Decoding Climate Complexity: Novel Approaches to S2S Forecasting through Advanced AI Technology

In the realm of Subseasonal-to-Seasonal (S2S) AI forecasting, the integration of advanced computer models, remote sensing, and data assimilation has yielded a rich repository of historical weather data. For instance, the ERA5 dataset spans nearly eight decades, providing intricate insights into hourly weather patterns, including fine-scale temperature and wind fluctuations at resolutions of about 30 kilometers. Concurrently, Landsat data dating back to 1972 offers detailed snapshots of Earth's surface, captured with a resolution of approximately 30 meters. Enhanced by records of atmospheric conditions (IR soundings) since 1979, this comprehensive dataset, which incorporates inputs from both the Global Fire Emissions Database (GFED) and the Fire Inventory from the National Center for Atmospheric Research (FINN) dataset, has been enriched with CO2 emission data from fires since 1998. This amalgamation of resources greatly facilitates precise wildfire tracking. Additionally, the compilation of global rainfall data from sources such as NASA TRMM (1997) to GPM presents a comprehensive overview of worldwide precipitation trends. These combined datasets form a cohesive framework that significantly enhances our comprehension of past weather, fire incidents, and rainfall patterns, ultimately advancing our grasp of the intricate mechanisms driving Earth's climate.

Capitalizing on recent strides in deep machine learning, we introduce a novel neural network operator paradigm based on wavelet decomposition that addresses the challenges of long-term predictions in S2S forecasting. This approach revolves around constructing long-term reanalysis datasets by aggregating monthly or annual data, with a specific emphasis in this presentation on predicting 2m Temperatures. Building on prior AI forecasting work by Nvidia involving Fourier Neural Operators, our model ingeniously integrates multiple attention mechanisms, including Daubechies Wavelet Decomposition, Convolutional Neural Networks enhanced by Transformers, and Adaptive Neural Operators, all primed to decipher intricate spatio-temporal patterns embedded in climate data. Notably, the transformer-based architecture excels in sequence modeling, making it adept at disentangling relationships across extended timeframes in weather data. Initial results exhibit a remarkable performance advantage over persistence and climatology in the context of monthly predictions. This improvement is evident in the Root Mean Square Error (RMSE) metrics and the Anomaly Correlation Coefficient (ACC). Also, we are conducting experiments with various families of wavelets to assess their impact on the accuracy of our forecasts. We also plan to take this forward using a new vision transformer model called the Swin Transformer. This convergence of deep machine learning expertise and the wealth of reanalysis data emerges as a potent instrument for precise and extended-range predictions, promising implications for immediate weather forecasts, and broader insights into climate behaviors.

Moreover, we emphasize the potential extension of this AI architecture to include Subseasonal-to-Seasonal CO2 emissions forecasting arising from Boreal Forest wildfires, which could provide significant insights into climate feedback mechanisms. The increasing frequency of moisture-related droughts and extreme temperatures is especially relevant in understanding the heightened risk of natural wildfires. Notably, the expanse of Boreal forests across North America and Eurasia represents the world's largest forest biome. Within this context, we highlight the applicability of our neural model in the realm of Subseasonal-to-Seasonal (S2S) predictions, explicitly focusing on forecasting summer CO2 emissions across global Boreal Forests. We aim to encompass CO2 emission forecasting to glean invaluable insights into climate feedback mechanisms. Moreover, for this presentation, our primary focus will be on Subseasonal-to-Seasonal (S2S) predictions, particularly concerning the forecast of 2m Temperature variations.