Partial Label Learning for Sea Ice Type Classification in the Arctic
Topics:
Keywords: Deep learning, remote sensing, convolutional neural networks, partial label learning
Abstract Type: Paper Abstract
Authors:
Behzad Vahedi, Department of Geography, University of Colorado Boulder, Boulder, USA
Andrew P Barrett, National Snow and Ice Data Center (NSIDC), CIRES, University of Colorado Boulder, Boulder, USA
Walter N Meier, National Snow and Ice Data Center (NSIDC), CIRES, University of Colorado Boulder, Boulder, USA
Siri Jodha Khalsa, National Snow and Ice Data Center (NSIDC), CIRES, University of Colorado Boulder, Boulder, USA
Farnoush Banaei-Kashani, Department of Computer Science and Engineering, University of Colorado Denver, Denver, USA
Morteza Karimzadeh, Department of Geography, University of Colorado Boulder, Boulder, USA
,
,
,
,
Abstract
Automated sea ice classification based on deep learning has great potential for automating operational ice mapping. To train supervised sea ice type classification algorithms, expert-generated ice charts are used as de-facto labels. These charts are comprised of polygons, each containing up to three different ice types and a partial concentration level associated with each type. This approach assigns multiple candidate labels (sea ice types) to each polygon, posing a challenge for conventional deep learning-based sea ice type classification algorithms. Such algorithms are trained using single-label training samples in which information about ice concentration and secondary ice types is ignored. Additionally, training datasets in sea ice classification are often imbalanced between classes. This skews the performance of deep learning algorithms towards producing higher accuracy on the majority classes, even if the minority classes are of more importance.
To address these challenges, we present a novel approach to sea ice classification by formalizing this problem as a partial label learning task with explicit confidence scores, while leveraging focal loss to address the class imbalance problem. Based on this approach, we treat all the polygon-level labels as candidate partial labels at the patch level, assign explicit confidence scores obtained from the corresponding sea ice concentrations to each candidate label, and then integrate them within a focal loss function to train a Convolutional Neural Network. We then tune the hyperparameters of focal loss for sea ice classification, providing guidance on what values to use for future sea ice classification applications.
Partial Label Learning for Sea Ice Type Classification in the Arctic
Category
Paper Abstract
