Introduction
While climate change research often focuses on rising mean temperatures, the volatility of temperature—its variance over time—is equally critical for ecological stability and human infrastructure. High volatility manifests as extreme weather events, such as heatwaves and cold snaps, which pose greater adaptive challenges than gradual mean shifts. However, estimating non-parametric trends in volatility across massive, multi-resolution global datasets presents significant computational hurdles. Standard optimization solvers struggle with the millions of observations and high spatial-temporal correlations inherent in climate data.
Methodology
This paper proposes a non-parametric model for variance estimation based on $\ell_1$-trend filtering. The model assumes that detrended temperature observations follow a normal distribution with a smoothly evolving variance. To handle the scale of global grids, the authors developed two algorithms based on the Alternating Direction Method of Multipliers (ADMM):
- Consensus ADMM: A distributed approach that decomposes the global space-time data into smaller "sub-cubes." Each sub-problem is solved independently before being reconciled into a global consensus, making it ideal for parallel computing environments.
- Linearized ADMM: A faster, single-machine alternative that utilizes proximal operators and the Lambert W function for element-wise updates, avoiding costly matrix inversions.
Datasets
The algorithms were validated using a synthetic $5 \times 7$ grid over 780 time steps to benchmark against ground truth. For real-world application, the authors utilized the ERA-20C Dataset, analyzing daily noon temperatures for the Northern Hemisphere from 1960 to 2010 on a $1^\circ \times 1^\circ$ grid.
Key Results and Insights
Computational Performance
In synthetic benchmarks, the Linearized ADMM demonstrated remarkable speed, reaching convergence in approximately 4 seconds on a single machine, compared to 2 hours for the Consensus ADMM. However, the Consensus approach remains the preferred choice for truly massive datasets that exceed the memory of a single node.
Validation and Climate Trends
Validation on synthetic data showed that the trend-filtering approach accurately captures seasonal cycles and underlying volatility shifts. As shown in the figure below, the choice of the regularization parameter $\lambda_t$ allows for varying degrees of smoothness in the estimated trend.
Applying the model to the ERA-20C data revealed a general decrease in temperature volatility across much of the Northern Hemisphere between 1961 and 2011. This finding supports the "Arctic amplification" theory, suggesting that as polar regions warm faster than the tropics, the reduced temperature gradient stabilizes weather patterns. A notable exception was Southeast Asia, which exhibited a distinct positive trend in volatility.
Discussion and Future Work
The authors highlight that variance is a robust metric for satellite-derived data, as it is less susceptible to the instrument biases that often plague mean temperature records. Future work will focus on a hybrid algorithmic approach—splitting problems into sub-cubes but solving them via the Linearized method—to further optimize performance. Additionally, the researchers aim to expand the analysis to include cloud reflectance and other multi-spectral global datasets.
Summary updated