Calibration of seasonal precipitation forecasts in Java (Indonesia) using bias-corrected precipitation and climate indices

03 August 2021

Seasonal rainfall forecasts help farmers make informed planning decisions about their livelihoods. Skilful rainfall forecasts can improve farming strategies in rain-fed agricultural production. In Indonesia, large-scale modes of climate variability have strong relationships with the seasonal rainfall. This makes them natural candidates for use as potential predictors in a statistical post-processing application. It is not known whether using climate indices as additional predictors in the statistical post-processing of ECMWF Seasonal Forecast System 5 (SEAS5) precipitation can improve skill. Lead author Dian Nur Ratri says "Indices of El Niño and the Indian Ocean Dipole are not needed as extra predictors to improve monthly precipitation forecasts for the first lead month in Java - Indonesia, except for September. However, for longer lead times in September and October, advanced statistical models that use only the climate indices are as skilful as models that use bias-corrected precipitation as the inputs".


Skilful forecasts of rainfall can improve farming strategies in rain-fed agricultural production systems and it is essential that these forecasts are also reliable and free of bias. Seasonal forecast models need to be calibrated and validated before they can be used in an optimal way, due to their inherent biases and dispersion errors. In our current study a more advanced statistical post-processing method is applied to investigate whether seasonal forecast skill can be further improved by using climate indices that are known to influence rainfall variability over Java island. 

There are two important modes of large-scale climate variability that influence precipitation in the region, namely El Niño-Southern Oscillation (ENSO) and the Indian Ocean Dipole (IOD). Another mode of climate variability that also significantly affects tropical weather is the Madden-Julian Oscillation (MJO) . We also consider SST around Java as one of the potential predictor variables in our current study.  We first compare the skill of a more advanced statistical post-processing method with a standard bias-correction method, as was applied in the region previously (Ratri et al. 2019). The second goal is to understand whether climate indices can improve forecast skill. This study focused on the dry months, i.e. July, August, and September plus October as the beginning of the wet season.


Using advanced statistical methods to improve seasonal precipitation forecasts

We use bias-corrected precipitation for the first 7-months lead times as the base input (Ratri et al. 2019), for the period of 1981 - 2010. The precipitation data are bias-corrected using empirical quantile mapping of daily precipitation reforecasts from the latest seasonal forecast model from ECMWF, "SEAS5". For the observations, as in Ratri et al. (2019) we again use the daily SA-OBS precipitation dataset (Van den Besselaar et al. 2017) over the period of 1981-2010. 

Statistical models learn relationships between observations and a set of potential predictor variables. As potential predictor variables we use the ensemble mean and standard deviation of raw and bias-corrected precipitation, the location of the grid cell (longitude and latitude), topography and ensemble mean SST around Java island (between 120 - 130◦E and 0 - 10◦S). Besides, we consider three climate indices as potential predictor variables in this study: the Nino3.4 index (from observations and the seasonal forecast), Dipole Mode Index (DMI) (from observations and the seasonal forecast), and an observed MJO index. 

We fit a single statistical model to all locations in Java, so that the model can learn from more data. Java has complex topography, with some mountains over 3000 m. We must first standardise the data so that we can use all locations in the one statistical model. Once the data is standardised, we use ensemble model output statistics (EMOS) to make seasonal precipitation forecasts. We test the inclusion of different predictor variables, and only keep the most important ones.

The CRPSS of the raw, bias-corrected and other post-processed forecasts.
Figure 1. Forecast skill measured by the continuous ranked probability skill score (CRPSS) of raw (red), EQM-corrected (pink), and the other post-processed forecasts valid in July, August, September and October.
The Brier skill score
Figure 2. BSS of EQM-corrected and other post-processed forecasts valid for the first lead month in July, August, September, and October. Box-and-whiskers indicate the BSS spread of the grid cells.


This study found that the post-processed precipitation only model (TP_only, in Figure 1) performs almost the same, and sometimes even better, than the post-processed models that have other additional potential predictor variables, for all months and most of the early lead times. SEAS5 has a reduced Equatorial Pacific cold tongue bias, which is accompanied by a more realistic ENSO amplitude and an improvement in ENSO prediction skill over the central-west Pacific (Johnson et al. 2019), and this suggests that SEAS5 is capturing the teleconnection well. Johnson et al. (2019) also highlighted that compared to System 4, SEAS5 is a better state-of-the-art seasonal forecast system which continues to display a particular strength in ENSO prediction. Besides, the strong negative correlation between SEAS5 ensemble mean precipitation and ensemble mean DMI indicates that that connection is also well represented in SEAS5, despite a bias in the DMI (Johnson et al. 2019). Therefore, we find that that Nino3.4 and DMI are not needed as extra predictors to improve monthly precipitation forecasts in Java for the first lead month, apart from the month July where the higher precipitation thresholds are more skilfully forecast if these climate indices are used as the only predictors (Fig. 2). However, for somewhat longer lead months, in September and October when there is more skill than climatology, the model that includes only Nino3.4 and DMI forecasts as potential predictors performs about the same compared to the model that uses only bias-corrected precipitation as a predictor.

This study found that simple bias-correction with EQM is hard to beat by a more advanced post processing method after the first lead month, except for October. For the first lead month the EMOS forecasts for higher precipitation amounts show higher skill than the EQM-corrected forecasts in all 4 months. This indicates that ECMWF SEAS5 precipitation contains much of the information from the predictors used here, but that forecasts for the first lead month (also regarding higher quantiles) can be improved by including information from climate indices.

The results were published in Weather and Forecasting.



Johnson et al., 2019: SEAS5: The new ECMWF seasonal forecast system. Geosci. Model Dev., 12, 1087–1117.