Beginning with Leslie Spiers’s (Reference Spier1917) work at Zuni in the early nineteenth century, survey has been a part of the archaeological tool kit (Ammerman Reference Ammerman1981). Since then, survey has changed considerably from its most traditional form—walking through a survey area and mapping or recording surface features. By the mid-twentieth century, archaeologists began studying settlement at the landscape scale, leading to new analytical approaches, including predictive modeling and site sensitivity analysis (Clarke Reference Clarke1977; Hodder and Orton Reference Hodder and Orton1976; Judge and Sebastian Reference Judge and Sebastian1988; Kvamme Reference Kvamme2006; Willey Reference Willey1953; Williams et al. Reference Williams, Thomas and Bettinger1973). Not long thereafter, aerial imagery was adopted to reveal features that were hard to see from the ground surface (Kincaid Reference Kincaid1983), and powerful technologies—from satellite imagery to geographic information systems (GIS)—were incorporated to accelerate the rate and interpretive insights gained from survey and site prediction (Kvamme Reference Kvamme2006; Parcak Reference Parcak2009; Romano and Tolba Reference Romano, Tolba, Kamermans and Fennema1996). The most recent evolution of archaeological survey involves computational predictive modeling and machine learning (Castiello and Tonini Reference Castiello and Tonini2021; Green et al. Reference Green, Altschul, Heilen, Cushman, Homburg, Lerch, Hayden and Trampier2012; Kvamme Reference Kvamme2006; Resler et al. Reference Resler, Yeshurun, Natalio and Giryes2021; Yaworsky et al. Reference Yaworsky, Vernon, Spangler, Brewer and Codding2020). In this article, we focus specifically on a form of machine learning called computer vision (CV), which uses neural networks to emulate humans’ ability to detect visual patterns in imagery through robust computation and which has grown rapidly in popularity within archaeology (Bellat et al. Reference Bellat, Figueroa, Reeves, Taghizadeh-Mehrjardi, Tennie and Scholten2025; Bickler Reference Bickler2021; Davis Reference Davis2019; El Naqa and Murphy Reference El Naqa and Murphy2015; Green et al. Reference Green, Altschul, Heilen, Cushman, Homburg, Lerch, Hayden and Trampier2012; Heilen et al. Reference Heilen, Leckman, Byrd, Homburg and Heckman2013).
Each evolution of archaeological survey involves balancing the proven value of established methods with the promise(s) of any new technology or technique. Often, such a balance involves identifying, and encouraging, the analytical practices that are shared between established and new methods. For example, remote survey requires balancing the intuition born from years of field experience with the unprecedented spatial scale of remote sensing. Fortunately, pedestrian and remote survey share many attributes. Whether a researcher is walking through a landscape or looking at aerial imagery, repeated exposure enables attunement to the inherent variation or patterns of an area of observation. Whether debating written descriptions in the field or the quality of a signature in an image, corroborating independent opinions allows for biases to be critically examined. Given the commonalities between remote and pedestrian survey, it is possible to preserve the best tendencies of pedestrian survey while capitalizing on the value of remote sensing through a research program that prioritizes independent interrogation of data by well-trained users (Casana Reference Casana, Lawrence, Altaweel and Philip2020, Reference Casana2021; Parcak Reference Parcak2009).
We aim to demonstrate a similarly balanced approach using image classification, a CV technique that makes individual classifications for entire images (e.g., classifying site presence in an image). Through this approach, we attempt to maintain the interpretive and analytically critical tendencies built from manually sorting through data with the promise of systemization and efficiency. To illustrate this approach, we conduct an image classification survey for residential sites in the US Southwest. Results show how image classification can be integrated into remote surveys for efficient and comprehensive, yet still flexible and interpretive, archaeological prospection. Our approach can be useful for similar prospection tasks, when planning and management decisions depend on understanding the regional distribution of particular site types. However, like all predictive models and emergent technologies, the limitations of an image classification survey are considerable, and the results of this study do not support the reduction or negation of traditional survey. Rather, this research supports the addition of critically developed machine-learning techniques as a valuable tool in an expanded archaeological tool kit.
