TL;DR
The authors introduce DAFA-LS, the first open-access satellite image time series (SITS) dataset designed specifically for detecting archaeological looting, featuring over 55,000 monthly images of 675 Afghan sites over an 8-year period. By benchmarking various computer vision approaches, they demonstrate that combining remote sensing foundation models with temporal attention mechanisms provides a massive boost in identifying subtle, illicit landscape changes amidst heavy seasonal noise.
What Problem Does This Paper Address?
Archaeological sites are invaluable windows into human history, but they are increasingly threatened by pillaging and conflict, particularly in unstable regions where ground-truth surveys are dangerous or impossible. While satellite imagery offers a viable alternative for remote monitoring, automating the detection of looting remains an open challenge.
Looting marks (such as small pits or bulldozer scars) are incredibly subtle and easily confused with irrelevant environmental changes like seasonal vegetation growth or shifting sands. Furthermore, progress in this niche field has been severely bottlenecked by a lack of data: prior studies relied on proprietary, non-multi-temporal, or completely private datasets, making it difficult for the broader machine learning community to contribute.
Datasets
To bridge this gap, the authors present the DAFA Looted Sites dataset (DAFA-LS). Curated in collaboration with expert archaeologists from the French archaeological delegation in Afghanistan (DAFA) and the startup Iconem, it contains:
- 55,480 RGB images captured monthly between 2016 and 2023.
- 675 unique Afghan archaeological sites, consisting of 135 looted and 540 preserved sites.
- 3 meters per pixel ground resolution sourced from Planet Basemaps.
- Coarse binary location masks manually annotated by experts to outline each site, helping models focus on the relevant area and avoid surrounding geographic biases (like nearby roads).
To prevent malevolent actors from using this dataset to locate and exploit intact sites, the authors intentionally added strong random noise to the spatial coordinates and applied geometric transformations to the distributed images.
Compared to existing datasets in the literature, DAFA-LS stands out as the only open-access dataset that provides both high temporal resolution (monthly) and a substantial number of sites.
Methods
The authors evaluated three distinct paradigms for satellite image time series classification:
- Single-frame methods: Standard architectures (like ResNets) and frozen foundation models (SatMAE, Scale-MAE, DOFA) trained only on 2023 images, operating under the assumption that looting marks remain visible in the final year's data.
- Pixel-wise multi-frame methods: Models (like TempCNN and DuPLo) that treat each pixel within the site mask as an independent time series, completely discarding spatial context.
- Whole-image multi-frame methods: Spatio-temporal architectures (like TSViT and UTAE) or foundation models paired with a Lightweight Temporal Attention Encoder (LTAE). This approach processes the entire image patch across all timestamps, capturing both spatial structures and temporal dynamics.
Key Insights and Ideas
- Spatial Context is Non-Negotiable: Whole-image multi-frame methods drastically outperformed pixel-wise methods. A looting pit cannot be identified by a single pixel's trajectory; the model must see the structural "scars" relative to the surrounding terrain.
- The Power of Foundation Models: Remote sensing foundation models—specifically DOFA (pretrained on over 8 million diverse satellite images)—provided a massive performance leap. Even though DOFA wasn't trained on this specific sensor resolution, its generalized representations transferred exceptionally well to this niche task.
- Temporal Attention Filters Out Seasonality: Incorporating the full time series via LTAE allowed the network to learn when to look. Ablation studies showed that the temporal attention mechanism naturally prioritized specific years (like 2020 and 2021, when major looting surges occurred) and downweighted confusing seasonal vegetation changes.
Key Results
The combination of the DOFA foundation model and the LTAE temporal encoder (DOFA+LTAE) delivered the best overall performance across all metrics:
- Overall Accuracy (OA): 78.7% (outperforming the best standard multi-frame baseline, UTAE, by 16.7 percentage points).
- F1-Score: 74.9%
- AUROC: 87.1%
Interestingly, while the single-frame DOFA model achieved a slightly lower false alarm rate (FAR) of 7.4%, incorporating the full time series (DOFA+LTAE) boosted the F1-score by +7.9 points, proving that temporal tracking is vital for minimizing missed looting events.
Limitations and Future Work
The primary limitation of DAFA-LS is its geographic homogeneity. Because all 675 sites are located in the arid, desertic regions of Afghanistan, the models trained on this dataset will likely struggle in regions with dense vegetation canopy (e.g., tropical jungles), where alternative modalities like SAR (Synthetic Aperture Radar) or LiDAR are required.
Additionally, the dataset only provides weak, series-level binary labels ("looted" vs. "preserved") rather than pixel-level annotations or exact timestamps of when the looting occurred. The authors suggest that exploring unsupervised change detection methods could be a promising next step to pinpoint exact looting dates without requiring tedious manual labeling.
Summary updated