Decadal temperature prediction is vital for assessing the impacts of climate change on food security, health, and biodiversity. Accurate predictions at a fine-scale (e.g., weekly) are particularly important for informing agricultural practices and mitigating the economic losses associated with extreme temperature fluctuations. Existing methods, such as simulation-based models (numerical weather prediction and global circulation models) and machine learning (ML) approaches, face challenges in handling the chaotic nature of temperature variations and the accumulation of prediction errors over long time horizons. Simulation-based methods struggle with uncertainty due to the chaotic nature of climate systems, while ML methods often require the latest ground truth data to avoid error accumulation. This necessitates a novel method capable of long-term, fine-scale temperature prediction while addressing these limitations.

Simulation-based methods, including numerical weather prediction (NWP) and global circulation models (GCMs), have been widely used for decadal temperature prediction. NWP models focus on regional scales and employ simplifying assumptions to manage computational complexity, while GCMs have evolved since the 1970s to incorporate broader climate systems. A major challenge for simulation-based methods is the inherent uncertainty due to the chaotic nature of climate systems. Ensemble methods, such as those used in the Coupled Model Intercomparison Project (CMIP), mitigate this uncertainty by combining predictions from multiple runs with slightly varied initial conditions. However, ensemble methods can be computationally expensive and struggle with the lack of up-to-date data for calibration during long-term predictions. Recent research has explored ML-based methods for temperature prediction, demonstrating promising results in short-term forecasts. However, these methods also suffer from limitations in long-term prediction due to error accumulation. Hybrid workflows combining ML and climate models have shown some success but often retain the computational costs and sensitivity to initial conditions of climate simulation models. This study aims to overcome these limitations by proposing a novel method that specifically addresses the error accumulation problem in decadal temperature prediction.

The proposed method comprises two main components: a dependency learning component (DLC) and an information tracking component (ITC), both implemented using multilayer perceptrons (MLPs). The DLC learns a predictive model from historical temperature data (1880-2010), capturing the spatiotemporal dependencies in the data. The ITC is designed to reduce prediction uncertainty. It tracks changes in temperature variation during the prediction phase (2011-2020) by providing probabilistic feedback on the next-step prediction error based on the current prediction. This feedback, quantified by the first-order difference of the temperature time series, is integrated into the model's objective function as a regularizer. The ITC essentially acts as a calibrator that prevents the amplification of initial errors. The model architecture is a five-layer MLP for the DLC and an encoder-decoder structure for the ITC, both using ReLU activation functions. Training uses data from 1880 to 2010, aiming to predict global weekly temperatures from 2011 to 2020. The objective function minimizes both the difference between predictions and ground truth, and the difference between the estimated prediction error (from the ITC) and the actual first-order difference of the ground truth. During prediction, the model uses its previous prediction as input for both DLC and ITC to generate the next step's prediction, using the ITC's feedback to compensate for potential errors. The evaluation metrics employed are Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Dynamic Time Warping (DTW). A sliding window approach with a window size of 50 is used during the prediction phase. Data used is the Global Daily Land (Experimental; 1880–Recent) data with a resolution of 1 degree × 1 degree Latitude–Longitude Grid, aggregated to weekly averages.

The proposed method outperforms several baseline methods in ten-year-ahead and five-year-ahead predictions of global weekly land surface temperature. These baseline methods include three classical sequence learning methods (LSTM, GRU, MLP), a state-of-the-art deep learning method for long sequence prediction (Informer), a statistical method (VMD-ARIMA), and simulation-based models from CMIP5 and CMIP6. Tables 1 and 2 show that the proposed method achieves significantly lower MAE, RMSE, and DTW values than the baselines. Figure 2 visually demonstrates the accuracy of the proposed method's predictions for the first week of January and July in 2020, showing a good match between predicted and actual temperatures on a global scale. Figure 3 and supplementary figures provide more detailed predictions for 12 countries selected based on their temperature variability, climate risk, sustainable development index, and vulnerability to temperature increases. The figures show that the proposed method accurately captures the complex and irregular temperature patterns in these countries, even over periods with high variability. The supplementary materials provide additional analyses showing correlations between predicted temperature sequences and their relationship to climate phenomena. The theoretical analysis included in the supplementary materials proves that the information tracking component effectively prevents exponential error accumulation, maintaining error at a constant level during the prediction phase.

The superior performance of the proposed method highlights the potential of data-driven approaches in decadal climate prediction. The incorporation of the ITC is key to its success, allowing the model to adapt to the dynamic and chaotic nature of temperature variations. The use of the first-order difference to supervise the ITC provides a way to capture and predict the changes in entropy of the temperature system. This suggests that the method’s ability to predict temperature accurately is intrinsically linked to its ability to track and forecast information changes within the system. The results show that the proposed method can achieve accurate long-term predictions even without real-time ground truth data for calibration. This addresses a major limitation of previous ML-based methods. The method's effectiveness in predicting temperature patterns across diverse geographical regions, with varied climate characteristics, further demonstrates its robustness and generalizability. The selection of countries for analysis ensures a comprehensive coverage of various climatic conditions and temperature patterns, adding strength to the findings.

This study presents a novel method for decadal weekly land surface temperature prediction that effectively addresses the challenges posed by the chaotic nature of climate systems. The integration of the dependency learning component and information tracking component enables accurate long-term predictions, outperforming existing approaches. The method's success demonstrates the potential of data-driven approaches in climate prediction. Future research could extend this method to incorporate additional climate factors and explore its applications in other domains.

The study focuses on land surface temperature, and the findings may not directly generalize to other climate variables. The model's performance relies heavily on the quality and availability of historical data. While the method successfully addresses error accumulation, potential biases in the training data could still affect the accuracy of predictions. Further investigation into model interpretability and uncertainty quantification could provide additional insights.