The Value of Image Classification for Large-Scale Survey and Prediction
For research involving the detection of signatures in static imagery, the two most common CV approaches depend on image classification, which assigns a single class to an individual image based on its overall content—or object detection—which identifies local objects of interest within an image (Argyrou and Agapiou Reference Argyrou and Agapiou2022; Bellat et al. Reference Bellat, Figueroa, Reeves, Taghizadeh-Mehrjardi, Tennie and Scholten2025; Bickler Reference Bickler2021; Davis Reference Davis2019; Fiorucci et al. Reference Fiorucci, Khoroshiltseva, Pontil, Traviglia, Bue and James2020, Reference Fiorucci, Verschoof-van der Vaart, Soleni, Saux and Traviglia2022; Jamil et al. Reference Jamil, Yakub, Azizan, Roslan, Zaki and Ahmad2022; Kadhim and Abed Reference Kadhim and Abed2023; Olivier and Verschoof-van der Vaart Reference Olivier and Vaart2021; Sevara et al. Reference Sevara, Pregesbauer, Doneus, Verhoeven and Trinks2016). To illustrate the differences between these approaches, imagine a research design that uses CV to identify ceramic sherds in a midden (Figure 1; see Agapiou et al. [Reference Agapiou, Vionis and Papantoniou2021] and Orengo and Garcia-Molsosa [Reference Orengo and Garcia-Molsosa2019] for examples).
Example of image classification and object detection approaches for an autonomous survey for ceramic sherds in high-resolution midden imagery. (Color online)
Figure 1 Long description
Insufficient visual information to describe this element accurately.
Both approaches conduct a remote survey, build an orthoimage, and then make predictions about the feature contents within the image (e.g., the presence and location of sherds). However, the methods produce slightly different outputs. An image classification approach produces a series of discretely classified subimages, each representing a portion of the midden and an associative probability of sherd presence (i.e., positive return) and absence (i.e., negative return). An object detection approach produces a series of masked locations (e.g., discrete polygons) containing a probability of sherd presence (all nonmasked locations are assumed to indicate sherd absence). Following model classification, some form of validation is conducted, often involving a human evaluating positive and negative returns to identify locations that actually contain a sherd. At this stage, it is imperative that the researcher possess a well-informed internal model of signal variations, patterns, and visual cues. Without these, the human validation is likely to misinform, mischaracterize, or misunderstand how the model is actually performing.
In best-case scenarios, well-informed researchers can use CV to survey large areas with unprecedented speed while also maintaining decision-making over model performance and outputs. This integration will help archaeologists produce more complete records of prehistoric settlement and human behavior, which has been a challenge due to complicated patchworks of land ownership and incomplete survey coverage. Yet, the integration of CV into archaeological research also poses risks. For example, in the coming decades, when professional experiences of the next generation are mediated through an increasing amount of automation, it may be harder to train human researchers who are well informed. Therefore, it is paramount that the discipline emphasizes repeated, firsthand exposure to material culture and landscapes (preferably in both field settings and remote sensing) and encourages CV-enabled work that keeps researchers in the loop. To illustrate such an approach, we conduct a research program involving iterative stages of researcher engagement, which results in a CV model that can identify residential sites in the Northern San Juan (NSJ) region of the northern US Southwest.
Case Study
The NSJ encompasses an area of approximately 16,000 km2 in southwestern Colorado, southeastern Utah, and northwestern New Mexico (Figure 2). Indigenous occupation of NSJ extends back thousands of years. Some of the most enduring surface signatures of Indigenous occupation date to around the mid-eighth through thirteenth centuries and are a result of ancestral Puebloans building aboveground masonry pueblos (i.e., room blocks). Pueblos in the NSJ have a consistent architectural template that is characterized by three central features—a pueblo, a circular subterranean structure (i.e., pit structure or kiva), and a midden—arrayed in a north–south orientation (Lipe Reference Lipe, Lipe and Hegmon1989; Prudden Reference Prudden1903). From the eighth through thirteenth centuries, political, social, and demographic changes influenced architectural preferences in the NSJ. However, the general form of pueblos persisted, so until ancestral Puebloan emigration circa AD 1280, north–south-aligned room blocks, pit structures / kivas, and middens were the most common architectural configuration for residences.
