Unsupervised semantic segmentation of radar sounder data using contrastive learning

Radar Sounders (RSs) are active sensors widely used for planetary exploration and Earth observation that probe the subsurface in a non-intrusive way by acquiring vertical profiles, called radargrams. Radargrams contain information on subsurface geology and are analyzed with neural networks for segmentation and target detection. However, most of these methods rely on supervised training, which requires a large amount of labeled data that is hard to retrieve. Hence, a need emerges for a novel method for unsupervised radargram segmentation. This paper proposes a novel method for unsupervised radargram segmentation by analyzing semantically meaningful features extracted from a deep network trained with a contrastive logic. First, the network (encoder) is trained using a pretext task to extract meaningful features (query). Considering a dictionary of possible features (keys), the encoder training loss can be defined as a dictionary look-up problem. Each query is matched to a key in a large and consistent dictionary. Although such a dictionary is not available for RS data, it is dynamically computed by extracting meaningful features with another deep network called the momentum encoder. Secondly, deep feature vectors are extracted from the encoder for all radargram pixels. After the feature selection, the feature vectors are binarized. Since pixels of the same class are expected to have similar feature vectors, we compute the similarity between the feature vectors to generate a cluster of pixels for each class. We applied the proposed method to segment radargrams acquired in Greenland by the MCoRDS-3 sensor, achieving good overall accuracy.