Clim Dyn DOI 10.1007/s00382-017-3577-1
Seasonal drought ensemble predictions based on multiple climate models in the upper Han River Basin, China Feng Ma1,2 · Aizhong Ye1,2 · Qingyun Duan1,2
Received: 14 April 2016 / Accepted: 3 February 2017 © Springer-Verlag Berlin Heidelberg 2017
Abstract An experimental seasonal drought forecasting system is developed based on 29-year (1982–2010) seasonal meteorological hindcasts generated by the climate models from the North American Multi-Model Ensemble (NMME) project. This system made use of a bias correction and spatial downscaling method, and a distributed time-variant gain model (DTVGM) hydrologic model. DTVGM was calibrated using observed daily hydrological data and its streamflow simulations achieved Nash–Sutcliffe efficiency values of 0.727 and 0.724 during calibration (1978–1995) and validation (1996–2005) periods, respectively, at the Danjiangkou reservoir station. The experimental seasonal drought forecasting system (known as NMME-DTVGM) is used to generate seasonal drought forecasts. The forecasts were evaluated against the reference forecasts (i.e., persistence forecast and climatological
forecast). The NMME-DTVGM drought forecasts have higher detectability and accuracy and lower false alarm rate than the reference forecasts at different lead times (from 1 to 4 months) during the cold-dry season. No apparent advantage is shown in drought predictions during spring and summer seasons because of a long memory of the initial conditions in spring and a lower predictive skill for precipitation in summer. Overall, the NMME-based seasonal drought forecasting system has meaningful skill in predicting drought several months in advance, which can provide critical information for drought preparedness and response planning as well as the sustainable practice of water resource conservation over the basin.
This paper is a contribution to the special collection on the North American Multi-Model Ensemble (NMME) seasonal prediction experiment. The special collection focuses on documenting the use of the NMME system database for research ranging from predictability studies, to multi-model prediction evaluation and diagnostics, to emerging applications of climate predictability for subseasonal to seasonal predictions.This special issue is coordinated by Annarita Martiotti (NOAA), Heather Archambault (NOAA), Jin Huang (NOAA), Ben Kirtman (University of Miami) and Gabriele Villarini (University of Iowa).
1 Introduction
* Aizhong Ye
[email protected] 1
State Key Laboratory of Earth Surface Processes and Resource Ecology, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China
2
Institute of Land Surface System and Sustainable Development, Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China
Keywords Seasonal drought forecasting system · NMME · DTVGM · Soil moisture · Han River basin
Under the influence of climate changes and intense anthropogenic activities, the frequencies and severities of climatic extremes (e.g., droughts and floods) are projected to increase in the future, which can lead to severe food and water security concerns in many regions of the world. China has been experiencing increased droughts over the past few decades and is especially vulnerable to increased drought risks under climate change (Wang et al. 2011, Su et al. 2008; Zhai et al. 2010). Unlike other natural disasters (e.g., floods), droughts often develop slowly, making it difficult to detect them until they have become severe and caused a large amount of damages. If droughts can be predicted a month or more ahead of time, then the impact may be mitigated. Well-planned drought preparations cannot be made without reliable predictions of likely future
13
Vol.:(0123456789)
droughts. Therefore, accurate predictions of future drought conditions are critical for drought preparedness and water resource management. There exists an extensive body of research on seasonal hydrological forecasting and applications to drought monitoring and predictions over the last five decades. Traditional seasonal hydrological forecasting is primarily based on empirical methods such as linear regression, which relates basin initial conditions and/or large-scale climate anomalies [e.g., Pacific decadal Oscillation (PDO), El Niño-Southern Oscillation (ENSO) and Pacific/North American (PNA)] to hydrologic variables (Garen 1992; Pagano et al. 2004; Yuan et al. 2015a). Since the emergence of conceptual and distributed hydrologic models, seasonal hydrological forecasting using physically-based models has become popular. An Ensemble Streamflow Prediction (ESP) system that utilizes a hydrologic modeling system to simulate the future hydrologic conditions, was put forward by Day (1985), which could overcome some limitations of the regressionbased method. However, the ESP system relies on land surface memory, and the premise is that historical records are representative of the possible future climatic conditions. In general, the skill of the ESP system