Improving the Skill of CFSv2 Seasonal Forecast over Northern South America

The precipitation is one of the most important and challenging variables to forecast in the field of water resources management. Being the main input in river discharge forecast models, the accuracy of precipitation forecasts is critical for several applications and decision-making in water security and drought management.

In this research, the seasonal precipitation and circulation variables forecasts from the Climate Forecast System version 2 (CFSv2) are evaluated over the northern region of South America, using the data from the Climate Forecast System Reanalysis (CFSR) and the Tropical Measurement Mission (TRMM) as observations. The skill evaluation considers the seasonal dependence, spatial distribution and the biases, parametric and non-parametric correlation coefficients, RMSE and other indices including the Skill Score (SS), Heidke Skill Score (HSS) and Linear Error in Probability Space (LEPS). The results show that the CFSv2 is able to reproduce the observed precipitation variability over this region with relatively good accuracy but has considerable biases. The performance of the model seems to be affected by the complex topography of the Andes, which is reflected in the lower skill results that are achieved over this region. The bias does not seem to increase with lead time, but the anomaly correlations decrease rapidly.

The bias in precipitation is corrected using the quantile mapping technique, which allows to improve the skill in terms of the amplitude rather than the variability. A new probabilistic non-parametric methodology to estimate precipitation from circulation variables is proposed based on the joint probability distribution of the circulation variables from the CFSv2 and the observed precipitation. The resulting precipitation forecast is considerably more skillful than the raw CFSv2 forecast and the methodology allows the generation of a probabilistic rather than deterministic predictions, which allows to give a range of precipitation variation exploring the uncertainty.

The skill of SST and climate variables forecasts in the tropical belt and the ability of the model representing the observed precipitation teleconnection with the Pacific and Atlantic Ocean is also analyzed given the influence of this oceans in the region of interest. The analyses show that the skill of SST and climate variables is considerably high in the east-central Pacific and over the tropical Atlantic. The relationship between these oceans and the climate in northern South America is well simulated at short lead times, while at longer lead times the correlation pattern is lost. In the Atlantic Ocean, the area of high correlations with precipitation match reasonably well with the observations, but the influence of this ocean seems to be stronger in the model over some regions of the domain.

Climate indices with strong teleconnections are identified and integrated in different hybrid statistical models with the post-processed precipitation from CFSv2, developing dynamical-statistical models that further improve the skill of seasonal forecasts. The impact of the choice of climate indices and the post-processing methodologies is also discussed to find the best combination for each region of the domain.