The El Niño–Southern Oscillation (ENSO) is an extremely complicated ocean–atmosphere coupling event, the development and decay of which are usually modulated by the energy interactions between multiple physical variables. In this paper, we design a multivariate air–sea coupler (ASC) based on the graph using features of multiple physical variables. On the basis of this coupler, an ENSO deep learning forecast model (named ENSO-ASC) is proposed, whose structure is adapted to the characteristics of the ENSO dynamics, including the encoder and decoder for capturing and restoring the multi-scale spatial–temporal correlations, and two attention weights for grasping the different air–sea coupling strengths on different start calendar months and varied effects of physical variables in ENSO amplitudes. In addition, two datasets modulated to the same resolutions are used to train the model. We firstly tune the model performance to optimal and compare it with the other state-of-the-art ENSO deep learning forecast models. Then, we evaluate the ENSO forecast skill from the contributions of different predictors, the effective lead time with different start calendar months, and the forecast spatial uncertainties, to further analyze the underlying ENSO mechanisms. Finally, we make ENSO predictions over the validation period from 2014 to 2020. Experiment results demonstrate that ENSO-ASC outperforms the other models. Sea surface temperature (SST) and zonal wind are two crucial predictors. The correlation skill of the Niño 3.4 index is over 0.78, 0.65, and 0.5 within the lead time of 6, 12, and 18 months respectively. From two heat map analyses, we also discover the common challenges in ENSO predictability, such as the forecasting skills declining faster when making forecasts through June–July–August and the forecast errors being more likely to show up in the western and central tropical Pacific Ocean in longer-term forecasts. ENSO-ASC can simulate ENSO with different strengths, and the forecasted SST and wind patterns reflect an obvious Bjerknes positive feedback mechanism. These results indicate the effectiveness and superiority of our model with the multivariate air–sea coupler in predicting ENSO and analyzing the underlying dynamic mechanisms in a sophisticated way.

A typhoon is an extreme weather event with strong destructive force, which can bring huge losses of life and economic damage to people. Thus, it is meaningful to reduce the prediction errors of typhoon intensity forecasting. Artificial and deep neural networks have recently become widely used for typhoon forecasting in order to ensure typhoon intensity forecasting is accurate and timely. Typhoon intensity forecasting models based on long short-term memory (LSTM) are proposed herein, which forecast typhoon intensity as a time series problem based on historical typhoon data. First, the typhoon intensity forecasting models are trained and tested with processed typhoon data from 2000 to 2014 to find the optimal prediction factors. Then, the models are validated using the optimal prediction factors compared to a feed-forward neural network (FNN). As per the results of the model applied for typhoons Chan-hom and Soudelor in 2015, the model based on LSTM using the optimal prediction factors shows the best performance and lowest prediction errors. Thus, the model based on LSTM is practical and meaningful for predicting typhoon intensity within 120 h.

The North Atlantic Oscillation (NAO) is the most significant mode of the atmosphere in the North Atlantic, and it plays an important role in regulating the local weather and climate and even those of the entire Northern Hemisphere. Therefore, it is vital to predict NAO events. Since the NAO event can be quantified by the NAO index, an effective neural network model EEMD-ConvLSTM, which is based on Convolutional Long Short-Term Memory (ConvLSTM) with Ensemble Empirical Mode Decomposition (EEMD), is proposed for NAO index prediction in this paper. EEMD is applied to preprocess NAO index data, which are issued by the National Oceanic and Atmospheric Administration (NOAA), and NAO index data are decomposed into several Intrinsic Mode Functions (IMFs). After being filtered by the energy threshold, the remaining IMFs are used to reconstruct new NAO index data as the input of ConvLSTM. In order to evaluate the performance of EEMD-ConvLSTM, six methods were selected as the benchmark, which included traditional models, machine learning algorithms, and other deep neural networks. Furthermore, we forecast the NAO index with EEMD-ConvLSTM and the Rolling Forecast (RF) and compared the results with those of Global Forecast System (GFS) and the averaging of 11 Medium Range Forecast (MRF) model ensemble members (ENSM) provided by the NOAA Climate Prediction Center. The experimental results show that EEMD-ConvLSTM not only has the highest reliability from evaluation metrics, but also can better capture the variation trend of the NAO index data.

