Deep learning methods help with sea ice classification based on Sentinel-1 synthetic aperture radar data. Image credit: NOAA (left) and Oscar Nord (right) via Unsplash.

PhD candidate Salman Khaleghian has been exploring new deep neural networks for sea ice classification using Sentinel-1 synthetic aperture radar (SAR). In his recently published paper, he and his colleagues have addressed a challenging issue in remote sensing based on deep semi-supervised learning. For sea ice classification, labeling the data precisely is very time consuming and requires expert knowledge.

Scarce training data is one of the most challenging issues in remote sensing data analysis due to the high cost connected to obtaining in-situ measurements and the time-consuming nature of the task. In this sense, semi-supervised learning is becoming increasingly important because it can combine the data that humans have carefully labeled with a large amount of available unlabeled data in order to improve the predictive performance of machine learning remote sensing approaches. This work proposes a novel deep semi-supervised learning (SSL) method for sea ice classification based on Sentinel-1 synthetic aperture radar (SAR) data.  

His method efficiently learns sea ice characteristics from a very limited number of labeled samples and a relatively large number of unlabeled samples. This method consists of two models, namely a Teacher and a Student model. The Teacher model is trained and generates pseudo-labels for unlabeled data considering a transductive label propagation method based on the manifold assumption in the feature space of the CNN, which are used to train the student model. They showed the benefits of this semi-supervised approach for sea/ice classification and its extensions to other remote sensing applications.

Access the full paper here.