Extent of the Northern San Juan region, with examples of ancestral Pueblo residential surface signatures (lower-left panels) taken from Mesa Verde National Park. (Color online)
Figure 2 Long description
The map shows the Northern San Juan region, highlighting ancestral Pueblo residential sites and Mesa Verde National Park. The main map outlines the Northern San Juan area across southwestern Colorado, southeastern Utah and northwestern New Mexico. A smaller inset map indicates the location of the region within the states of Utah, Colorado, Arizona and New Mexico. The map includes labeled areas such as National Park Service Monuments and the Ute Mountain Ute Reservation. The lower-left panels display close-up images of ancestral Pueblo residential surface signatures, characterized by circular and rectangular features, likely representing room blocks and kivas. These are shown at a scale of 50 meters. The bottom right includes a topographic map of Mesa Verde National Park with a scale of 5 kilometers. An arrow in the bottom right indicates north. The map provides a detailed view of the region′s archaeological and geographical features.
Residential features are an important line of evidence for archaeological research in the NSJ; room blocks with a single pit structure / kiva are commonly used as a proxy for single households (see Bocinsky et al. Reference Bocinsky, Rush, Keith and Timothy2016; Kohler and Varien Reference Kohler and Varien2012; Lipe Reference Lipe, Lipe and Hegmon1989; Schwindt et al. Reference Schwindt, Bocinsky, Ortman, Glowacki, Varien and Kohler2016; Varien et al. Reference Varien, Ortman, Kohler, Glowacki and Johnson2007), and larger residences with communal architecture evidence the social, political, and demographic anchors of ancestral Pueblo networks (Glowacki Reference Glowacki2015; Glowacki and Ortman Reference Glowacki, Ortman, Kohler and Varien2012; Lipe and Ortman Reference Lipe and Ortman2000; Varien Reference Varien1999). Despite more than four decades of intensive archaeological research in the NSJ (see Blinman et al. Reference Blinman, C. Dean, Robert, Mary and Linda1984; Breternitz Reference Breternitz1993; Glowacki Reference Glowacki2015; Kohler and Varien Reference Kohler and Varien2012; Ryan Reference Ryan2023; Schwindt et al. Reference Schwindt, Bocinsky, Ortman, Glowacki, Varien and Kohler2016; Varien et al. Reference Varien, Ortman, Kohler, Glowacki and Johnson2007), gaps remain in the survey record. Moreover, the region contains a complicated array of land management, including various state and federally managed lands, conservation and wilderness areas, and private property, which means that a comprehensive pedestrian survey of the region is nonexistent and unlikely to become a reality.
Fortunately, room blocks and pit structures / kivas can be clearly visualized in 1 m lidar-derived digital elevation models (DEMs), especially after visual transformations (Figure 2). The existence of common architectural forms that can be detected in remote sensing make the NSJ an ideal case study for an image classification CV-enabled survey. Many neighboring regions—including the Cibola, Kaibab, Gallinas, Rio Grande, and Mogollon regions—also benefit from highly visible and relatively well-preserved surface signatures, so they could be equally conducive for CV-enabled survey. However, because the majority of our field-based research is in the NSJ, it is only within the NSJ that we are confident that signatures in remote sensing correspond to features on the ground. It is also worth noting that our approach is unlikely to work in other research contexts where sites lack quality surface signatures or require invasive techniques for site identification (such as shovel testing). Therefore, this case study is illustrative of a best-case application, rather than the utility of CV approaches everywhere.
Methods
The workflow (Figure 3) used in this study involved four primary steps: (1) train a model to classify the presence or absence of a site within an image using known site signatures from Mesa Verde National Park (MVNP), (2) deploy the model over a subset of the NSJ, (3) retrain the model using results from the first deployment, redeploy the model over the entire NSJ, and (4) examine outputs. The details involved in each step are reviewed below.
