Learning to Predict Land Cover from Bad Examples
Keywords: Machine learning, land cover classification, weak supervision, label noise
Abstract: One of the core tasks of remote sensing is to create (and update at ever higher frequencies) maps of the Earth’s land cover, which is traditionally cast as a (supervised) machine learning problem. Given the difficulty to create actual ground truth for the annotation of training samples on the one hand and the availability of several global land cover datasets with rather low resolution and limited accuracies on the other hand, it is high time to pair recent progress in machine learning from noisy samples with the latest developments in remote sensing-based land cover classification.
This thematic session will address the problem of learning well-generalizing models for land cover prediction from bad examples, where “bad” relates to both a resolution that is lower than the measurement satellite data (i.e. to the concept of weak supervision), and to erroneous annotations (i.e. to the concept of label noise). In particular, it invites results that have been achieved based on the SEN12MS dataset, which was published on the ISPRS Workshop “Munich Remote Sensing Symposium” in September 2019.
Chairs: Michael Schmitt, Signal Processing in Earth Observation, Technical University of Munich
Jan Dirk Wegner, EcoVision Lab, ETH Zurich
Supported by: Ribana Roscher, Institute of Geodesy and Geoinformation, University of Bonn
Ronny Hänsch, Microwaves and Radar Institute, German Aerospace Center (DLR)
Naoto Yokoya, Center for Advanced Intelligence Project, RIKEN
Pedram Ghamisi, Helmholtz-Zentrum Dresden-Rossendorf, Helmholtz Institute
Freiberg for Resource Technology
The session should be open to submissions by anyone, especially those who are working with the SEN12MS dataset or participate in the IEEE-GRSS Data Fusion Contest. In addition, we will invite the following two colleagues to submit a paper based on their experience and previous work relevant to the topic:
• Giles M. Foody, School of Geography, University of Nottingham, UK – Land cover accuracy assessment
• Caleb Robinson, Georgia Institute of Technology – Label super-resolution networks