HTML
--> --> -->The relationship between ENSO and the East Asian monsoon is complex and not fully understood. However, it has long been clear that a strong El Ni?o peaking in winter is likely to be followed by above-average rainfall in China the following summer (e.g., Stuecker et al., 2015; He and Liu, 2016; Xie et al., 2016; Zhang et al., 2016a, b), although this response is not symmetric under La Ni?a conditions (Hardiman et al., 2017). The extreme El Ni?o event of 1997/98 was followed by devastating floods in the Yangtze River basin (Zong and Chen, 2000; Ye and Glantz, 2005; Yuan et al., 2017): thousands of people died, millions were made homeless, and the economic losses ran into billions of CNY. In the subsequent years, much work has gone into better water management and flood prevention, and into improving both the accuracy and communication of climate forecasts, to prevent such a disaster happening again.
Seasonal rainfall forecasts across China have long been produced based on statistical relationships with large-scale climate phenomena, rather than forecasting precipitation directly from dynamical models. For example, (Zhu et al., 2008) and (Li and Lin, 2015) both examined the skill of 500 hPa geopotential height (z500) data, from multi-model ensembles of dynamical seasonal forecast systems, for forecasting summer monsoon and Yangtze river valley rainfall, respectively. (Tung et al., 2013), however, found that using sea level pressure performed better than using z500 when forecasting station-scale summer rainfall in southern China. (Kwon et al., 2009), (Peng et al., 2014), (Wu and Yu, 2016), (Xing et al., 2016) and (Zhang et al., 2016b) all investigated different statistical approaches to forecasting summer precipitation in China, based on various observational indices derived from sea surface temperatures (SSTs), air temperature and pressure. In many cases, these showed an improvement over dynamical models. (Wang et al., 2013) found that both dynamical models and a statistical model based on SSTs and pressure were able to predict the variability in the West Pacific subtropical high, which was itself shown to be a good predictor of East Asian summer monsoon rainfall. Statistical downscaling techniques have also been shown to improve predictions of summer precipitation in China over global dynamical forecast models (e.g., Ke et al., 2011; Liu and Fan, 2012).
Recent advances in the dynamical seasonal forecast system developed at the UK Met Office, GloSea5 (MacLachlan et al., 2015), have resulted in the development of operational and prototype climate services for the UK in many sectors (e.g., Svensson et al., 2015; Palin et al., 2016; Clark et al., 2017). Recent work has shown that GloSea5 also has useful levels of skill for various processes in China (Bett et al., 2017; Lu et al., 2017), including for summer precipitation over the Yangtze River basin (Li et al., 2016), without having to use statistical models based on larger-scale drivers.
In parallel to these findings, (Golding et al., 2017) demonstrated that there was a clear demand from users for improved seasonal forecasts for the Yangtze, both from the flood risk and hydropower production communities. The very strong El Ni?o that developed during the winter of 2015/16 (Zhai et al., 2016) provided a perfect opportunity to develop a trial operational seasonal forecast using GloSea5 for the subsequent summer of 2016.
We therefore produced forecasts for the upcoming three-month period each week, from February (forecasting March-April-May) to the end of July 2016 (forecasting August-September-October); our focus, however, was on forecasting the June-July-August (JJA) period, as that was where (Li et al., 2016) had demonstrated skill. In the last week of each month, a forecast for the coming season was issued by the Met Office to the China Meteorological Administration (CMA).
In this paper, we describe the observed rainfall in the Yangtze region in summer 2016, and assess how the real-time forecasts for May-June-July (MJJ) and JJA performed, with a range of lead times from zero to three months. We describe in section 2 the datasets used, and in section 3 our forecast production methodology. In section 4 we compare the forecasts to the observed behavior, and in section 5 discuss possible future developments.
Using this configuration, GloSea5 runs operationally, producing both forecasts and corresponding hindcasts (intended to bias-correct the forecasts). Each day, two initialized forecasts are produced, running out to seven months. To produce a complete forecast ensemble for a given start date, the last three weeks of individual forecasts are collected together to form a 42-member lagged forecast ensemble.
At the same time, an ensemble of hindcasts is produced each week. As described by (MacLachlan et al., 2015), three members are run from each of four fixed initialization dates in each month (the 1st, 9th, 17th and 25th), for each of the 14 years covering 1996-2009. The full hindcast ensemble is made by collecting together the four hindcast dates nearest to the forecast start date, yielding a 12-member, 14-year hindcast. Note that the hindcast was extended at the end of April 2016 to cover 23 years (1993-2015).
