| El Ni(?)o-Southern Oscillation(ENSO)is a basin-scale sea-air interaction phenomenon manifested as changes in sea surface temperature and pressure over the equatorial Pacific Ocean.It is one of the major drivers of interdecadal climate variability on Earth,with significant impacts on global climate,agriculture,ecosystems,health,and society,and can even induce war.Therefore,earlier,and more accurate prediction of ENSO to minimize these impacts is valuable.However,predicting the intensity of ENSO remains a challenge due to its complexity.The increasing diversity of ENSO since 2000 has led to an increasing interest in the types of ENSO events.Since the climate impacts of different ENSO types vary greatly,e.g.,different types of ENSO events have different impacts on winter temperatures in the U.S.and climate in East Asia,accurately predicting ENSO event types is important to improve the quality of climate forecasts.ENSO has complex phenomena,incomplete theories,unstable initial conditions and unavoidable data bias,so existing dynamical models and traditional statistical models cannot accurately predict ENSO.In recent years,deep learning methods are active in various fields with their powerful feature extraction capabilities,among which several studies have been conducted in the field of weather forecasting.However,there are still limitations in previous explorations in this area,mainly including the following: first,the amount of data in previous studies done is small.Secondly,the performance of the deep learning methods applied in this field today is average.Third,there are various types of ENSO in recent years,and there is the problem of sample imbalance.Fourth,deep learning algorithms lack interpretability and cannot correlate prediction results with physical mechanisms.In this paper,we address the above problems by extending the dataset and integrating various models and technical tools to predict the intensity and type of ENSO events and make reasonable explanations of the model prediction results based on visualization techniques.The research work in this paper has:The problem of insufficient historical observation data is addressed by using extended data and transfer learning.Expanded data is to increase the amount of data by fully applying simulated data generated by multiple members of multiple models.Transfer learning is to make the forecast results closer to the real values by acting the knowledge learned from simulated data to the real data in some way.Two novel models are proposed to address the existing problem of inadequate forecasting performance of traditional algorithms.One is a model consisting of convolutional neural network and Transformer,which has both inductive bias characteristics and the ability to focus on global features,thus achieving better forecasting results for both ordinary and super El Ni(?)o events;The other is a deep residual convolutional neural network combining residual structure and transfer learning to further improve the prediction capability.For the sample imbalance problem,introduce a weighting mechanism and heterogeneous migration learning methods for its comprehensive prediction of type.The former makes the model focus more on the hard-to-classify samples during training by reducing the weights of easy-to-classify samples.The latter enables the model to obtain stronger predictive power by acting the predictive power of the index on the prediction types.To address the lack of interpretability,the regions of interest to the model are visualized by the gradient-weighted class activation mapping method and compared with real physical mechanisms,and the results show that proposed model can focus on important regions and can explore the underlying physical mechanisms.This study not only can well predict the intensity and type of ENSO,but also can analyze the physical mechanism of ENSO and its potential mechanism by integrating scientific hypothesis and visualization technology technique,which provides new avenues for the application of deep learning to the study of earth science. |
