Abstract:  Sea ice plays a critical role in influencing various aspects of  atmosphere, ecology, oceans, and even human activities, making the task of sea ice retrieval (SIR) immensely important. At present,  machine learning methods based on global navigation satellite system  reflectometry (GNSS-R) technology have benn applied  in the field of sea ice remote sensing, achieving good results. However, the existing algorithms are hindered by over-reliance on labels, structural redundancies, and limited effectiveness in dealing with sample imbalances. To overcome these challenges, the anomaly detection driven semi-supervised multi-task retrieval algorithm for sea ice based on GNSS-R is proposed in this paper. The approach of semi-supervised anomaly detection is applied to the field of remote sensing SIR, which solves the retrieval problem in the case of over-reliance on labels and sample imbalance. The proposed algorithm is composed of two primary components: the data preprocessing module and the GNSS-R-based sea ice multi-task retrieval network (SIMTRN). SIMTRN contains a sea ice detection (SID) module and a sea ice concentration retrieval (SICR) module, and utilizes the inter-module data feature sharing mechanism to realize  SID and SICR tasks based on a single network, which solves the problem of structural redundancy. In addition, for the SICR task, this paper considers the relationship between different atmospheric and meteorological features and sea ice concentration, and improves the prediction accuracy by fusing with sea ice delay Doppler maps’ features. Experimental results demonstrate that SIMTRN achieves a SID accuracy of 0.973 3 and a SICR correlation coefficient of 0.969 9, outperforming four existing machine learning algorithms: The backpropagation-learning multilayer perceptron neural network, the convolutional neural network based on single-layer convolution, the support vector based on feature selection, and the convolutional neural network based on double-layer convolution. Additionally, this study confirms that the proposed algorithm possesses superior generalization capability to the existing methods.

Key words: sea ice, anomaly detection, GNSS-R, remote sensing, semi-supervised learning