decreases significantly after 1 month (Shukla et al. 2013). More recently, there is a trend in developing climate-model-based seasonal hydrological forecasting systems, which aims to benefit from the predictive knowledge of the coupled atmosphere-oceanland general circulation models (CGCMs) and land surface hydrologic models (Yuan et al. 2015b). Those systems have also shown meaningful predictive skill over many regions, where seasonal hydrologic forecasts and extreme predictions were generated by driving hydrological models with bias-corrected meteorological forecasts from CGCMs and refined initial land surface hydrologic conditions (Yuan et al. 2013). In addition, seasonal hydrological forecasting system based on multi-CGCMs has also been getting attention, which provided more accurate evaluation of hydrologic predictability (Mo and Lettenmaier 2014; Yuan et al. 2015b). To date, a number of seasonal hydrological forecasting systems for drought predictions have been established and operated globally at many scientific research institutions and operational agencies. For instance, Princeton University developed a Drought Monitoring and Hydrologic Forecasting system to provide seasonal soil moisture and drought probability forecasts over the United States (Luo and Wood 2008; Yuan et al. 2013). The system (http:// hydrology.princeton.edu/forecast) is based on climate prediction products from the National Centers for Environmental Prediction (NCEP) Climate Forecast System version 2 (CFSv2) and VIC hydrological model. The system was modified to form an African Drought Monitor system (ADM; Sheffield et al. 2014), which provides drought
13
F. Ma et al.
predictions out to 6 months by combining the VIC hydrological modeling, satellite remote sensing products and CFS seasonal climate forecasts. The system with multiple CGCMs (e.g., North American Multi-model Ensemble (NMME) project) also provides seasonal hydrological forecasting of droughts and wet spells for major global river basins (Yuan et al. 2015a). A recently launched European research project, IMPREX, is built to improve predictions and management of hydrological extremes through climate services (van den Hurk et al. 2016; http://www.imprex. eu). In addition, more organizations and projects have been initiated to focus on drought forecasting and warning, including the Hydrologic Ensemble Prediction Experiment (HEPEX) and Hydrology and Water Resources Programme (HWRP). There are a few operational and experimental seasonal hydrological forecasting systems for drought prediction in China. For instance, Xiao et al. (2016) built a probabilistic seasonal drought forecasting model based on copula functions in the Huai River basin. Yuan et al. (2016) and Yuan (2016) developed an experimental seasonal hydrological forecasting system over the Yellow River basin by using NMME climate forecast models and the VIC land surface hydrological model. However, most seasonal drought forecasting systems are built from large-scale land surface hydrological models, which need to be improved given the small-scale catchment features and changes in land use and land covers in basins with intense human activities (Wood et al. 2011). Further, regardless of the great progress in seasonal drought forecasting, a great challenge still remains in terms of how to effectively disseminate the information from the system to the decision makers, which requires cross-disciplinary dialogue and collaboration (Yuan et al. 2015b). The Han River is the largest tributary of the Yangtze River of China, flows through the China’s heart, linking the Yangtze and Yellow River basins. Because of its clear water quality and convenient location, the upper Han River is the water source of the Middle Route of China’s SouthNorth Water Transfer Project (Zhang 2005), which supplies water from the Danjiangkou reservoir into North China, including Beijing and Tianjin, to relieve water shortages in the region (Li et al. 2009). During the last half century, precipitation has obviously dropped, and drought frequencies showed an increasing trend in the Han River (Meng and Yin 2012). Under the influence of climate change, precipitation, terrain factors and intensive anthropogenic activities, drought disasters have occurred more frequently in the upper Han River basin (Ren et al. 2013). Therefore, in this paper, we present and evaluate an experimental seasonal drought forecasting system for soil moisture drought prediction over the upper Han River Basin in China using climate prediction products from
Seasonal drought ensemble predictions based on multiple climate models in the upper Han River…
multiple CGCMs participating in the NMME project (Kirtman et al. 2014). The next section briefly describes the data and methodology for the seasonal hydrological forecasting system, and the assessment of the predictions for soil moisture hindcasts and realistic drought events is presented in Sect. 3. The conclusions and discussion are then summarized in Sect. 4.
