AI in Archaeology

AI in Archaeology

Artificial intelligence is revolutionizing archaeology — the study of human history through material remains. From detecting buried sites with satellite imagery to deciphering ancient scripts, AI enables archaeologists to analyze data at scales and resolutions previously impossible, while preserving access to fragile cultural heritage. The field sits at the intersection of computer vision, natural language processing, remote sensing, and cultural informatics.

Remote Sensing and Site Detection

Satellite and Aerial Imagery Analysis

Archaeological sites often leave subtle surface signatures — crop marks, soil discolorations, geometric earthworks — visible from above but invisible at ground level. Deep learning dramatically accelerates prospection.

LiDAR analysis in dense forest: LiDAR (Light Detection and Ranging) strips away forest canopy to reveal hidden structures. CNNs detect Maya temples, Cambodian reservoirs, and Amazonian geoglyphs in LiDAR point clouds:

import numpy as np
import torch
import torch.nn as nn
from torchvision import models


class LiDARSiteDetector(nn.Module):
    """Detect archaeological features in LiDAR Digital Elevation Models."""

    def __init__(self, num_classes: int = 2):
        super().__init__()
        self.backbone = models.resnet50(weights="IMAGENET1K_V1")
        self.backbone.conv1 = nn.Conv2d(1, 64, 7, 2, 3, bias=False)  # grayscale DEM
        self.backbone.fc = nn.Linear(2048, num_classes)

    def forward(self, dem: torch.Tensor) -> torch.Tensor:
        return self.backbone(dem)


# Example: classify 256×256 DEM tiles as site / no-site
dem_tile = torch.randn(4, 1, 256, 256)  # batch of 4 LiDAR tiles
model = LiDARSiteDetector(num_classes=2)
logits = model(dem_tile)

In 2022, AI analysis of LiDAR data over northern Guatemala revealed over 1,000 previously unknown Maya structures, including causeways and reservoirs spanning hundreds of kilometers.

Multi-Spectral and SAR Analysis

Synthetic Aperture Radar (SAR) penetrates dry sand to detect buried walls and channels invisible in optical imagery. Combined with multispectral bands, machine learning models identify:

• Subsurface soil moisture anomalies indicating buried architecture
• Plough-zone scatters suggesting ancient occupation
• Ancient road networks connecting known sites

# Fusion of optical + SAR for site prediction
import segmentation_models_pytorch as smp

model = smp.Unet(
    encoder_name="resnet34",
    in_channels=8,      # 4 optical bands + 2 SAR polarizations + DSM + NDVI
    classes=1,          # binary: site / background
    activation="sigmoid",
)

Artifact Classification and Dating

Ceramic and Lithic Classification

Pottery is the most abundant archaeological artifact. AI classifies ceramics by period, culture, and function from images of sherds, enabling rapid sorting of excavation finds:

from torchvision import transforms, models
import torch.nn.functional as F

# Fine-tune EfficientNet on labeled ceramic assemblage
transform = transforms.Compose([
    transforms.Resize(224),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

classifier = models.efficientnet_b3(weights="IMAGENET1K_V1")
classifier.classifier[1] = torch.nn.Linear(1536, n_ceramic_types)

3D Artifact Analysis

Photogrammetry and structured light scanning produce 3D meshes. Point cloud deep learning (PointNet++, DGCNN) classifies artifact typologies and detects joins between fragments.

import torch
from pointnet2 import PointNet2SSG

model = PointNet2SSG(num_classes=15)   # 15 lithic reduction stages
# Input: (B, N, 3) point cloud of flint artifact surface
preds = model(point_cloud)

Radiocarbon Calibration with Bayesian ML

OxCal and BChron use Bayesian statistical models to convert radiocarbon measurements into calibrated calendar dates, accounting for atmospheric $^{14}$C variation curves. Machine learning extensions handle stratigraphic ordering constraints automatically.

Paleogeographic and Landscape Reconstruction

Predictive Site Modeling

Logistic regression and random forests trained on environmental variables (elevation, slope, proximity to water, soil type) predict site locations for field survey prioritization:

from sklearn.ensemble import GradientBoostingClassifier
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler

pipeline = Pipeline([
    ("scaler", StandardScaler()),
    ("clf", GradientBoostingClassifier(n_estimators=300, max_depth=5, learning_rate=0.05)),
])

# Features: elevation, slope, TWI, distance to water, aspect, geology code
pipeline.fit(X_train, y_train)   # y: 1 = known site, 0 = background
probabilities = pipeline.predict_proba(X_new)[:, 1]

Paleoclimate Reconstruction

Pollen records, stable isotopes, and faunal assemblages are fed to ML models that reconstruct past temperature and precipitation, linking human settlement patterns to climate change events like the 8.2 kya event or the Bronze Age collapse.