Computational workflow for detection of ancestral Pueblo residential surface signatures in lidar-derived digital elevation model (DEM) imagery.
Figure 3 Long description
The flowchart outlines a computational workflow for detecting ancestral Pueblo residential surface signatures in lidar-derived digital elevation model (DEM) imagery. It consists of four main steps: 1. Initial Model Training and Deployment: This involves data preparation using USGS DEM and MVNP site signatures, building training data and deploying a model to classify site presence. 2. Initial Output: Using QGIS, the model output is visualized and potential sites are identified. 3. Retrain and Redeploy: The model is retrained with expanded training data, including new site signatures and redeployed. 4. Output: The final model output is visualized in QGIS and potential sites are identified for further analysis. Each step includes specific tasks and tools, such as R, Python and QGIS, to process and analyze the data effectively.
Step 1
Mesa Verde National Park was defined as the training area due to dense ancestral Pueblo occupation, as well as a long history of research and cultural landscape management that has resulted in a detailed site record. Images of known site locations in MVNP (∼4,000 images) were derived from publicly accessible 1 m lidar-based DEMs transformed using the Sky-View Factor from the Relief Visualization Toolbox in QGIS (Kokalj and Somrak Reference Kokalj and Somrak2019; QGIS Development Team 2026; US Geological Survey 2025; Verbovšek et al. Reference Verbovšek, Popit and Kokalj2019; Zakšek et al. Reference Zakšek, Oštir and Kokalj2011). The Sky-View Factor was conducted with a search radius of 10 m, 16 search directions, and a vertical exaggeration factor of 5. An automated procedure was developed in R (R Core Team 2025) to produce each image, which was centered on an individual site location and bounded by an area of 250 × 250 m. All images were manually evaluated to ensure they contained a clear signature. Given the site density in MVNP, some images contained multiple sites. In other instances, where a site was located on the edge of an image or grid, multiple images referenced the same site. Therefore, the number of images representing sites is not an exact reflection of site count or density. Additionally, most signatures correspond to sites that date to later occupation (ca. AD 950–1250), and as a result, the model has a significant temporal bias. For example, Indigenous occupation of the NSJ began thousands of years ago, and much of this history—from Paleoindian and Archaic to historic occupations—left material traces that are not detectable to this survey (Copeland and Fike Reference Copeland and Fike1988; Pitblado Reference Pitblado1998; Tune Reference Tune2020). Furthermore, there are important components of ancestral Puebloan occupation that are unaccounted in this survey, such as cliff dwellings, due to the site being occluded from aerial imaging.
After image creation and evaluation, our database included 1,108 site images. Three thousand five hundred locations within the park that did not contain a site were randomly identified, and corresponding images were produced to represent the “non-site” class. Non-site images were also evaluated to ensure that no signatures were present.
Training images from both classes were fed into a PyTorch pipeline, a python-based open-source machine-learning framework widely used in computer vision (Paszke et al. Reference Paszke, Gross, Massa, Lerer, Bradbury, Chanan and Killeen2019). Training processes leveraged foundational supervised learning techniques, beginning with feeding 80% of the labeled images—both site and non-site examples—into pretrained deep-learning algorithms that rely on stochastic optimization (Kingma and Ba Reference Kingma and Ba2014). Initial training resulted in a model that performed at 93.5% accuracy, correctly classifying 437 of the 467 test images (i.e., images from the initial training dataset that were not fed into the model during training).
