Fig. 5: Detection of changes with a sequence of multi-spectral medium-resolution imagery.
From: Meta-learning to address diverse Earth observation problems across resolutions
This use-case shows the port of Beirut, Lebanon, where an explosion event caused substantial damage on August 4, 2020. a A total of 70 Sentinel-2 images where we use the first and last five images as support set to define the classes pre-event and post-event. task-model, which is fine-tuned on these examples. The meta-model in METEOR, shown in (b), is fine-tuned on these two classes and predicts all remaining images in the sequence. In (c), we show the resulting probability for the post-event class, where it remains low (0.6%) until August 3, 2020, and sharply rises probability of 84.5% on the following image of August 8, 2020. The crater and damaged buildings from the event caused the sudden increase in this probability score, as revealed by the occlusion sensitivity analysis drawn in (d). A further comparison to MOSAIKS and SSLTRANSRS is placed in Supplementary Fig. 5.
