The Tibetan Plateau (TP) is known as one of the most sensitive regions to climate change, and it has experienced accelerated warming in recent decades. However, to what degree the TP warming amplification relates to remote forcing such as sea ice loss in the Arctic sea ice remains unclear. Here, we found that the decline of sea ice concentration over the Barents-Kara Sea (BKS) could account for 18–32% of the winter warming over the TP by comparing observational data and ensemble experiments from an atmospheric general circulation model. The reduced BKS sea ice and resultant upward turbulent heat fluxes can intensify a Rossby wave train propagating equatorward to the TP. As a result, the enhanced southwesterlies towards the TP strengthen the warm advection over most parts of the TP and lead to TP warming. In addition, an atmospheric teleconnection between the Arctic and the TP also exists in the interannual variability. That is, a tripole mode in air temperature, with warm centers in the Arctic and TP but a cold center in the mid-high latitudes of the Eurasian continent in between. Our results imply that the BKS sea ice loss could intensify such a tripole mode and thus enhancing the winter TP warming.

Accurate prediction of the East Asian summer monsoon (EASM) is beneficial to billions of people’s production and lives. Here, a convolutional neural network (CNN) and transfer learning are used to predict the EASM. The results of the constructed CNN regression model show that the prediction of the CNN regression model is highly consistent with the reanalysis dataset, with a correlation coefficient of 0.78, which is higher than that of each of the current state-of-the-art dynamic models. The heat map method indicates that the robust precursor signals in the CNN regression model agree well with previous theoretical studies and can provide the quantitative contribution of different signals for EASM prediction. The CNN regression model can predict the EASM one year ahead with a confidence level above 95%. The above method can not only improve the prediction of the EASM but also help to identify the involved physical predictors.