Understanding the Changes in Polar Areas with Time Series Satellite Imagery

Kurzfassung

Studying the changes on the Earth’s surface is always an important research question in remote sensing. Many methods and algorithms have been proposed by researchers to understand the dynamics on regions of Earth using Satellite Image Time Series (SITS), as SITS contain huge amount of information. However, some regions are more challenging because of lack of ground truth data. This contribution presents an explainable machine learning method to produce transparent, trustworthy and interpretable results to understand the dynamics in the Polar region using SAR Sentinel-1 SITS. We propose an explainable artificial intelligence approach to discover the hidden information about the surface cover dynamics over the 2018 and 2019. Visualizations are produced to support explainability of the results. The specific objectives are, to create semantic labels for 24 study months (one image for each month of the selected period) by forwarding some limited amount of domain knowledge, and then to analyze the time series evolution based on these semantic labels. The domain knowledge used in classification consists of a labeled dataset from an Active Learning research activity [1], and knowledge about the mutual closeness of the surface cover classes from another research project [2]. We semantically labeled the images acquired over 24 months of study and each image is tiled into 6,400 patches of size 256×256 pixels, giving a total of 153,600 patches. We retrieve 8 semantic classes as: Black border, Glaciers, Icebergs, Mountains, Old ice, First year ice, Young ice, and Water group. The class Water group combines 4 close classes that were defined in [1]: Water body, Melted snow, Water current, and Floating ice. The classification achieved an average accuracy of 95%. The land cover change analysis was performed using a Bayesian model called Latent Dirichlet Allocation (LDA) which has been demonstrated to be an explainable artificial intelligence model by concerned literature [3]. LDA was originally developed for text classification and enables us to retrieve change classes by observing the semantic labels. The change classes are further described with color-coded change signatures. The pattern and homogeneity of the change signatures indicate the nature and amount of the changes in the respective change class. References: [1] C.O. Dumitru, V. Andrei, G. Schwarz, and M. Datcu, “Machine Learning for Sea Ice Monitoring from Satellites”, International Archives of the Photogrammetry, Remote Sensing & Spatial Information Sciences, 2019, Vol. XLII-2/W16, pp. 83-89. [2] H2020 ExtremeEarth project, available online: http://earthanalytics.eu/. [3] C. Karmakar, C.O. Dumitru, G. Schwarz, and M. Datcu, “Feature-Free Explainable Data Mining in SAR Images Using Latent Dirichlet Allocation”, IEEE JSTARS, 2021, Vol. 14, pp. 676-689.