Explainable deep learning for insights in El Niño and river flows

The El Niño-Southern Oscillation (ENSO) is a major driver of interannual climate variability globally, significantly influencing regional hydrology and weather patterns. Predicting ENSO and its downstream impacts on river flows is crucial for economic and societal planning, particularly in regions heavily reliant on river resources. However, current methods, including physics-based numerical simulations and data-driven models, struggle to accurately predict ENSO at interannual, decadal, and multidecadal timescales, limiting our ability to forecast ENSO’s hydrometeorological effects. Challenges in ENSO forecasting stem from various factors. Data limitations, such as the use of arbitrary rectangular regions to define ENSO indices, fail to capture the complexity of the phenomenon. ENSO is understood to be part of a larger system of interrelated SST oscillations which impact regional hydrometeorology. Furthermore, the relationship between ENSO and river flows is inherently nonlinear. This complexity necessitates methods capable of leveraging the complete information content from global SST data and identifying complex geographic dependence structures, encompassing both proximity-based dependence and long-range teleconnections. Existing methods like visual comparison, correlation analysis, mutual information, and linear regression often require heuristic feature selection and struggle with complex spatiotemporal features. Recent successes in climate science with deep learning methods offer a promising alternative, but the black-box nature of these models hinders interpretability and understanding of the underlying mechanisms. This paper aims to bridge this gap by applying explainable deep learning (XDL) techniques to gain insights into the ENSO-river flow relationship.

The literature highlights the importance of understanding ENSO's impact on river flows and the challenges in accurately predicting this complex relationship. Studies have shown the influence of ENSO on flood timings in Africa, interannual variability of flow in major rivers like the Ganges, Amazon, and Congo, and its broader effects on regional climate and hydrology. Previous research has explored the use of deep learning (DL) for improving ENSO prediction and complex networks (CN) for understanding teleconnections. However, there is a notable gap in understanding the predictive mechanisms, especially the “black box” nature of DL models, the limitations of simple ENSO indices in representing a complex phenomenon, and the challenges of translating DL-based ENSO predictions into river flow predictions. Several studies have emphasized the nonlinear nature of these relationships, calling for advanced techniques to capture complex dependencies and teleconnections. The use of XDL, combined with CN analysis, promises to enhance our predictive capability and understanding by providing insights into the spatiotemporal dynamics of ENSO and its impact on river systems. Existing methods such as sparse linear regression, while offering some interpretability, have limitations when dealing with high-dimensional spatiotemporal data.

The study employed a convolutional neural network (CNN) to predict monthly Amazon and Congo River flows using monthly SST data from Earth System Models (ESMs) and reanalysis datasets. The CNN architecture comprised four convolutional layers followed by three fully connected layers, utilizing ReLU activation functions and 2D max pooling. The input data consisted of global SST images, with different numbers of channels depending on the source (32 for ESMs, 3 for reanalysis, and 1 for averaged ESMs/reanalysis). The network was trained using the Adam optimizer, a squared loss function, and hyperparameters optimized through a validation set. Model performance was compared against an ensemble of machine learning models using only Niño 3.4 region indices (mean SST, anomaly). To address the black-box nature of deep learning, saliency maps were generated to visualize the importance of different SST regions for river flow predictions. Specifically, cyclical saliency maps (Cyclic-SM) were utilized to account for temporal periodicity in the data, creating yearly and seasonal saliency maps to provide better understanding of the spatiotemporal relationships. Additionally, complex network theory was applied to analyze the correlation structure of global SST data. Degree maps and teleconnections were constructed based on Pearson's correlation coefficients, identifying regions with high SST correlation and long-range teleconnections. Both ESM and reanalysis SST data were analyzed to compare model coupling strength and identify potential differences. The datasets used include multiple ESM simulations from NASA Earth Exchange (NEX), three reanalysis datasets (Hadley-OI SST, COBE SST, and ERSSTv5), and river flow data from UCAR. Preprocessing steps involved data alignment, interpolation to a common spatial resolution (1° longitude by 1° latitude), and handling missing data. The time period spanned from January 1950 to December 2005, with data split into training, validation, and testing sets.

The CNN model incorporating global SST data significantly outperformed models relying solely on Niño 3.4 indices in predicting three-month rolling mean river flows for both the Amazon and Congo rivers. Models using the wider SST region (41.5°S–37.5°N, 50.5°E–9.5°W) captured both the phase and amplitude of annual river flow fluctuations and components of interannual variation better than those utilizing only the Niño 3.4 region. The superior performance suggests that information beyond the canonical ENSO region is crucial for accurate prediction. Saliency maps revealed that the predictive power of ESMs originates primarily from the ENSO and Indian Ocean Dipole (IOD) regions, highlighting the significant link between these phenomena and regional hydrology. The Amazon River flow predictions exhibited salient areas in the tropical Pacific and Indian Oceans, while the Congo River predictions showed similar, albeit weaker, patterns. Reanalysis data resulted in more diffuse saliency maps, indicating weaker relationships between the predictor and predictand in reanalysis than in ESM data. Complex network analysis showed that ESM SST data exhibited a stronger correlation structure, with numerous teleconnections between the tropical Pacific, Indian, and Atlantic Oceans, particularly along the equator. The correlation structure in reanalysis SST was weaker, with fewer long-distance connections. This difference in coupling strength between ESMs and reanalysis/observations is consistent with existing literature and suggests that data-driven approaches can potentially quantify and bridge this gap. Further analyses of the histograms of connection distances highlighted qualitative differences in correlation structure among individual ESMs.

The findings demonstrate the substantial advantages of utilizing XDL methods in combination with climate network analysis for enhanced prediction and understanding of ENSO-driven river flows. The superior predictive performance of the CNN model using global SST data over models based solely on ENSO indices highlights the importance of considering broader spatiotemporal patterns in SST for improved hydrological forecasting. The interpretability provided by saliency maps and the insights gained from complex network analysis offer valuable physical insights into the underlying mechanisms governing ENSO-river flow teleconnections. The observed differences in correlation structure between ESMs and reanalysis data emphasize the potential of data-driven methods to bridge the gap between model simulations and observations in climate science. The results suggest that incorporating additional data and data-driven technologies could lead to a more comprehensive understanding of causal relationships in Earth systems and inform climate adaptation strategies by providing more robust and reliable projections of river flows.

This study successfully demonstrated the use of explainable deep learning to improve the prediction of ENSO-driven river flows and enhance our understanding of the underlying mechanisms. The integration of global SST data, saliency maps, and complex network analysis provided valuable insights into the spatiotemporal dynamics of ENSO and its teleconnections with river systems. Future research could focus on expanding the analysis to include additional river basins, exploring other climate indices beyond ENSO, and integrating other environmental factors influencing river flows. Furthermore, refining XDL techniques and developing more sophisticated network models could improve prediction accuracy and robustness.

The study was limited by the availability of data, with the time period spanning from January 1950 to December 2005. This restricted dataset may limit the generalizability of the findings to other periods. While the CNN model showed significant improvement over traditional methods, further research is needed to fully evaluate its performance against other sophisticated machine learning techniques. Additionally, while the study utilized multiple ESM and reanalysis datasets, the inherent uncertainties and biases associated with these datasets could influence the findings. Finally, focusing primarily on two major river basins may not be fully representative of all ENSO-driven river flow variability across the globe.