Ancient Script and Language Decipherment

Handwriting Recognition for Ancient Scripts

CNNs and transformers trained on cuneiform, hieroglyphic, and Linear B tablets accelerate transcription of millions of still-unread tablets:

from transformers import TrOCRProcessor, VisionEncoderDecoderModel

processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")
model = VisionEncoderDecoderModel.from_pretrained("cuneiform-ocr/trocr-cuneiform-v1")

pixel_values = processor(tablet_image, return_tensors="pt").pixel_values
generated = model.generate(pixel_values)
text = processor.batch_decode(generated, skip_special_tokens=True)

Partially Deciphered Scripts

The Linear A script (Minoan, undeciphered) and Proto-Elamite are targets for unsupervised and semi-supervised ML approaches. Google DeepMind’s Ithaca model assists epigraphers in restoring damaged Greek inscriptions and attributing texts to geographic origin and historical period with human-AI collaboration.

Network Analysis of Corpora

Graph-based NLP models map semantic networks within ancient corpora — detecting conceptual categories, scribal schools, and intertextual borrowing across thousands of clay tablets.

Bioarchaeology and Physical Anthropology

Skeletal Analysis

Deep learning on CT scans and photographs of skeletal remains estimates:

• Age at death: bone density and epiphyseal fusion stage
• Sex: pelvic morphology classification (95%+ accuracy with 3D CNNs)
• Pathology: identifying fracture patterns, nutritional deficiency markers, infectious disease

Ancient DNA and Population Genetics

Variational autoencoders and admixture models applied to ancient DNA (aDNA) datasets (like those from the Allen Ancient DNA Resource) reconstruct population movements, admixture events, and demographic histories across millennia.

Stratigraphy and Excavation Support

Real-Time Object Detection During Excavation

YOLOv8 models deployed on tablets at the trench-side detect and log artifact occurrences in real time, automatically tagging finds with spatial coordinates from total station data.

from ultralytics import YOLO

model = YOLO("yolov8m.pt")
model.train(data="artifacts.yaml", epochs=100, imgsz=640)

# During excavation: live inference on camera feed
results = model(frame, conf=0.5)
for det in results[0].boxes:
    log_find(det.xyxy, det.cls, gps_coords)

Photogrammetric Documentation

Structure from Motion (SfM) pipelines (COLMAP, Agisoft Metashape) reconstruct 3D site models from drone and hand-held photography. AI automates feature matching and sparse reconstruction steps.

Digital Heritage and Preservation

3D Reconstruction of Destroyed Sites

AI reconstructs damaged or destroyed heritage sites from historical photographs using neural radiance fields (NeRF) and Gaussian splatting:

• Palmyra (destroyed by ISIS): reconstructed from tourist photos using NeRF
• Notre-Dame: structural damage assessed from photogrammetric AI models
• Pompeii: building façade colorization from grayscale archival photographs using conditional GANs

Forgery Detection

Generative models distinguish authentic artifacts from modern fakes by analyzing thermoluminescence data, isotope ratios, and stylometric features in an ensemble classifier.

Key Datasets and Resources

Ethical Dimensions

• Looting risk: published site prediction models could guide illicit excavation — responsible disclosure is essential
• Indigenous sovereignty: communities may object to AI analysis of their ancestors’ remains or sacred sites
• Colonial heritage: many museum collections were acquired under colonial power — AI-assisted repatriation claims analysis is emerging
• Interpretive bias: training data reflects historically dominant narratives; underrepresented cultures may be systematically misclassified

Summary

AI in archaeology spans the full workflow — from satellite prospection and LiDAR analysis to ceramic classification, script decipherment, and digital preservation. Deep learning, remote sensing fusion, and NLP are transforming a traditionally manual discipline into one capable of processing entire landscapes and archival corpora in hours rather than decades. The greatest challenge lies not in the technical capabilities but in ensuring that AI-assisted archaeology remains grounded in archaeological theory, respects indigenous communities, and does not inadvertently endanger the very heritage it seeks to illuminate.