Assessing seasonal climate predictability using a deep learning application: NN4CAST

Abstract

Seasonal climate predictions are essential for climate services, with changes in tropical sea surface temperature (SST) representing the most influential oceanic drivers. SST anomalies can affect the climate in remote regions through various atmospheric teleconnection mechanisms, and the persistence/evolution of those SST anomalies can give seasonal predictability to atmospheric signals. Dynamical models often struggle with biases and low signal-To-noise ratios, making statistical methods a valuable alternative. Deep learning models are currently providing accurate predictions, mainly in short-range weather forecasts. Nevertheless, the black-box nature of this methodology makes it necessary to ensure its explainability. In this context, we present NN4CAST (Neural Network foreCAST), a Python deep learning pipeline designed to assess seasonal predictability, with built-in tools for evaluating model skill and performing basic spatial diagnostics based on empirical orthogonal functions. Starting from the raw datasets, NN4CAST performs all methodological steps: preprocessing, training and evaluation, enabling researchers to rapidly explore the predictability of a target variable and identify its main potential drivers. This flexible framework allows for the quick testing of predictive skill from different sources of predictability, making it a valuable asset for climate services. Although NN4CAST can use different variables to feed the model, we illustrate its application to reproduce tropical and extratropical teleconnections by training the model with Pacific SST anomalies. We show that NN4CAST can provide skilful predictions across timescales, from modelling variables at lag 0 that capture observed relationships to producing seasonal forecasts at longer leads, both in regions with linear SST-Atmosphere coupling (tropics) and in highly non-linear remote regions (such as Europe). Two key examples are the prediction of SST anomalies in the tropical Atlantic region during boreal spring from previous winter SSTs, and the modelling of precipitation anomalies over the European continent in boreal fall from the contemporaneous Pacific SSTs. The former exemplifies a predominantly linear ENSO-Tropical North Atlantic teleconnection, whereas the latter involves a highly non-linear and non-stationary ENSO-Euro-Atlantic teleconnection. Our results demonstrates NN4CAST's potential to determine and quantify the influence of specific potential drivers on a target variable, offering a useful tool for improving climate predictability assessments. NN4CAST enables the attribution of predictions to specific input features, helping to identify the relative importance of different sources of predictability over time and space. In summary, NN4CAST offers a powerful framework to better characterize and understand the complex, non-linear and non-stationary remote climate interactions.