2 Data and methodology 2.1 Study area The Han River has a total length of 1530 kilometers and its drainage basin covers a total area of approximately 159 × 103 km2 (Li and Zhang 2012; Shen and Liu 1998; Yang et al. 1997). Its upper river basin (105°29′–112°E, 31°20′–34°10′N; approximately 95,200 km2 in area) extends from the head of the basin to Danjiangkou reservoir and lies in the north sub-tropic monsoon climatic zone (Li et al. 2008). The average precipitation is ranges between 700 and 1800 mm/year, of which approximately 80% falls during the rainy season of May to October period (Yang et al. 1997). The average runoff is 41.1 × 109 m3/year and has large inter-annual and intra-annual variability (Jin and Guo 1993; Yang et al. 1997). Based on a 500 m DEM, the upper Han River basin has been divided into 380 subbasins (418 km2 on average) using a Geographic Information System (GIS). The hydrologic process was performed on each sub-basin, which is the minimum hydrologic unit of the hydrological model.
2.2 Observed data and NMME hindcasts Station-based daily meteorological observations were bilinearly interpolated to 0.5-degree resolution, and temperature was first adjusted to the corresponding elevation based on a constant lapse rate (6 °C/km). Figure 1 exhibits the geographical locations of the meteorological stations and Danjiangkou reservoir. The latitude and longitude of the Danjiangkou reservoir are 32.56°N and 111.49°E, respectively. Hindcasts and forecasts from the NMME project (Kirtman et al. 2014), in support of intraseasonal to seasonal to interannual predictions, have been widely used for hydroclimate forecasting (Becker et al. 2014; Delsole et al. 2014; Mo and Lettenmaier 2014; Tian et al. 2014; Yuan and Wood 2012, 2013). The predictive skills of precipitation and meteorological drought based on the NMME hindcasts have been investigated in China, where most of the models perform well (Ma et al. 2015, 2016). In this study, climate hindcasts (including monthly precipitation and temperature) with total 71 ensemble members during the period 1982–2010 from six models (Ma et al. 2015) were used as inputs to drive the hydrological model. For any of these models, the lead-0 month refers to forecasts with a 0.5-month lead time, lead-1 month refers to 1.5month lead time, and so on. The hindcasts are interpolated into 68 grids at a 0.5 degree resolution, and then bias-corrected and downscaled to 380 sub-basins using a bias correction and spatial downscaling approach (Wood et al. 2002) and the inverse quadratic distance weighing method.
Fig. 1 Locations of the meteorological and hydrological stations in the upper Han River basin, China. Blue represents the lowest elevation whereas red represents the highest elevation
13
2.3 Experiment design A schematic diagram of the seasonal forecasting system for drought prediction in the upper Han River Basin is provided in Fig. 2. The central element is the distributed time-variant gain model (DTVGM) hydrological model that converts meteorological information into land surface states such as soil moisture conditions. For each month, the monthly soil moisture is converted into percentiles to produce soil moisture index, and the 20th percentile is used as the threshold to determine if a region is in drought. The validation datasets for soil moisture are soil moisture from the DTVGM offline simulation forced by observed precipitation (Yuan et al. 2015a). 2.4 Hydrologic models The DTVGM, which takes advantage of both nonlinear and distributed hydrological models, is proposed to simulate diverse hydrologic processes under changing climatic conditions (Xia et al. 2005; Ye et al. 2014). The DTVGM model (Xia et al. 2005) includes the following major parts. (1) Preprocessor for land surface parameters and hydrological condition, including the spatial distribution of the drainage characteristics and hydrometeorological forcing variables. In this process, the potential evapotranspiration can be simulated by Hargreaves method (Arnold et al. 1997; Pereira et al. 1999), and then the actual evapotranspiration is obtained based on the energy balance. (2) Runoff
Fig. 2 Schematic diagram for seasonal drought predictions in the upper Han River Basin, China
13
F. Ma et al.
generation simulation. There are three runoff components on each hydrologic unit (sub-basin in this paper), including surface runoff on land surfaces, sub-surface runoff from surface soil layers, and base flows from deep soil layers. The DTVGM model is designed based on water balance and calculated as: (1) where P is precipitation, W is soil moisture, E is evaporation, Rs is surface runoff, Rss is sub-surface runoff, Rg is base flow, and i is a period of time. (3) Flow routing simulation. The routing model used in this process is the kinematic wave model. The flow route can be treated as the river stream network, which is ranked from the outlet of the sub-basin to upstream according to the flow direction (Ye et al. 2005), and the flow routing is undertaken inversely. A detailed description of the DTVGM model can be found in Xia et al. (2003). The DTVGM model has performed well in simulating hydrological processes and the complex relations between land surface change and runoff variation over different basins (Xia et al. 2005; Ye et al. 2010). The calibration of the DTVGM model is described in Sect. 2.5.
