Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Detecting Looted Archaeological Sites from Satellite Image Time Series

Elliot Vincent, Mehraïl Saroufim, Jonathan Chemla, Yves Ubelmann, Philippe Marquis, Jean Ponce, Mathieu Aubry

Abstract

Archaeological sites are the physical remains of past human activity and one of the main sources of information about past societies and cultures. However, they are also the target of malevolent human actions, especially in countries having experienced inner turmoil and conflicts. Because monitoring these sites from space is a key step towards their preservation, we introduce the DAFA Looted Sites dataset, \datasetname, a labeled multi-temporal remote sensing dataset containing 55,480 images acquired monthly over 8 years across 675 Afghan archaeological sites, including 135 sites looted during the acquisition period. \datasetname~is particularly challenging because of the limited number of training samples, the class imbalance, the weak binary annotations only available at the level of the time series, and the subtlety of relevant changes coupled with important irrelevant ones over a long time period. It is also an interesting playground to assess the performance of satellite image time series (SITS) classification methods on a real and important use case. We evaluate a large set of baselines, outline the substantial benefits of using foundation models and show the additional boost that can be provided by using complete time series instead of using a single image.

Detecting Looted Archaeological Sites from Satellite Image Time Series

Abstract

Archaeological sites are the physical remains of past human activity and one of the main sources of information about past societies and cultures. However, they are also the target of malevolent human actions, especially in countries having experienced inner turmoil and conflicts. Because monitoring these sites from space is a key step towards their preservation, we introduce the DAFA Looted Sites dataset, \datasetname, a labeled multi-temporal remote sensing dataset containing 55,480 images acquired monthly over 8 years across 675 Afghan archaeological sites, including 135 sites looted during the acquisition period. \datasetname~is particularly challenging because of the limited number of training samples, the class imbalance, the weak binary annotations only available at the level of the time series, and the subtlety of relevant changes coupled with important irrelevant ones over a long time period. It is also an interesting playground to assess the performance of satellite image time series (SITS) classification methods on a real and important use case. We evaluate a large set of baselines, outline the substantial benefits of using foundation models and show the additional boost that can be provided by using complete time series instead of using a single image.
Paper Structure (38 sections, 7 figures, 4 tables)

This paper contains 38 sections, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Archaeological looting detection
  4. SITS-based visual detection of archaeological looting.
  5. Automatic looting detection.
  6. Satellite image time series classification
  7. Satellite image time series datasets.
  8. Satellite image time series classification approaches.
  9. Dataset
  10. Sites identification and characterization.
  11. Test set definition.
  12. Training and validation splits.
  13. Benchmark
  14. Task and evaluation
  15. Problem formalization.
  16. ...and 23 more sections

Figures (7)

  • Figure 1: DAFA Looted Sites (DAFA-LS) contains monthly satellite image time series (SITS) of Afghan archaeological sites acquired between 2016 and 2023. We show the location of preserved (a) and looted (b) sites, adding strong random noise to their coordinates to prevent misuse of the data. Test sites are marked with a star ($\star$). We also show images (d-f) from two sites (c): Ancient Balkh (in blue, top row) has been preserved from looting unesco2004balkh, while Dilberjin (in red, bottom row) suffered irreparable damage.
  • Figure 2: Examples of time series and coarse location masks. For each archaeological site, we show the September image for each year from 2016 to 2023 and the corresponding coarse location mask. The 4 top rows show looted sites (red squares) and the 4 bottom rows show preserved sites (blue squares).
  • Figure 3: Ablation of DOFA+LTAE. We evaluate DOFA+LTAE on either month-specific (a) or year-specific (b) sub-time series of DAFA-LS. In other words, instead of taking all monthly time stamps as input, we only take images of a given month (across all years) or of a given year.
  • Figure A1: Example of visible looting marks.
  • Figure A2: Example of failure cases. We show all the time series for which our best baseline (DOFA+LTAE) predicts the wrong label with a confidence higher than 95%.
  • ...and 2 more figures