Spatiotemporal characteristics of the vertical structure of predictability and information transport in Northern Hemisphere

Based on nonlinear prediction and information theory, vertical heterogeneity of predictability and information loss rate was obtained over the Northern Hemisphere. In seasonal to interannual time scales, the predictability is low in lower troposphere and high in mid-upper troposphere. But within mid-upper subtropics troposphere over some ocean area there is a relatively poor prediction. The conclusions fit the seasonal time scale too. When it goes to the interannual time scale, the predictability is high in lower troposphere and low in mid-upper troposphere contrary to the formers. And on the whole the interannual trend is more predictable than the seasonal trend. The average information loss rate is low over mid-east Pacific, west of North America, Atlantic and Eurasia, and the atmosphere over other places have a relatively high information loss rate in all time scales. Two channels were found steadily over the Pacific Ocean and Atlantic Ocean in subtropics. There exist unstable channels as well. The four-season impact on predictability and information communication was studied too. The predictability is low no matter what season data are removed and each season plays an important role in the existence of the channels except the winter. The predictability and teleconnections are paramount issues in atmospheric science, and the teleconnections may be established by communication channels. So this work is interesting since it reveals the vertical structure of predictability distribution and channel locations, the contribution of different time scales to them and their variations with different seasons.

💡 Research Summary

The paper investigates the vertical and horizontal distribution of atmospheric predictability and information transport over the Northern Hemisphere by combining a nonlinear prediction framework with concepts from information theory. Using the NCEP/NCAR reanalysis from 1966 to 2008, the authors extract monthly mean fields on a 5° × 5° grid at twelve pressure levels (100–1000 hPa). Seasonal (30‑90 day) and interannual (1‑3 year) components are separated with the Zheng‑Dong filtering method.

A nonlinear prediction model based on phase‑space reconstruction (embedding dimension = 5, delay = 2) is trained on a 40‑year (480‑month) segment and then used to forecast 12 months ahead in a moving‑window fashion. Predictability is quantified by the Pearson correlation coefficient (r) between predicted and observed values; r ≈ 1 indicates perfect predictability, r ≈ 0 indicates no skill.

Information loss is assessed through the Kolmogorov entropy K, approximated from the dynamics, and the loss rate λ = K/τ (τ being the embedding delay). Low λ implies that information is retained longer and can be transmitted more efficiently across the atmosphere.

Key findings:

1. Vertical predictability structure – In the seasonal time scale, predictability is low in the lower troposphere (100‑700 hPa) and higher in the mid‑upper troposphere (600‑100 hPa). This reflects the dominance of large‑scale planetary waves that persist longer aloft. In contrast, at the interannual scale the pattern reverses: the lower troposphere becomes more predictable while the mid‑upper layers lose skill, suggesting that ocean‑driven phenomena (e.g., ENSO, PDO) imprint their signal primarily on the lower atmosphere.

2. Spatial patterns of information loss – The average λ is markedly low over the mid‑east Pacific, the western side of North America, the Atlantic basin, and Eurasia, indicating these regions act as “information highways.” Two persistent channels of low information loss are identified in the subtropical Pacific (approximately 20°‑30° N, 150°‑180° E) and the subtropical Atlantic (approximately 20°‑30° N, 80°‑60° W). Additional unstable channels appear intermittently, likely linked to transient dynamical regimes.

3. Seasonal influence on channels – When each season’s data are removed in turn, the two main channels remain robust for spring, summer, and autumn, but disappear when winter data are omitted. This points to winter‑time circulation features (e.g., the Arctic Oscillation, jet‑stream configuration) being essential for maintaining the identified pathways.

4. Methodological considerations – The study fixes the embedding parameters (n = 5, τ = 2) without systematic optimization, potentially limiting model performance for different pressure levels. Using only Pearson r as a skill metric may under‑represent nonlinear forecast skill. The Kolmogorov entropy is not directly computed but inferred from approximations, which could introduce bias into λ estimates. The coarse spatial resolution (5°) and limited vertical sampling (12 levels) restrict the ability to resolve finer‑scale processes such as mesoscale convection.

5. Implications – By revealing that predictability shifts vertically with time scale and that specific low‑loss corridors exist, the work supports the view that atmospheric teleconnections are mediated not merely by wave propagation but also by preferential information pathways. The finding that interannual predictability is generally higher than seasonal predictability, especially in the lower troposphere, underscores the importance of ocean‑atmosphere coupling for medium‑range forecasting.

Overall, the paper provides a novel, integrated perspective on how predictability and information transport vary with altitude, geography, and time scale. Future work could improve upon the present analysis by employing higher‑resolution datasets, testing alternative nonlinear predictors (e.g., machine‑learning ensembles), and directly estimating Kolmogorov entropy to refine the information‑loss diagnostics. Incorporating the identified channels into operational forecast models may enhance skill for both seasonal and interannual predictions.