In this letter, a Physics-Informed Red Tide (PIRT) forecast model considering causal-inferred predictors selection is proposed. Specifically, the directed acyclic graph-graph neural network (DAG-GNN) method is first applied to quantify the causality among multiple ocean-atmosphere-biology variables for selecting the most significant predictors of the red tides (or other chlorophyll variations). Then, the encoder-decoder model consisting of an Energy Attention Module (EAM) is built for daily red tide forecasting. The multisourced multivariate dataset during 2010–2020 covering the East China Sea serves to train and evaluate PIRT. The experimental results demonstrate that the predictors in the learned causal graph are closely related to the occurrence and decay of red tides, which exhibits high physical interpretability. PIRT has a superior forecasting skill, the predictions of which are with highly consistent spatial patterns, especially in extreme events. The seven-lead-day forecast errors for chlorophyll are within 0.9 $\text{mg}{\cdot}\text{m}^{-3}$ , which is much better than the other models. This also indicates that PIRT can be used as a reliable tool to study the ecology of the East China Sea.

Climate downscaling is a way to provide finer resolution data at local scales, which has been widely used in meteorological research. The two main approaches for climate downscaling are dynamical and statistical. The traditional dynamical downscaling methods are quite time- and resource-consuming based on general circulation models (GCMs). Recently, more and more researchers construct a statistical deep learning model for climate downscaling motivated by the single-image superresolution (SISR) process in computer vision (CV). This is an approach that uses historical climate observations to learn a low-resolution to high-resolution mapping and produces great enhancements in terms of efficiency and effectiveness. Therefore, it has provided an appreciable new insight and successful downscaling solution to multiple climate phenomena. However, most existing models only make a simple analogy between climate downscaling and SISR and ignore the underlying dynamical mechanisms, which leads to the overaveraged downscaling results lacking crucial physical details. In this paper, we incorporate the a priori meteorological knowledge into a deep learning formalization for climate downscaling. More specifically, we consider the multiscale spatial correlations and the chaos in multiple climate events. Depending on two characteristics, we build up a two-stage deep learning model containing a stepwise reconstruction process and ensemble inference, which is named climate downscaling network (CDN). It can extract more local/remote spatial dependencies and provide more comprehensive captures of extreme conditions. We evaluate our model based on two datasets: climate science dataset (CSD) and benchmark image dataset (BID). The results demonstrate that our model shows the effectiveness and superiority in downscaling daily precipitation data from 2.5 degrees to 0.5 degrees over Asia and Europe. In addition, our model exhibits better performance than the other traditional approaches and state-of-the-art deep learning models.

The global impact of an El Niño–Southern Oscillation (ENSO) event can differ greatly depending on whether it is an eastern Pacific (EP)-type event or a central Pacific (CP)-type event. Reliable predictions of the two types of ENSO are therefore of critical importance. Here we construct a deep neural network with multichannel structure for ENSO (named ENSO-MC) to simulate the spatial evolution of sea surface temperature (SST) anomalies for the two types of events. We select SST, heat content and wind stress (i.e., three key ingredients of Bjerknes feedback) to represent coupled ocean–atmosphere dynamics that underpin ENSO, achieving skilful forecasts for the spatial patterns of SST anomalies out to 1 year ahead. Furthermore, it is of great significance to analyse the precursors of EP-type or CP-type events and identify targeted observation sensitive areas for the understanding and prediction of ENSO. Precursors analysis is to determine what type of initial perturbations will develop into EP-type or CP-type events. Sensitive area identification is to determine the regions where initial states tend to have the greatest impacts on the evolution of ENSO. We use the saliency map method to investigate the subsurface precursors and identify the sensitive areas of ENSO. The results show that there are pronounced signals in the equatorial subsurface before EP events, while the precursory signals of CP events are located in the northern Pacific. It indicates that the subtropical precursors seem to favour the generation of the CP-type El Niño and that the EP-type El Niño is more related to the tropical thermocline dynamics. Furthermore, the saliency maps show that the sensitive areas of the surface and the subsurface are located in the equatorial central Pacific and the equatorial western Pacific respectively. The sensitivity experiments imply that additional observations in the identified sensitive areas can improve forecasting skills. Our results of precursors and sensitive areas are consistent with the previous theories of ENSO, demonstrating the potential usage and advantages of the ENSO-MC model in improving the simulation, understanding and observations of the two ENSO types.