This operational hindcast is not intended to be used for skill assessments: with only 12 members, skill estimates would be biased low (Scaife et al., 2014). However, a separate, dedicated hindcast was produced for skill assessment, with 24 members and 20 years. Using that hindcast, we find a correlation skill of 0.56 for summer Yangtze rainfall, statistically indistinguishable from the previous value of 0.55 found by (Li et al., 2016).
We use precipitation data from the Global Precipitation Climatology Project (GPCP) as our observational dataset. This is derived from both satellite data and surface rain gauges, covering the period from 1979 to the present at a spatial resolution of 2.5° (Adler et al., 2003). The verification we present here uses version 2.3 of the data (Adler et al., 2016). Only version 2.2 was available when we started our operational trial, although we have confirmed that the choice of version 2.2 or 2.3 makes negligible difference to our forecasts or results.
Figure1. Forecasts for MJJ (produced 25 April 2016) and JJA (produced 23 May 2016), as labeled, using GPCP observations. Observation/hindcast points are color-coded according to their observed winter ENSO index: red points are El Ni?o years, blue points are La Ni?a years, and gray points are neutral. The horizontal width of the green forecast bars is the standard error on the ensemble mean, i.e., the forecast ensemble spread divided by the number of ensemble members. The 75% and 95% prediction intervals are shown as gray shading. The variability in the observations is indicated by the pink horizontal dotted lines, at 1 and 2 standard deviations. The correlation r between hindcast and observations is marked on each panel (coincidentally the same when rounded).We therefore implemented a simple precipitation forecasting methodology, based entirely on the historical relationship between the hindcast ensemble means and the observed precipitation, averaged over the Yangtze River basin region (25°-35°N, 91°-122°E), following (Li et al., 2016), for the season in question. The prediction intervals, derived from the linear regression of the hindcasts to the observations (e.g., Wilks, 2011), provide a calibrated forecast probability distribution.
This is illustrated in Fig. 1, where we show the precipitation forecasts issued in late April for MJJ, and in late May for JJA. The distribution of hindcasts and observations is shown as a scatter plot, with the ensemble mean forecast also included as a green circle. The uncertainty in the linear regression (gray) determines the forecast probabilities (green bars). The GloSea5 data are shown in standardized units——that is, the anomaly of each year from the mean, as a fraction of the standard deviation of hindcast ensemble means. The observations on the vertical axis are presented as seasonal means of monthly precipitation totals. The relationship with ENSO is indicated though color-coding of the hindcast points: years are labelled as El Ni?o (red) or La Ni?a (blue) according to whether their Oceanic Ni?o Index (http://www.cpc.noaa.gov/products/analysis_monitoring/ensostuff/ensoyears.shtml), based on observed SST anomalies in the Ni?o3.4 region, is above 0.5 K or below -0.5 K, respectively.
Forecasts like those shown in Fig. 1 were produced each Monday starting in February 2016, using the forecast model runs initialized each day of the preceding three weeks to generate the 42-member ensemble, and the four nearest weeks of hindcast runs for the 12-member hindcast ensemble. The forecast produced near the end of each month was issued to the CMA: the MJJ release was produced on 25 April and the JJA release on 23 May.
It is important to note that, due to the linear regression method we employ, our forecast probabilities are explicitly linked to both the hindcasts and the observations. The correlations between hindcasts and observations are biased low due to the smaller size of the hindcast ensemble compared to the forecast ensemble——a larger hindcast ensemble would not necessarily alter the gradient of the linear regression, but would reduce its uncertainty. Our forecast probabilities are therefore conservative (likely to be too small).
The forecast information provided was designed to show very clearly and explicitly the uncertainties in the forecast system, to prevent overconfidence on the part of potential decision-makers. In addition to the scatterplot showing the forecast and the historical relationship (Fig. 1), we also provided the probability of above-average precipitation as a "headline message". This was accompanied by a contingency table showing the hit rate and false alarm rate for above-average forecasts over the hindcast period. For the MJJ and JJA forecasts, these are shown in Tables 1 and 2.
To assess the sensitivity of our results to individual years, we have performed leave-one-out cross-validation for the MJJ and JJA forecasts. We find that the correlation between hindcasts and observations in the case of each left-out year does not vary much: 75% of the cases have correlations between 0.41 and 0.47. However, leaving out 1998 does reduce the performance, as expected: the correlation over the remaining 22 years in that case reduces to r=0.37 (MJJ) and r=0.24 (JJA), and the observed value falls outside the 95% prediction range of the forecasts; our procedure does require similar signals to be present in the hindcast period in order to calibrate the forecasts. Note that this cross-validation procedure is not directly analogous to our actual forecasts: with only 12 members per year, the hindcast ensemble means are much more uncertain than our actual 42-member forecasts for 2016, and our cross-validation does not account for this.