Pi + Wi = Wi+1 + Rsi + Ei + Rssi + Rgi
2.5 Calibration of the DTVGM model The DTVGM model was run from 1967 to 2010, and the simulations in the first twelve years were dropped for spinup. Eighteen years (1978–1995) of daily hydrological data for the Danjiangkou station were selected to calibrate the
Seasonal drought ensemble predictions based on multiple climate models in the upper Han River…
model and data for the last ten years (1996–2005) were used for validation. The five major parameters of the DTVGM runoff model selected for calibration were the runoff coefficient when soil is saturated (g1, 0 < g1 < 1), the soil water parameter (g2, g2 > 0), the infiltration rate (fc, 0–30 mm/h), the sub-surface runoff coefficient (Kr, 0 < Kr < 1), and the groundwater runoff coefficient (Kg, 0 < Kg < 1). The optimized DTVGM runoff parameters were then used to generate simulated runoff to calibrate the DTVGM routing model, and the Manning roughness coefficient (n, 0.001 < n < 0.15; Huggins and Monke 1966) was selected for calibration. Figures 3 and 4 show the comparison results between the simulated and observed streamflows during the calibration
and validation periods, respectively. A Nash–Sutcliffe efficiency (NSE) value of one corresponds to a perfect match between the simulated and reference streamflows, and a value of less than zero occurs when the climatology is better than the model simulated streamflow. The averaged NSE during the calibration and validation periods were 0.727 and 0.724, respectively. The values of correlation coefficient (R) during the calibration and validation periods were 0.89 and 0.86. The simulated and observed total water volumes were similar, with a balance coefficient of 0.905 during the calibration period and 1.039 during the validation period, both values being close to one. Furthermore, the simulated streamflows matched the observations reasonably well (Figs. 3, 4) during the calibration
Fig. 3 Observed (red) and DTVGM simulated (black) daily streamflow datasets during the calibration period (1978–1995) at Danjiangkou reservoir on the upper Han River
Fig. 4 Observed (red) and DTVGM simulated (black) daily streamflow datasets during the validation period (1996–2005) at Danjiangkou reservoir on the upper Han River
13
and validation periods. The aforementioned fitting statistics demonstrates that the DTVGM model is reasonably accurate for long-term simulations of hydrological conditions in the upper Han River Basin. 2.6 Bias Correction of the NMME climate hindcasts The Bias Correction approach (Wood et al. 2002), which originally was developed to process monthly datasets, was adopted in this study to correct the biases in the NMME climate model precipitation forecasts. In this study, the bias correction was based on two datasets: observed meteorological datasets (1957–2011) and hindcasts from a given NMME model member (1982–2010). First, both the observed data and hindcasts were interpolated to a resolution of 0.5 × 0.5. For each grid cell, given month, and variable (precipitation, average temperature, maximum temperature, and minimum temperature), Cumulative Distribution Function (CDF) curves for both observed datasets and model hindcasts were then constructed in a leave-one-out (i.e., the target year) cross validation mode. For low precipitation, an Extreme Value Type III (Weibull) distribution where a lower bound is known to be zero was employed. For high precipitation, an Extreme Value Type I (Gumbel) function was applied. For temperature, a normal distribution was employed. The paired CDFs were related to form a ‘quantile map’ through a probability threshold, which can be used to assess the bias between the observed data and model hindcasts. The adjustment for bias removal involved the following steps: (1) identifying the hindcast value at any time step and finding the related rank probability from the ‘quantile map’; (2) identifying the corresponding observed value having the same rank probability in the ‘quantile map’; (3) the bias correction is then carried out by replacing the hindcasts value with the observed value. Finally, the bias-corrected hindcasts are expressed as anomalies (shift for temperature and percentage for precipitation) regarding the monthly averaged observed values for the 55-yr climatology period. With the downscaled and bias-corrected monthly variables, daily time series were generated by randomly sampling the observed daily series and rescaling to match the monthly hindcasts (Luo and Wood 2008) for each member. Both the daily meteorological hindcasts and observations were then downscaled to 380 sub-basins in the upper Han River Basin to force the DTVGM model to produce soil moisture hindcasts and references, using the inverse quadratic distance weighing method. In this paper, soil moisture indexes for each sub-basin are defined as percentile values of the current soil moisture versus the 29-year soil moisture climatology at each calendar month (Sheffield et al. 2004), which were obtained from running the DTVGM model using observed atmospheric
