Accurate prediction of independent sea ice motion in arctic shipping lanes is of great guiding significance for ensuring navigation safety， assessing navigation navigability and dynamically correcting shipping lanes. However， the traditional optical flow method cannot meet the requirement of "spatio-temporal prediction + semantic segmentation". In this paper， the sea ice motion data set SeaiceMoving was made based on MERSI-Ⅱ image and a sea ice motion prediction algorithm based on Multiloss-SAM-ConvLSTM was proposed， introducing weighted FDWloss based on SAM-ConvLSTM to enhance the acquisition of spatial semantics of each node. Aiming at the imbalanced sample distribution， we discussed the offset effect of back-end segmentation threshold. The optimal segmentation threshold is determined by grid search method， which improves the overall prediction result of sea ice motion. The experimental results indicate that the Kappa coefficient， IOU coefficient and Dice coefficient of the proposed method are 0.75， 0.61 and 0.76 respectively. Compared with SAM-ConvLSTM， they are improved by 0.1， 0.12 and 0.1 respectively. Furthermore， the proposed method can improve the position prediction and shape recovery ability of sea ice after motion and reduce the "adhesion" of sea ice. In addition， the algorithm can still effectively predict the sea ice motion under the interference of thin clouds， which can provide more accurate technical support for the dynamic planning and route correction of the Arctic route.