Following initial training, the model was deployed across a roughly 10,000 km2 area that encompasses the central part of the NSJ. The area was drawn beyond the extent of the NSJ to understand how the model performs in landscapes with significantly different topographies from that of MVNP. The deployment area was divided into grid points at approximately 150 m intervals, and static images for each grid point were derived in R. Each image was centered on an individual grid point and bounded by an area of 250 × 250 m. Through this approach, adjacent images had a high amount of overlap, ensuring that any site on the edge of an individual image was likely to be fully included in adjacent imagery and therefore easily viewed during survey. In total, 436,211 static images were derived for the subarea. The model was then deployed to classify likelihood of a site signature in every image. An image with a confidence value of 0 indicates a 0% likelihood of site presence, whereas a value of 1 indicates a 100% likelihood of site presence, according to the model. Early results produced a wide range of confidence estimates, and although many were low (<10%), there were more than 100,000 images with >50% of site presence, which after a cursory review were shown to consistently be false positives (FP; incorrect classification of site presence). Therefore, a very high threshold of 95% confidence was used to differentiate between images that were sorted into “site” and “non-site” classes (e.g., any image with a >95% likelihood of site presence was sorted into the “site” class, and all others were sorted into the “non-site” class). Developing the initial deployment data and conducting the first survey took approximately 13 hours of total computing time, using the 19.2-petaflops Derecho supercomputer at the National Center for Atmospheric Research (NCAR) Wyoming Super Computing Center (NWSC).
Step 2
The initial model classified 7,711 images as containing a site signature (a 1.8% positive return rate). To validate these results, all images classified as a “site” were manually examined to determine if a site’s signature was present or not. Site signatures were corroborated against satellite imagery, available site records, and/or legacy geodatabases to increase the likelihood that remotely identified site locations corresponded to legitimate cultural resources. For all validated site signatures, the location of the site was recorded in a geodatabase, and the image was tabulated as a true positive (TP). Images that were classified as a site but that did not include a signature were recorded as false positives (FPs). Validating negative returns involved a less comprehensive protocol, given the time costs of manually validating the other 428,500 images. Instead, approximately 1,000 “non-site” images were randomly selected and validated as an estimate of the true negative (TN; the percentage of instances in which an image was classified as “non-site” and corresponded to an area without a signature) and false negative (FN; the percentage of instances in which an image was classified as “non-site” but corresponded to an area with a signature) rate. When an FN image was identified, the location of the site was appended to the geodatabase.
Initial survey results indicated significant room for improvement. Estimated recall, or the proportion of sites that were classified correctly, was high (0.92), which means that the model performed relatively well in detecting site locations. However, precision, or the portion of positively identified images that correspond to actual sites, was exceedingly low (0.15). Consequently, although the model detected most site locations, it also routinely assigned high confidence of site presence in an image that did not contain a site. As a result, the actual performance of the model was very low (F1 = 0.26, MCC = 0.05; for metric equations and discussion, see Chicco and Jurman Reference Chicco and Jurman2020; Gravel-Miguel et al. Reference Gravel-Miguel, Snitker, Hirniak, Peck and Fetterhoff2025). Due to low performance, additional steps were taken to improve the quality of the model.
Step 3
To improve the model, the training dataset was expanded using the validated data from Step 2 as well as an image augmentation technique. Specifically, translation augmentation was used to derive multiple, slightly offset images for each site from the initial MVNP training dataset and each site identified during the validation processes in Step 2 (i.e., TP and FN identified in Step 2). This process increased the number of images in the “site” class to 3,484. Training images for the “non-site” class were expanded to 7,982 images and included a wider range of non-site locations from across the central part of the NSJ.
The initial model was retrained on the expanded training data, using the same general workflow outlined in Step 1 and resulting in a model that performed at 92.8% accuracy during training. Next, images for the entire NSJ region were derived (n = 774,836) and classified. Again, a 95% confidence threshold was used to differentiate between images that were sorted into “site” and “non-site” classes. Processing was conducted on the NCAR Derecho supercomputer, requiring about 22 hours of cumulative processing time to retrain the model and deploy it for the entire NSJ.
Step 4
Outputs from Step 3 were validated following the protocol laid out in Step 2. However, an additional step was taken to better estimate the TN rate. Five thousand “non-site” images were randomly pulled from the “non-site” class and visually inspected to identify if a signature was present. Results from the final deployment demonstrate marked improvement (Table 1).
Classification Outputs of the Complete Survey of the NSJ.
