Behzad Vahedi听received the Outstanding Student Presentation Award (OSPA) from the American Geophysical Union (AGU) for his presentation at the 2021 Fall Meeting. This award is for the presentation titled "A Comparison Of Classic Deep Learning Architectures For Sea Ice Classification From SAR". Behzad's advisor Morteza Karimzadeh, post doc听Dr. Benjamin Lucas, and collaborators in the National Snow and Ice Data Center (NSIDC) and CU Denver were Behzad's co-authors on this presentation.听According to AGU, "This honor is awarded for only the most exceptional presentations during AGU Fall Meeting 2021."
Presentation Abstract
During the last decade, advances in the state-of-the-art deep learning models, in particular convolutional neural networks, have facilitated significant improvements听in image recognition tasks. In fact, on the benchmark ImageNet dataset, the state of the art is now recognized as performing better than human. As a result, many adjacent tasks, including image recognition in remote sensing, have adopted these state-of-the-art models with little investigation into their transferability. For instance, the common image datasets鈥攆rom which pre-trained model weights are derived or modern architectures are evaluated on鈥攃ontain R-G-B images of everyday items such as animals, symbols, and vehicles. Needless to say, this is very different from the contents of a standard optical or radar image acquired by a satellite.
In this work we explore this idea of transferability in the context of the classification of sea ice type from Sentinel-1 SAR imagery. We compare the performance of a basic CNN with 4 significant models from the field of computer vision鈥擜lexNet, ResNet, VGG, and Inception鈥攊n the task of classifying sea ice in a region of the Chukchi Sea, a sea of the Arctic Ocean. We extend these experiments further to compare models that have been pre-trained on the ImageNet dataset with models where the parameters are randomly initialized, to demonstrate whether pre-trained models are beneficial for this application. The performance of models is compared using overall accuracy and F-1 score. Finally, we hypothesize why some models perform better on our dataset than the others, and we conclude by explaining how the results inform our model choice for future sea ice classification projects.