13
F. Ma et al.
forcings during the period 1982–2010. The soil moisture percentile-based drought index can be used in drought forecasting because of its quantitative nature as a spatially continuous field (Luo and Wood 2007). Hit rate, False alarm rate, bias score and brier score (Wilks 2011) were used to assess the forecast skill of soil moisture drought from different aspects. The soil moisture index maps from the references and NMME models’ ensemble means (EMs) were then compared and drawn to show the skills in predicting specific droughts in the upper Han River Basin.
3 Results and discussion Figure 5 shows the forecast skill of seasonal precipitation hindcasts from the NMME ensemble mean, which were the input data for soil moisture forecasting, at the first lead over the upper Han River basin. Herein, spring is from March to May, summer is from June to August, autumn is from September to November, and winter is December to February in the next year. For each season, the hindcast period is 1982–2010 (29-year), so a correlation of 0.21 (0.18) is statistically significant at the 5% (10%) level. The precipitation forecasts at lead-0 shows the highest skill during the spring and autumn, with an average correlation coefficient of 0.71, and the forecast skill declines as the forecast lead time increases. The high skills appear in the southern part during spring, and in western part during autumn. For the prediction forecasts during the summer, the grand NMME has the lowest but moderate skill, with an averaged correlation higher than 0.24 (p < 0.05). However, the eastern part of the upper Han River basin shows almost no skill during summer. Until the lead-3 month, the precipitation hindcasts also have high skill, with an averaged correlation higher than 0.6, during the spring and autumn (not shown). The differences in Pearson’s correlation for the seasonal precipitation at lead-0 between NMME ensemble mean forecasts and reference forecasts (i.e. climatological mean forecasts during the period of 1967–1981) are shown in Fig. 6. NMME shows generally higher predictive skill than climatological forecasts except for the western and eastern parts of the upper Han River basin during summer, indicating some challenges in predicting precipitation during summer over those regions. Most of the significant improvements appear in southeastern part of the upper Han River basin during autumn, and much of the upper Han River basin during winter. The soil moisture hindcasts begin at the first of each calendar month of every year, over the 29-year period 1982–2010 and run for 4 months, with 71 ensemble members for NMME-DTVDM. The predictions were very skillful at capturing the drought areas when a wide range of regional droughts occurred over the upper Han River basin,
Seasonal drought ensemble predictions based on multiple climate models in the upper Han River…
Spring
Summer 0.85 0.8 0.75 0.7 0.65
Autumn
Winter
0.6 0.5 0.4 0.21 0.18
Fig. 5 Pearson’s correlation of seasonal precipitation forecasts from the NMME ensemble averaged among 71 members at the first lead over the upper Han River basin during the period of 1982–2010
(a) Spring
(b) Summer
0.4
0.35
0.3
0.25
0.2
(c) Autumn
(d) Winter
0.15
0.1
0.05
0
Fig. 6 Differences in Pearson’s correlation of seasonal precipitation between the NMME ensemble averaged among 71 members at the first lead and climatological forecasts over the upper Han River basin during the period of 1982–2010
