Description
The problem of coastline erosion is of global concern. Acquisition and processing of useful earth
observation data about coastal changes is crucial to accurate change monitoring [1]. With the
availability ofsophisticated machine learning techniques, it is possible to accurately detect and analyze
patterns of changes in coastal regions. One important aspect here is the explainability of the machine
learning model used to predict changes and the possibility to incorporate human expertise in the
process of detection [2]. In this research, we use an explainable artificial intelligence model to discover
data patterns in Sentinel-2 time-series images to describe changes over a 7-year study period. Timeseries imagery was acquired every month from January 2018 to September 2023, covering 4,694 cloudfree locations along the North Sea and Baltic Sea coastlines, each spanning 5 km x 5 km. These locations
were selected using farthest point sampling to ensure representative coverage. The imagery was
further divided into smaller scenes of 1.28 km x 1.28 km, and active learning techniques were
employed to minimize labeling efforts. We have used Latent Dirichlet Allocation (LDA), a Bayesian
generative model recently established as explainable model [1]. Being a probabilistic model, LDA is
able to output certainty score for its predictions. We use the LDA as an unsupervised explainable model
to create interpretable intermediate visual outcomes that support model explainability, while certainty
scores of each prediction enhances trust. These interpretable outcomes are used by the domain expert
to assess quality of the outcomes. Two kinds of visualizations are produced: 1) visual topic maps -LDA
retrieved visual topics depicting latent data patterns, often perceived by humans as visual objects 2)
change class maps and change signature maps - maps showing which land cover classes (e.g wavebreaking zones, dry sand, inter-tidal area, vegetation) have gone through most changes ( we produce
histograms showing percentage of change per class per year, and also over the whole study period );
change signatures describe the nature of change in every class. We conclude the research by validating
our results by domain experts.
References:
1. A. Fejjari, G. Valentino, J. A. Briffa and S. D'Amico, "Detection and Monitoring of Maltese
Shoreline Changes using Sentinel-2 Imagery," 2023 IEEE International Workshop on
Metrology for the Sea; Learning to Measure Sea Health Parameters (MetroSea), La Valletta,
Malta, 2023, pp. 52-56, doi: 10.1109/MetroSea58055.2023.10317486.
2. C. Karmakar, C. O. Dumitru, G. Schwarz and M. Datcu, "Feature-Free Explainable Data Mining
in SAR Images Using Latent Dirichlet Allocation," in IEEE Journal of Selected Topics in Applied
Earth Observations and Remote Sensing, vol. 14, pp. 676-689, 2021, doi:
10.1109/JSTARS.2020.3039012.
3. C. Karmakar, C.O. Dumitru, N. Hughes and M. Datcu, "A Visualization Framework for
Unsupervised Analysis of Latent Structures in SAR Image Time Series", IEEE Journal of
Selected Topics in Applied Earth Observations and Remote Sensing, 16, pp. 5355-5373, 2023.
4. L. Du, Y. Wang, W. Xie, Z. Wang and J. Chen, "A Semisupervised Infinite Latent Dirichlet
Allocation Model for Target Discrimination in SAR Images With Complex Scenes," in IEEE
Transactions on Geoscience and Remote Sensing, vol. 58, no. 1, pp. 666-679, Jan. 2020, doi:
10.1109/TGRS.2019.2939001.
5. Sentinel-2 mission. Available online: https://sentinel.esa.int/web/sentinel/missions/sentinel
