Do CFSv2 Seasonal Forecasts Help Improve the Forecast of Meteorological Drought over Mainland China?

Yang Lang, Lifeng Luo, Aizhong Ye, Qingyun DuanWater 2020 · DOI

TL;DR

This paper investigates whether the NCEP Climate Forecast System version 2 (CFSv2) can improve meteorological drought forecasts in China compared to a simple persistence forecast. The study finds that CFSv2 generally enhances drought prediction skills, particularly at longer lead times and in the humid southeastern regions, though its contribution varies significantly by season and location, with some areas showing no or even negative improvements.

What Problem Does This Paper Address?

Drought is a significant natural hazard in China, impacting agriculture, water resources, and the economy. Accurate seasonal drought prediction is crucial for minimizing these impacts. While dynamical climate models like CFSv2 offer potential for improved forecasts, their actual added value over simpler methods, especially in specific regions like China, needs thorough evaluation. This paper aims to quantify the extent to which CFSv2 can improve meteorological drought forecasts in China, moving beyond general global assessments to a regional focus. It specifically addresses how CFSv2's contributions vary geographically and seasonally, and whether there are instances where it performs worse than a basic persistence forecast.

Methods

The study evaluates the skill of CFSv2 forecasts for meteorological drought using the six-month Standardized Precipitation Index (SPI6) as the drought indicator. The primary benchmark for comparison is a persistence forecast, which assumes future precipitation anomalies will be the same as recent past anomalies. CFSv2 forecasts are bias-corrected to account for systematic model errors before being combined with observed precipitation to create SPI6 forecasts at various lead times (0 to 6 months).

Forecast skill is quantified using two main metrics:

  • Temporal Anomaly Correlation (TAC): Measures the linear relationship between forecast and observed SPI6 anomalies.
  • Root Mean Square Error (RMSE): Measures the average magnitude of the errors between forecast and observed SPI6.

To assess the improvement offered by CFSv2, skill scores are calculated, comparing CFSv2's performance against the persistence forecast. A positive skill score indicates improvement, while a negative score suggests the CFSv2 forecast is inferior. Additionally, spatial anomaly correlation (SAC) is used to evaluate how well forecasts capture observed spatial patterns of drought. The analysis considers both overall performance and variations across different lead times, seasons, and geographical regions within China.

Datasets

  • CFSv2 Reanalysis Data: A 25-year (1982-2006) reforecast dataset from the NCEP Climate Forecast System version 2 (CFSv2) was used. This dataset includes monthly mean precipitation from 9-month forecast runs.
  • Observational Precipitation Data: Daily precipitation analysis data for East Asia, developed by Xie et al. [24], covering the period 1961-2007 at a 0.5° × 0.5° resolution. This dataset was used for bias correction and for evaluating both the persistence and CFSv2 forecasts.

The study uses the 25-year common period (1982-2006) for evaluation. The persistence forecast serves as the baseline benchmark, and CFSv2 forecasts are evaluated against it.

Key Insights and Ideas

  • Persistence is a Strong Baseline: The SPI6 persistence forecast demonstrates surprisingly good skill across China, especially at short lead times (1-3 months). This highlights the inherent predictability of precipitation anomalies due to their temporal persistence.
  • CFSv2 Adds Value, Especially Later: While persistence is strong initially, CFSv2 forecasts consistently improve drought prediction skills in most regions, particularly at longer lead times (beyond 3 months). This suggests dynamical models become more critical when the inherent persistence of precipitation starts to wane.
  • Regional and Seasonal Nuances are Critical: The benefit of CFSv2 is not uniform. Humid southeastern regions show the most pronounced improvements, while arid northwestern regions sometimes see degradation in forecast skill when CFSv2 is used. Seasonal cycles of precipitation and inter-annual variability strongly influence both persistence skill and CFSv2's added value.
  • "Negative Contribution" Zones Exist: The study identifies specific regions and time periods where CFSv2 forecasts perform worse than the simple persistence forecast. This is a crucial finding, suggesting that blindly applying dynamical model forecasts without considering their regional and seasonal limitations can be detrimental.
  • Spatial Patterns Matter: Beyond just predicting the magnitude of drought, CFSv2 also shows a tendency to provide spatially coherent forecasts, contributing positively to capturing drought patterns.

Key Results

  • Overall Improvement: CFSv2 forecasts generally outperform persistence forecasts for meteorological drought prediction in China, especially at longer lead times.
  • Lead Time Dependency: At 1-month lead time, persistence is already skillful (TAC > 0.8). CFSv2 provides marginal improvements. However, at 3- and 5-month lead times, CFSv2's contribution becomes more significant, particularly in the humid southeast.
  • Regional Differences: Humid regions (e.g., southeast) show larger improvements in TAC and reductions in RMSE with CFSv2. Arid regions (e.g., northwest) sometimes experience a degradation of skill.
  • Seasonal Variations: Forecast skill and CFSv2 contributions vary significantly by season. For instance, persistence skill is generally higher during dry seasons and lower during wet seasons.
  • Maximum Lead Time: Figure 6 illustrates the maximum lead time (in months) for which CFSv2 shows better skill than persistence. This varies widely, with some areas benefiting for up to 5 months, while others show little to no benefit.
  • Negative Contributions: Specific initialization periods and target months can lead to CFSv2 performing worse than persistence, particularly in certain regions. For example, forecasts initialized in January for February-June periods sometimes showed poor performance.

Limitations and Future Work

  • Deterministic vs. Probabilistic: The study treated CFSv2 forecasts as deterministic (using ensemble means). A full evaluation using probabilistic metrics (considering ensemble spread) could offer additional insights.
  • Bias Correction Limitations: While bias correction was applied, it might not fully address all model deficiencies, especially in regions with poor CFSv2 performance.
  • Poor Model Performance Instances: The paper notes instances of "abnormally poor performance" from CFSv2 in specific initialization periods. Identifying the causes of these issues is crucial for model improvement.
  • Data Requirements: Advancing seasonal drought forecast capability in China requires high-quality observations of precipitation and temperature.
  • Future Directions: The authors suggest that understanding these regional and seasonal variations is critical for applying CFSv2 forecasts to other applications like seasonal hydrological predictions. Future work should focus on addressing the identified issues, potentially through improved bias correction, ensemble forecasting techniques, and more targeted model development.

Summary updated

Abstract

Seasonal forecasts from dynamical models are expected to be useful for drought predictions in many regions. This study investigated the usefulness of the Climate Forecast System version 2 (CFSv2) in improving meteorological drought prediction in China based on its 25-year reforecast. The six-month standard precipitation index (SPI6) was used as the drought indicator, and its persistence forecast served as the benchmark against which CFSv2 forecasts were evaluated. The analysis found that the SPI6 persistence forecast shows good skills in all regions at short lead times, and CFSv2 forecast can further improve those skills in most regions. The improvement is particularly pronounced at longer lead times and over the humid regions in the southeast. This study also examined the seasonality and regionality of persistence forecast skills and CFSv2 contributions, and reveals regions where CFSv2 forecast shows no or sometimes even negative contributions.
BibTeX
@inproceedings{lang2020do,
  author = {Yang Lang and Lifeng Luo and Aizhong Ye and Qingyun Duan},
  title = {{Do CFSv2 Seasonal Forecasts Help Improve the Forecast of Meteorological Drought over Mainland China?}},
  booktitle = {Water},
  year = {2020},
  doi = {10.3390/w12072010},
}
Figure