Informatik-Kolloquium

Dr. Marco Körner, from the Technical University of Munich is giving a talk on ” Why you should NOT only look once: Multi-temporal models for analyzing optical remote sensing data”

Abstract:

Due to their impressive performance, the use of feed-forward deep convolutional neural network models became state-of-the-art for semantic segmentation tasks considering persistent objects in natural and man-made scenarios. While this persistence assumption is valid in many cases—e.g., street scenes, façades, documents, etc.—, the use of more complex models is called for when the objects to be detected show inherent temporal variability.

The domain of remote sensing and earth observation, in particular, has to face this issue, as the varying overall appearance of detection targets causes context-free modeling schemes to be very likely to fail.

I will share our considerations on the potential of context-sensitive models (i.e., recurrent neural networks and long short-term memory networks) for the special use case of vegetation modeling from remote sensing satellite data. By doing so, we utilize these phenological patterns as valuable features in crop detection application scenarios.