Weakly-Supervised Estimation of Auroral Motion Field Via Iterative Pseudo Ground Truth Learning.

The small scale auroral structure is a far less explored field. Provided by auroral images which record the vivid auroral behaviors with satisfied temporal and spatial resolution, we are devoted to studying local auroral motions and the fine scale auroral activities. In order to estimate the auroral motion field, the method of optical flow is introduced to analyze auroral motion. However, the technology requires expensive dense annotations while training the network. Leveraged between the strong learning ability of the fully-supervised deep learning methods and the uncertainty of auroral data, we propose an iterative ground-truth learning approach to mine the pixel-level pseudo ground truth for auroral motion. Specifically, we first train a fully-supervised estimator on synthetic data via the Recurrent All-Pairs Field Transforms (RAFT) algorithm. The reconstructability and robustness of the estimated motion field are used as the criteria to measure applicability of the fully-supervised estimator for auroral images. Then, the mined motion fields as pseudo ground truths are in turn fed into the RAFT algorithms to fine-tune the fully-supervised estimator again, which is iterated until the high-quality pseudo ground truths for auroral data are found. Experiments on auroral data from the Yellow River Station demonstrate the effectiveness of our method. More and more pseudo ground truths of auroral data are used to gradually improve the estimated motion field results by refining the contextual features of auroral images. With iterative pseudo ground truth learning, estimated errors can be reduced effectively.