Table 1 Long description
The table reports image classification results for two deployments, listing total images classified, how many were validated, confusion-matrix counts, and performance metrics. Deployment 1 classified 436,211 images but only 8,811 were validated; within validated images it has 1,143 true positives, 6,568 false positives, 1,000 true negatives, and 100 false negatives, yielding recall 0.92, precision 0.15, accuracy 0.24, and F1 0.26. Deployment 2 classified 774,836 images with 10,197 validated; within validated images it has 3,364 true positives, 1,833 false positives, 4,991 true negatives, and 9 false negatives, giving recall greater than 0.99, precision 0.65, accuracy 0.82, and F1 0.79. Compared with deployment 1, deployment 2 greatly reduces false positives and false negatives and improves overall accuracy and F1. A third row estimates deployment 2 performance if the validated true and false negative rates also apply to unvalidated images, producing very large true negative counts, a false positive rate below 0.01, accuracy greater than 0.99, and F1 0.68. Interpretation should note that most images are unvalidated, so the scaled results depend on the assumption that validated performance generalizes to the unvalidated set.
Results
The performance of the retrained model was evaluated based on the TP, FP, TN, and FN counts for all validated images (n = 10,197), as well as TN and FN counts for the entire dataset, assuming that the validated TN and FN rates were constant for all unvalidated images (n = 764,639). In both instances, F1 and MCC scores improved (0.68–0.79). In addition, precision improved from 0.15 to 0.65, and recall ranged from 0.71 to >0.99. However, some of the performance results can be misleading. For instance, recall, when measured only from the validated images, presents inflated performance, because it does not account for the likelihood of additional FN in nonvalidated images. Similarly, when looking at the entire dataset, model accuracy is >0.99. This too is misleading, because it is overly influenced by the high number of assumed TN images (n = 763,263) but does not account for the imbalance in classes (e.g., many more non-site than site images). Because archaeological sites are sparsely distributed, one could throw a dart at a map and as long as one said the dart was not touching a site, one would be right almost every time. In this way, the CV model can be seen as good as (or no better than) a model that assigns only a single class.
Among images classified with >50% confidence of site presence, most had a very high confidence (>95%) of site presence, suggesting that the model discerned site presence using a well-tuned, sensitive threshold (Figure 4). Among all positive returns (n = 5,197) more than 3,300 accurately corresponded to a site, resulting in the identification of 4,905 pit structure / kiva depressions (Figure 5). These results produce a settlement distribution that is comparable to our understanding of ancestral Pueblo settlement in the NSJ (Figure 5). For example, sites were autonomously located in most areas of known occupation—along major rivers, atop southward-sloping mesas such as Mesa Verde, and among the complex canyon systems that range from Bears Ears National Monument to Canyon of the Ancients National Monument (Figure 5). Additionally, outputs resemble existing databases, with the densest concentrations of ancestral Pueblo occupation occurring in the middle of the NSJ on either side of the Colorado/Utah border (see Glowacki Reference Glowacki2015; Glowacki et al. Reference Glowacki, Coffey, Varien and Ryan2023). Correspondence was measured by comparing the results of the CV survey to data from the Village Ecodynamics Project (Kohler and Varien Reference Kohler and Varien2012; Schwindt et al. Reference Schwindt, Bocinsky, Ortman, Glowacki, Varien and Kohler2016), which is one of the most comprehensive geodatabases from the NSJ. Among a subset of the larger VEP sitesFootnote 1 (n = 70), slightly more than 75% were identified by the model. The most informative accuracy metrics (F1, MCC; Table 1) and the performance of the model against know site distributions suggest that the model performed at about 70% accuracy when deployed across the NSJ.
Distribution of confidence values for all images. All images receiving a value ≥0.95 confidence were attributed a “site” classification. (Color online)
Spatial distribution of all images classified with a high site presence. Warmer heat signature indicates a higher likelihood of classified site presence, and cooler heat signature indicates lower likelihood of classified site presence. Gray dots indicate location of probable site signature that was manually validated. Precise site locations are not visualized in Reservation/Trust Land due to data sovereignty concerns. (Color online)
Figure 5 Long description
The map of Northern San Juan illustrates the likelihood of model classifying site presence, ranging from low to high. Areas with high likelihood are concentrated in the central region, while low likelihood areas are in the western part. A verified site signature is marked in the eastern section. Reservation/Trust Land is indicated centrally. The map includes a scale bar showing 0 to 50 kilometers and a legend explaining the gradient from low to high likelihood. An arrow points north for orientation.