13
F. Ma et al.
but over-predicted them when the droughts were not severe (not shown). The drought areas are defined as the total area of subbasin where the monthly average soil moisture value is less than the threshold at the 20th percentile. Over the upper Han River basin, there are decreasing conformities with lead times between the predicted soil moisture drought and estimated soil moisture drought areas, with correlation coefficient values of 0.72, 0.54, 0.45, and 0.33 from 0 to 3 lead months, respectively. The NMME-DTVGM forecasts are then compared with reference forecasts based on historical simulations of 1967–1981. Figure 7 exhibits the differences in hit rate for soil moisture drought between NMME ensemble mean forecasts and reference forecast. The reference forecast is persistent forecast, where the initial soil moisture (mean states during the period of 1967–1981 in this study) is used throughout the forecast horizon. The hit rate was calculated as the ratio of correct drought predictions to the number of times drought occurred (that is, the detectability; Wilks 2011). For example, a hit rate value of 0.5 indicates that 50% of the soil moisture droughts in the hindcast period can be predicted by the seasonal drought forecasting system. The hit rate was more than 0.6 at shorter lead months, which indicates that more than 60% of the soil moisture droughts can be predicted by NMME-DTVGM at least 1 month ahead (not shown). Here, positive difference in hit rate indicates that NMME-DTVGM forecast is better than persistent forecast. NMME-DTVGM is generally more skillful than persistent forecast over most regions, except for eastern part of the upper Han River basin during summer. Spring
Summer
Lead−0
This is consistent with the phenomenon that almost no predictive skill and little improvement in NMME precipitation forecast also happen in summer (Figs. 5b, 6b). The biggest improvement occur in the winter, which may attribute to the biggest improvement of precipitation in winter (Fig. 6d). During cold season, more than 40% of soil moisture droughts can even be predicted 4 months ahead by NMME-DTVGM, in spite of low precipitation skill, due to the long memory from the initial conditions (such as soil moisture and snow, not shown). Other study (Shukla and Lettenmaier 2011) also provided similar results, and demonstrated that the predictability of soil moisture had strong seasonality. The improvement of NMME-DTVGM is still significant until lead-3, however, the difference decreases with lead time, which may be because of the influence of initial conditions and the decreasing predictability of climate forecasts with leads. Figure 8 shows the differences in False alarm rate (FAR) for soil moisture drought as in Fig. 7. The FAR was the fraction of drought predictions that fail to occur (Wilks 2011). The FAR was less than 0.5 at shorter lead time, and increased with leads (not shown). The NMME-DTVGM shows higher hit rate and lower FAR at the first two leads, indicating more accurate drought predictions (not shown). Here, negative difference in false alarm rate indicates that NMME-DTVGM forecast is better than persistent forecast. Figure 8 shows that negative differences still last until lead-3 month, indicating lower false alarm rate of NMMEDTVGM than persistent forecast. Unlike hit rate, the differences in false alarm rate between NMME-DTVGM and Autumn
Winter
1
0.9
0.8
0.7
Lead−1
0.6
0.5
0.4
Lead−2
0.3
0.2
0.1
Lead−3
Fig. 7 Differences in hit rate for soil moisture drought between the NMME and reference forecast, at different seasons and forecast leads during the hindcast period (1982–2010). The NMME were the
13
0
ensemble mean soil moisture from all of the models in the study, the reference forecast is persistent forecast
Seasonal drought ensemble predictions based on multiple climate models in the upper Han River… Spring
Summer
Autumn
Winter 0
Lead−0
−0.1
−0.2
−0.3
Lead−1
−0.4
−0.5
Lead−2
−0.6
−0.7
−0.8
Lead−3
−0.9
−1
Fig. 8 Differences in false alarm rate for soil moisture drought as in Fig. 7
persistent forecast are small after lead-1, indicating fastincreasing false alarm rate in NMME-DTVGM. In addition, the bias score for soil moisture drought of NMMEDTVGM is shown in Fig. 9. The bias score represents the ratio of predicted drought events to the actual drought events, with unbiased forecasts of one. The NMMEDTVGM usually overforecasts the soil moisture drought over the upper Han River basin, with the value of bias score greater than one. The phenomenon of overforecasting is
Spring
Summer
more severe during spring and winter. During the winter, there is underforecasting at the lead-3 month. The persistent forecast also overforecasts or underforecasts the soil moisture drought irregularly (not shown). Hit rate, false alarm rate, and bias score are just one aspect of forecast skill, and the Brier Score (BS; Wilks 2011) is widely applied to assess the probabilistic drought forecasting performance (Ma et al. 2015). A high accuracy rate receives a smaller BS, and a perfect forecast
Autumn
Winter
6
5
Lead−0
4
3.5
3
Lead−1
2.5
2
Lead−2
1.5
1
0.5
Lead−3
0
Fig. 9 Bias score for soil moisture drought of the NMME at different seasons and forecast leads during the hindcast period (1982–2010)
