北极海冰范围退缩已对区域和全球气候变化、北极通航性乃至地缘政治格局产生了深远影响，开展北极海冰预测具有重要意义。本文使用美国冰雪数据中心（NSIDC）发布的海冰密集度、海冰冰龄遥感资料以及ERA5再分析资料，结合前人研究和海气耦合模式结果选择预测参量方案，开展了月尺度的海冰冰情预测。比较了支持向量机（SVR）、深度森林（DF）、LightGBM （LGB）、XGBoost （XGB）和CatBoost （CAT）等5种机器学习算法和以树模型LGB、XGB和CAT作为基模型，以贝叶斯回归、岭回归、套索回归和深度森林作为元模型的4种堆叠式集成学习模型，以及深度神经网络（DNN）、卷积神经网络（CNN）、时空卷积网络（ConvLSTM）3种深度学习模型在2000年测试集上对海冰范围和密集度空间分布的预测效果。结果表明：在海冰密集度预测中，ConvLSTM表现最优，套索堆叠集成学习模型预测效果次之。集成学习模型相较于三种单一树模型在预测效果上有约1%~4%的提升。在海冰范围预测中，堆叠式集成学习模型的预测效果最好。本研究为开展机器学习海冰预测奠定了重要基础。

The rapid retreat of Arctic sea ice has attracted extensive international attention， with far-reaching impacts on navigation periods， regional and global climate change， and geopolitical Pattern. It is of great significance to predict Arctic sea ice. In this study， the sea ice concentration， ice age data from NSIDC and ERA5 reanalysis were used to select prediction parameter schemes based on previous studies and coupled Ocean-Atmosphere model results. This study focused on the monthly-scale sea ice prediction. Five machine learning algorithms， including support vector machine （SVR）， Deep forest （DF）， LightGBM （LGB）， XGBoost （XGB） and CatBoost （CAT）， and four stacked ensemble learning models using Bayesian regression， ridge regression， Lasso regression and deep forest as meta models， three tree models LGB， XGB and CAT as base models，as well as deep neural network （DNN）， convolutional neural network （CNN） and spatio-temporal Convolution network （ConvLSTM） were compared and evaluated for the predictions of sea ice concentration and extent in 2000 test set. The results show that ConvLSTM has the best performance in the prediction of sea ice concentration， and lasso stacking ensemble learning model takes the second place. The ensemble learning model improved the prediction performance by about 1% to 4% compared with the three single tree models. In sea ice extent prediction， stacking ensemble learning model performs best. This study has laid an important foundation for the monthly-scale sea ice prediction based on the machine learning.