Project Summary

• This exploratory study assessed the applicability of Geospatial Foundation Models (GFMs), specifically TerraMind and RemoteCLIP, for archaeological site identification and segmentation under label scarcity.
• A 7-band multimodal dataset was built to map archaeological settlement mounds (tells) in Central Iraq, combining Sentinel-2 optical (RGB–NIR), Sentinel-1 SAR (VV–VH), and the Copernicus Digital Elevation Model (COP-30, resampled to 10 m), with annotations from expert archaeologists.
• GFMs were evaluated against task-specific baselines (vanilla U-Net and Vision Transformer) trained from scratch across reduced label percentage (100% down to 5%).
• Preliminary results on a mini-dataset (~20% of the total study area coverage) suggest advantages from GFMs in low-data regimes, with multimodal inputs (SAR + optical + elevation) outperforming single-modality baselines.

Development Tools

• PANGAEA evaluation protocol: a benchmarking framework and codebase by Marsocci et al. for evaluating GFMs under label-scarcity regimes.
• TerraMind: a multimodal GFM developed by IBM Research and ESA Φ-lab.
• Data access and sources: The research uses Copernicus data collections as base data, accessed through the Google Earth Engine Python API. Ground-truth annotations are from Chiricallo et al., 2026.

Development Outputs

Building on these initial outcomes, ongoing and future research will focus on:

1. Scaling up: Expanding training and evaluation to the complete multimodal dataset across the full study area.
2. Comprehensive metrics: Incorporating object-level (F1-score) and tile-level evaluation metrics alongside standard pixel-wise IoU.
3. Domain shift testing: Evaluating model transferability and robustness on an out-of-distribution test region in Iran featuring distinct geomorphological characteristics.
4. Precomputed embeddings and task expansion: Exploring the applicability of GFMs and precomputed geospatial embeddings (e.g., TESSERA and AlphaEarth) to other archaeological prospection tasks.

This exploratory project was conducted at ESA Φ-lab under the supervision of Dr. Peter Naylor. Preliminary findings were presented to ESA Φ-lab researchers during a final project presentation, while subsequent project updates were shared via a poster presentation at ESA Φ-Innovation on June 24, 2026.

Project Description

Deep learning has become a valuable tool for Earth Observation analysis, but its adoption in archaeological prospection remains constrained by data scarcity and class imbalance. Traditional supervised deep learning approaches typically require large, high-quality labeled datasets, which are often unavailable in archaeological research or are time-consuming to produce. Archaeological objects are rare and sparsely distributed, vary in appearance and seasonal visibility, and may exhibit fuzzy boundaries due to gradual transitions between archaeological space and the surrounding landscape.

This exploratory work investigates GFMs as a promising tool for archaeological remote sensing, assessing whether their globally pretrained knowledge can transfer effectively to specialized tasks of archaeological site identification and segmentation under limited label availability.

The case study focuses on ancient settlement mounds, i.e., tells, in Central Iraq (Southern Mesopotamian Floodplain). Tells are Neolithic and Bronze Age artificial mounds in Near and Middle Eastern landscapes, formed by the accumulated and stratified remains of ancient settlements. They are characterized by an elliptical shape, a distinct topographic prominence above the surrounding plain, and specific spectral and microwave backscatter signatures linked to distinctive soil composition (anthrosols) and land use.

To capture these characteristics, a 7-band multimodal dataset was built using the Google Earth Engine Python API, consisting of:

• RGB + NIR bands from Sentinel-2 Level-2A multitemporal mean reflectance images (years 2021–2023).
• VV + VH bands from Sentinel-1 IW GRD multitemporal mean amplitude images (years 2021–2023).
• Heightmap band from the Copernicus Digital Elevation Model (COP-30 DEM), resampled to 10 m resolution.
• Ground truth annotations from prior work (Chiricallo et al., 2026).

Following the PANGAEA evaluation protocol, **TerraMind** (base and large configurations) and RemoteCLIP models were evaluated as feature extractors with frozen encoder weights coupled to a trainable UPerNet decoder. These were benchmarked against supervised U-Net and Vision Transformer baselines trained from scratch on the same multimodal inputs, using Intersection over Union (IoU) for the target class (i.e., Tell Site pixels) as the evaluation metric. To examine performance under label scarcity, training was performed across reduced subsets (100%, 80%, 50%, 25%, 10%, and 5% of the original training labels). Model training and experiments were conducted on the computing workstations of the Center for Cultural Heritage Technology (Istituto Italiano di Tecnologia).

Examples of inferences across the validation areas for three of the tested models. The Sentinel-2 images are visualized using the B4, B3, and B2 bands (RGB) while Sentinel-1 images are visualized using the VV, VH, and VV bands. The COP-30 DEM is shown as a grayscale image. In the visualization, purple regions represent the intersection (overlap) between ground truth (blue) and predicted masks (red)

The results reported here represent an initial exploratory phase conducted on a mini-dataset (~20% of the total study area coverage) and evaluated on validation data to establish feasibility and performance trends. The main findings are:

• Label efficiency: TerraMind consistently outperforms baselines across all regimes, from standard to very low-data settings, demonstrating strong representation quality and robust feature extraction capabilities for archaeological objects even with minimal training examples. These results suggest that foundation models can reduce the amount of manual labeling needed to map archaeological sites from satellite data.
• Multimodality: While topographic elevation (COP-30 DEM) is a primary driver for identifying mound relief, integrating optical (Sentinel-2) and radar (Sentinel-1) bands provides key spectral and texture cues, improving both overall site detection and boundary segmentation quality over single-modality baselines.