13
F. Ma et al.
has BS = 0. The BS is often converted to Brier skill score (BSS), normalizing the score by that of a reference forecast, for example, climatological forecast. A BSS of 1 indicates a perfect forecast, a value of 0 indicates skill of reference forecast, while the negative value should indicate lower skill than reference forecast. Figure 10 shows the brier skill score (BSS) of soil moisture drought predicted by NMMEDTVGM, using climatological forecast as reference forecast, which consists of the samples from historical simulation during 1957–1981. The biggest improvement occur in autumn, while there is little improvement in spring despite higher predictive skill for precipitation (Fig. 5c). The effects of initial condition on soil moisture during spring are expected to larger than other seasons. This is another example of predicting uncertainty associated with both the climate forecasts and the initial conditions. The 1997 drought over the upper Han River basin was one of the most severe seasonal droughts in the past 50 years (Tao et al. 2015; Xu 1998). Other than the drought monitor (http://www.cma.gov.cn/tqyb/v2/product/), however, few studies predicted the 1997 drought over the upper Han River basin. In this study, to further evaluate the skill of the seasonal drought forecasting system, the capability of NMME-DTVGM for predicting the 1997 soil moisture drought was tested. Comparison of the predicted soil moisture index from the ensemble mean of the NMME-DTVGM forecasts and estimated soil moisture index for August-October 1997 at different lead times is shown in Fig. 11. Over the upper Han River basin, the soil moisture drought extended to the entire river basin in September. Most droughts are associated with anomalies in Spring
Lead−0
Summer
precipitation. The precipitation anomalies for the antecedent 3 months of August–October 1997 from observations over the upper Han River basin are shown in Fig. 12a–c. During the 3 months, the negative precipitation anomalies resulted in soil moistures below the 20th percentile (Fig. 11a–c). By comparing with the referenced soil moisture index, the predictions 1–2 months ahead over the south were satisfactory for predicting the soil moisture drought over the region during the 3 months (August–October, 1997). Meanwhile, the precipitation of antecedent 3 months from the NMME mean hindcasts also shows negative anomalies over the southern part (Fig. 12a0–a2, b0–b2, c0–c2). The predicted soil moisture conditions 1 month ahead matched well with the estimated soil moisture in terms of the area of the drought in September. However, the NMME-DTVGM under-predicted the area and severity of drought in August and predictions at lead-2 month during the 3 months. The under prediction may be attributed to the positive precipitation anomalies in the NMME mean hindcasts (Fig. 12a0–a2, b0–b2, c0–c2). However, the inconformity between the dry soil moisture index (Fig. 11a0–a1, b0–b1) and precipitation positive anomalies (Fig. 12a0–a1, b0–b1), such as northern part, may attribute to the memory of the initial soil moisture conditions. In fact, predictions from the same initial conditions (ICs) produced similar soil moisture conditions (Fig. 11a0, b2, c3), which also indicate that ICs are a major source of hydrologic predictability (Yuan et al. 2015b). In addition, our system also can indicate the region where soil moisture is wetter than normal, such as the northeastern part of the upper Han River basin in August (Fig. 11a), but it over-predicted the severity. The Autumn
Winter
1
0.9
0.8
0.7
Lead−1
0.6
0.5
0.4
Lead−2
0.3
0.2
0.1
Lead−3
0
Fig. 10 Brier skill scores (BSS) for seasonal soil moisture drought predicted by NMME-DTVGM during the hindcast period (1982–2010). The reference forecast is climatological forecast
13
Seasonal drought ensemble predictions based on multiple climate models in the upper Han River… Estimated SM (a)
1 Month Lead Predicted SM (a0)
2 Month Lead Predicted SM (a1)
3 Month Lead Predicted SM 100
(a2)
90
AUG 1997
80
70
(b)
(b0)
(b1)
(b2)
60
50
SEP 1997
40
30
(c)
(c0)
(c1)
(c2)
20
10
OCT 1997
0
Fig. 11 Predicted soil moisture index for the 199708, 199709, and 199710 forecasts at different forecast leads (1–3 months), compared with the simulated soil moisture index obtained using the observed meteorological forcing
Lead−0
JJA 1997 P Anomaly
OBS (a)
(a0)
Lead−2
Lead−1 (a1)
(a2)
mm/day 6
5
4
3
(b)
(b0)
(b1)
(b2)
2
JAS 1997
1
0
−1
−2
(c)
(c0)
(c1)
(c2)
ASO 1997
−3
−4
−5
−6
Fig. 12 Average precipitation anomalies for JJA (June–August), JAS (July–September), and ASO (August–October) 1997 from observations (a–c) and the NMME mean hindcasts at different leads
precipitation anomalies in the northeastern part also are positive during June–August 1997 (Fig. 12a). The precipitation anomaly patterns corresponded well with the soil moisture conditions, which indicates the considerable contributions of the climate forecasts to hydrological predictability. In addition, the climatic factors that led to drought during 1997 over the Han River basin have been presented in Xu (1998).