Figure2. Observed precipitation from GPCP (version 2.3) for May, June, July and August (as labeled), in standardized units with respect to the 1993-2015 period. The Yangtze box used for the forecasts is marked as a red rectangle, with a pink polygon showing the physical Yangtze River catchment. Major rivers are marked in blue.
Figure3. Mean precipitation for 2016-MJJ in the GPCP observations and GloSea5 forecast signal, (as labeled), in standardized units. The GloSea5 data have been regridded to match the lower-resolution observations.
Figure4. Mean precipitation for 2016-JJA in the GPCP observations and GloSea5 forecast signal, (as labeled), in standardized units. The GloSea5 data have been regridded to match the lower-resolution observations.Figures 3 and 4 show the three-month mean precipitation anomalies for MJJ and JJA respectively, for both GPCP and the forecast averages from the GloSea5 model output. While we do not expect the spatial patterns to match in detail [considering the skill maps of (Li et al., 2016)], the overall signal is similar to the observations, with stronger anomalous precipitation in the eastern region in MJJ, and closer-to-average precipitation in JJA.
We examine our forecasts for the Yangtze basin box more quantitatively in Figs. 5 and 6, where we show the variation with lead time of the hindcast-observation correlation, the 2016 forecast signal, and the probability of above-average precipitation, for MJJ and JJA respectively. Neither the hindcast-observation correlation nor the forecast signal vary significantly with lead time; indeed, they are remarkably consistent back to three months before the forecast season, and when the 23-year hindcast is introduced at the end of April.
The forecasts did a good job of giving an indication of precipitation in the coming season. For MJJ, the forecast gave a high probability of above-average precipitation (80%), and it was observed to be above average. In JJA, the mean precipitation was observed to be slightly below average, due to the strong drier-than-average signal in August, although it was within a standard deviation of the interannual variability. While our forecast marginally favored wetter than average conditions (65% probability of above-average rainfall), it was correctly near to the long-term mean, and the observed value was well within the forecast uncertainties.
Figure5. Time series showing the behavior of the MJJ forecasts and hindcasts with lead time: (a) Correlation between observations and the operational hindcasts available each week. The final point was produced using 23 years, whereas only 14 were available before that. The shading indicates 95% confidence intervals using the Fisher Z-transformation. (b) The forecast signal shown as 95% and 75% prediction intervals (boxes) and the ensemble mean (blue line). The observed precipitation is marked as an orange horizontal line from May. The observed historical mean and standard deviation over the hindcast period are marked as a dashed line and orange shading respectively. (c) The forecast probability of above-average precipitation. The final forecast issued for MJJ, produced 25 April, is highlighted with a gray vertical bar.
Figure6. Time series showing the behavior of the JJA forecasts and hindcasts with lead time, following the same format as Fig. 5. In (a) (hindcast-observations correlation), the line becomes thicker when 23 years of hindcasts are available. We mark with a blue cross and error bar the correlation skill derived from the assessment hindcast (see text for details). The final forecast issued for JJA, on 23 May, is highlighted across all panels.Our verification has shown that our forecasts gave a good indication of the observed levels of precipitation for both MJJ and JJA averages over the large Yangtze Basin region. A greater degree of both spatial and temporal resolution——splitting the basin into upper and lower sections, and producing additional forecasts at a monthly timescale——would of course be preferable to users. However, smaller regions and shorter time periods may well be less skillful, so further work is needed to assess how best to achieve skillful forecasts in these cases.
One significant improvement would be to increase the ensemble size of the hindcast. During 2017 the GloSea5 system was changed from three hindcast members per start date to seven. This could result in noticeable improvements in forecasts like those described here, as the hindcast-observations relationship will be less uncertain, especially when a predictable signal is present, such as from El Ni?o. Improvements in the underlying climate model, such as to parametrized convective precipitation, and the simulation of the monsoon and features like the MJO, could also improve the forecast skill.
We will be issuing forecasts again in 2017. However, unlike 2016, in 2017 there are no strong drivers, such as El Ni?o. Nevertheless, understanding the behavior of the forecast system under such conditions will be informative, for both the users and the producers of the forecasts. Ultimately, trial climate services such as this help to drive forecast development, improve understanding of forecast uncertainties, and promote careful use by stakeholders in affected areas.
