Achiving the Precipitation Best prediction giving the NAO index

The boreal winter precipitation forecast, acumulated from November to March, can be improved by considering the NAO index. The first step is to find the best estimation of winter NAO, given by two Seasonal Forecast System (SFS). The second step is to employ the enhanced NAO index pdf to produce weights for a SFS (it could be the same or a different SFS from the previous one). The third step is to apply this weights to a precipitation field. The methodology has been proved to improve the skill on precipitation forecast in the iberian peninsula given the relation between the winter precipitation and the NAO index in seasonal time scale (Sánchez-García, E., Voces-Aboy, J., Navascués, N., & Rodríguez-Camino, E. (2019). Regionally improved seasonal forecast of precipitation through Best estimation of winter NAO, Adv. Sci. Res., 16, 165174,

This document is aim to ilustrate a practical use of the functions included in CSTools applying this methodology.

Loading packages and data

Open an R sesion and load library CSTools:


The required data to applied this methodology are:

Given the memory limitations, the following example uses synthetic data.

The first SFS is a dynamical model containing 25 ensemble members and its output will be save in the object NAO_hind1 for the 20 years reference period and NAO_fcst1 for a next season forecast. The second SFS is a empirical model, so, the NAO_hind2 and NAO_fcst2 are characterized by a mean and standard deviation saved in the 'statistics' dimension.

The synthetic data is created by running the following lines:

NAO_obs <- rnorm(3, sd=3)
dim(NAO_obs) <- c(time = 3)
NAO_hind1 <- rnorm(3 * 6, sd=3)
dim(NAO_hind1) <- c(time = 3, member = 6)
NAO_hind2_mean <- rnorm(3, sd=3)
NAO_hind2_sd <- rnorm(3, mean=5, sd=1)
NAO_hind2 <- cbind(NAO_hind2_mean,NAO_hind2_sd)
dim(NAO_hind2) <- c(time=3, statistic=2)

The acumulated precipiation field (only march for memory limiations) is loaded by running:

lonlat_prec$data <- apply(lonlat_prec$data, c(1, 2, 3, 5, 6), sum)

1- Best Estimate Index NAO

The function PDFBest performes the next steps:

pdf_hind_best <- BEI_PDFBest(NAO_obs, NAO_hind1, NAO_hind2, index_fcst1 = NULL,
                             index_fcst2 = NULL, method_BC = 'none',
                             time_dim_name = 'time', na.rm = FALSE) 

2- Compute weights using the Best Estimation of Index NAO

An array of weights is calculated for a SFS. This SFS could be a different SFS than the ones in section 1. The function WeightIndex computes these weights for each ensemble's member based on the best NAO PDF estimate.

weights_hind <- BEI_Weights(NAO_hind1, pdf_hind_best)
names(dim(weights_hind))[1] <- 'sdate'

The expected dimensions of these weights are 'member' and temporal dimension.

3- Apply weights to a precipitation field

The function CST_BEI_Weighting computes the ensemble mean or the terciles probabilities for a climate variable. The corrected precipitation field is obtained by running:

em_prec <- CST_BEI_Weighting(lonlat_prec, weights_hind, type = 'ensembleMean', time_dim_name = 'sdate')
prob_prec <- CST_BEI_Weighting(lonlat_prec, weights_hind, type = 'probs', )

Comparison and visualization

The original model output can be compared against the BEI corrected field. To do this a equiprobability weights are created and they are applied to the original precipitation field:

aweights_raw <- rep(1/6, 3 * 6)
dim(aweights_raw) <- dim(weights_hind)

em_prec_raw <- CST_BEI_Weighting(lonlat_prec, aweights_raw, type = 'ensembleMean', time_dim_name = 'sdate')
prob_prec_raw <- CST_BEI_Weighting(lonlat_prec, aweights_raw, type = 'probs', time_dim_name = 'sdate')

A map with the probability that the total precipitation will be in the lower/normal/upper tercile based on the Best Estimate Index NAO could be obtained using the 'PlotEquiMap' function or 'PlotMostLikelyQuantileMap' function from 'CSToools' package.

The following figures show the probabilities for lower tercile for precipitation from November to March 2012/13 for ECMWF S5 system applying the methodology exposed or not, obtained using real data:



In a similar way we can plot the map with the probability that the total precipitation from November 2012 to March 2013, for example, will be in the lower tercile from ECMWF Seasonal Forecast System 5 (raw) to compare results running the code: