Giving meaning to unsupervised EO change detection rasters: a semantic-driven approach

The field of Earth Observation (EO) change detection has been fostered with new sources of satellite image data coupled with the development of deep learning algorithms. However, the output of these algorithms lacks context. Contextual knowledge explaining a detected change is required to better analyze those images and understand the phenomenon that caused the change. This paper presents a semantic-driven data integration approach that supports the generation of a knowledge graph from a raster change file and from various data sources of events that may explain the changes. The output graph represents spatial and temporal features for areas affected by a high change, as well as various kinds of contextual data useful for explaining the detected changes. We validate the approach with a real-case scenario of fire monitoring. We process changes detected between pairs of Sentinel-2 images located on the same tiles, with contextual data such as administrative units, tweets, and thermal sensors. We show the added value of the proposed approach for i) explaining change detection and ii) validating the results from unsupervised deep learning algorithms.