Forecasting near-surface air temperature via SARIMA and LSTM: A regional time-series study

Abstract

Accurate modeling of near-surface air temperature (AT) trends is critical for assessing global and regional climate risks, particularly in light of the intensifying warming signals observed across the northern hemisphere and the tropics. This study proposes a robust and computationally efficient framework for forecasting near-surface AT across the global, the northern hemisphere, the southern hemisphere, and the tropics using two complementary time-series modeling techniques: seasonal autoregressive integrated moving average (SARIMA) and long short-term memory (LSTM) networks. The models are trained to capture both structured seasonal patterns and nonlinear temporal dynamics by leveraging the ERA5 reanalysis dataset (1970–2024) and incorporating preprocessing steps such as detrending and Z-score normalization. SARIMA consistently outperformed LSTM across most domains, particularly in the global region, achieving lower RMSE (0.0967 °C) and higher correlation (R = 0.9975), reflecting its superior capacity for linear and seasonal signal extraction. Quantitatively, SARIMA demonstrates 5%–10% lower RMSE and slightly higher correlation than LSTM across all domains, underscoring the statistical significance of its performance advantage. Projected near-surface AT anomalies by 2050 reveal a marked warming trend, with the SARIMA model estimating a global anomaly of +1.078 °C and a northern hemisphere anomaly of +1.474 °C, closely aligning with IPCC-reported trajectories and exceeding CMIP5 RCP4.5 projections. The findings underscore SARIMA’s reliability for short- to mid-term near-surface AT forecasting and LSTM’s potential for future hybrid modeling schemes. This work fills a critical methodological gap by integrating statistical rigor with scalable deep learning, offering enhanced fidelity for regional climate adaptation planning.