the effect of severely underestimating prediction error (Roberts et al., 2017).”

• cites Roberts et al., 2017 sourceNot checked — no full text of the source

Across the southern rim of Amazonia, over 1,000 earthwork sites comprised of fortified settlements, mound villages, and ditched enclosures with geometric designs known as geoglyphs have been discovered.

“Across the southern rim of Amazonia, over 1,000 earthwork sites comprised of fortified settlements, mound villages, and ditched enclosures with geometric designs known as geoglyphs have been discovered.”

• Also stated in: § Introduction · boosted

  “Southwest Amazonia (Upper Madeira River) has over 1,000 earthwork sites clearly visible in satellite imagery (Pärssinen et al., 2020; Saunaluoma & Schaan, 2012).”

• cites Pärssinen et al., 2020 sourceCheck pending
• cites Saunaluoma & Schaan, 2012 sourceCheck pending

The authors confirmed the geolocations of all 1,100 earthwork sites in satellite imagery.

“Starting with an online dataset (http://jqjacobs.net/archaeology/geoglyph.html), we confirmed the geolocations of all 1,100 sites in satellite imagery, including the well-documented earthwork clusters in Acre (Saunaluoma, Pärssinen & Schaan, 2018), Upper Xingu (Heckenberger et al., 2003), Upper Tapajós (de Souza et al., 2018), and Llanos de Moxos in Northern Bolivia (Prümers et al., 2022).”

Model testing is done with spatial cross-validation.

“Model testing is done with spatial cross-validation, and the best model at the optimal spatial scale of 1 km has an Area Under the Curve of 0.91.”

The lower misclassification of earthworks likely results from the tighter spatial clustering of earthworks making them more predictable in space.

“This likely results from the tighter spatial clustering of earthworks making them more predictable in space.”

LiDAR studies show expansion of settlements and structures extending considerably farther in geographic extent than previously established through conventional fieldwork.

“Like our machine learning model, these LiDAR studies show expansion of settlements and structures extending considerably farther in geographic extent than what had been previously established through conventional fieldwork.”

• cites de Souza et al., 2018 sourceCheck pending
• cites Iriarte et al., 2020bNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Stenborg, Schaan & Figueiredo, 2018 sourceNot checked — no full text of the source

Several geoglyphs were discovered in Acre, Brazil from LiDAR data originally purposed to develop aboveground biomass models in forests.

“Several geoglyphs were discovered in Acre, Brazil from LiDAR data originally purposed to develop aboveground biomass models in forests (Wagner et al., 2022)”

• cites Wagner et al., 2022Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

LiDAR mapping of ancient archaeological features hidden under forest canopy has proven to be an important discovery tool.

“LiDAR mapping of ancient archaeological features hidden under forest canopy has proven to be an important discovery tool.”

• cites Chase et al., 2011 sourceNot checked — no full text of the source

ADEs and other sites show some geographic overlap making it more difficult for the model to separate the two.

“ADEs and other sites show some geographic overlap making it more difficult for the model to separate the two, although the overall model performance is still quite high.”