One of the most compelling results from this work is that the research design encouraged the flexibility and interpretability of traditional survey without overly sacrificing the efficiencies of CV. Throughout the research process, tens of thousands of images were manually examined, allowing lab members to build their own internal models and intuitions about site signatures in lidar data. Also, the recursive nature of the workflow—which involved repeated training, validation, and deployment—meant that lab members were consistently sharing insights and cross-examining validation classes, leading to a more critical and trustworthy research output. This process directly reflects the same dynamics that play out in field research, including working through model results, discussing interpretive outcomes, and evaluating accuracy through robust interrogation of the data. At the same time, remote surveys using this approach can be accomplished at an unprecedented speed; tens of thousands of square kilometers can be surveyed autonomously and validated remotely in a matter of days.
However, it is worth noting that the benchmarks achieved in this study were not immediate. In total, approximately 100 person hours were spent writing code, validating model results, and establishing the computational workflow. Although this is a considerable investment, it represents a fraction of the cumulative costs of conducting the same survey on foot or through a manual review of the entire digital dataset.
Despite these benefits, there is much this approach (and any other form of CV-enabled survey) cannot do. Not only are vast periods of occupation left unassessed, but critical components of traditional archaeological surveys—such as evaluating National Register of Historic Places (NRHP) eligibility—cannot be accomplished from a computer. The results of our study show only ancestral Puebloan occupations, and more specifically, occupation of sites that likely date to Pueblo I–Pueblo III (AD 700–1300). This 600-year window represents only a fraction of the approximately 13,000 years of occupation in the NSJ. Furthermore, our results are also not reflective of the variety of site types within this window (e.g., kilns, storage features, agricultural features, rock imagery, to name a few). Therefore, these results emphasize the utility of CV-enabled survey as a tool that can amplify but should not replace traditional archaeological techniques and protocol. As just one example of how CV-enabled survey can be used to augment traditional methods and objectives, results of this survey (such as the heatmap in Figure 5) could be fed into the site sensitivity models that inform large-scale management plans or land-use decisions (Doelle et al. Reference Doelle, Barker, Cushman, Heilen, Herhahn and Rieth2016).
Conclusion
As is often the case with powerful technologies, adoption can outpace the development of best practice. In pursuit of best practice, we have demonstrated a CV-enabled survey that identified ancestral Pueblo residential sites with relative accuracy, but not without active engagement by the researcher. As with other recent CV applications in archaeology, the results of this study reiterate how CV can be used to conduct region-scale survey efficiently and systematically. The case study also highlights how image classification techniques can be integrated into a remote survey design while maintaining the methodological tendencies that enable interpretable and trustworthy results.
Acknowledgments
We are deeply indebted to the Mesa Verde National Park Cultural and Natural Resources archaeologists and naturalists. Their expertise and support—especially that of Christine McAllister, Elizabeth Dickey, Carole Graham, Sheldon Baker, and Kay Barnett—have been instrumental in this project. We also thank Katherine Peck for helping develop ideas related to this research. The research was carried out under National Park Service permit MEVE-2024-SCI-0005 and was made possible through the National Center for Atmospheric Research Wyoming Supercomputing Center.
Funding Statement
This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. All authors were supported in part by internal funding from the University of Wyoming.
Data Availability Statement
All code used to conduct the analysis is available at https://github.com/sfield2/NSJ_computervision/. Lidar-derived DEMs are available at https://apps.nationalmap.gov/lidar-explorer/#/. Due to sensitive site locations, all site location data used to train the model and site location outputs from the model are not available. The model output is available on request from the corresponding author (Sean.Field@uwyo.edu).
Competing Interests
The authors declare none.
Open access