4 Conclusions In this study, a seasonal hydrological forecasting system was established using high resolution and fine calibration procedures. Its performance in predicting soil moisture droughts was assessed over the upper Han River basin. Based on 18 years (1978–1995) of observed streamflow data at the Danjiangkou reservoir and an observed forcing
13
dataset interpolated from meteorological stations. The DTVGM land surface hydrological model was calibrated over 380 sub-basins in the upper Han River basin. During the calibration and validation (1996–2005) periods, the averaged Nash–Sutcliffe efficiency (NSE) were 0.727 and 0.724, respectively. For the predictions, the 29-year (1982–2010) NMME climate hindcasts were bias corrected and downscaled to the river basin scale to drive the DTVGM land surface hydrologic model, and the evaluation was based on the 29-year soil moisture hindcasts and a realistic drought case. It was found that the model-based seasonal drought forecasting system provided a reliable prediction of soil moisture conditions, especially at shorter lead times. When soil moisture droughts widely occurred over the upper Han River basin, the system was able to capture the area of soil moisture below the 20th percentile threshold well. However, the system always over-predicted the drought areas when the soil moisture droughts were not severe. In terms of the hit rate (HR), False alarm rate, bias score and brier score (BS), the system exhibited good predictive skill against reference forecast for predicting soil moisture droughts in the upper Han River basin, especially during the cold–dry season at shorter lead times. During the cold-dry season, the system was even able to predict droughts 4 months in advance. Because of great influence of initial conditions during spring and lower predictive skill for precipitation during summer, little improvements against reference forecast are found during these seasons. The NMME-DTVGM generally overforecasts the soil moisture drought events. A forecasting of the AugustOctober 1997 widespread drought shows that NMMEDTVGM was able to predict the spatial pattern and severity of the soil moisture drought quite well at the first two lead months. The major cause of drought in August–October 1997 may have been the lack of precipitation before the drought occurred, which was influenced by circulation anomalies. In addition, wetter-than-normal soil moisture conditions can also be captured by NMME-DTVGM. Both initial land surface conditions (ICs) and climate forecast skill influence hydrologic predictability. Overall, the seasonal hydrological forecasting system with NMME-DTVGM ensemble mean established in this study demonstrated encouraging performance for predicting soil moisture drought hindcasts and well predictions for the August-October 1997 drought. Other than previous studies (e.g. Yuan et al. 2016; Yuan 2016), our study was focused on the drought prediction ability in humid basins, and also tried to use a different hydrological model, which is proposed to simulate diverse hydrologic processes based on the sub-basins, to predict droughts. A website will be established to provide drought monitoring and real-time drought forecasts (e.g., soil moisture and streamflow)
13
F. Ma et al.
several months in advance, using the seasonal hydrological forecasting system based on NMME real-time climate forecasts (including the six models in this study and other added models in NMME, such as COLA-RSMAS-CCSM4, GFDL-CM2p5-FLOR-A06, and GFDL-CM2p5-FLORB01; http://iridl.ldeo.columbia.edu/SOURCES/.Models/. NMME/) or/and extended NMME-II datasets with higher temporal resolutions. Combining the seasonal drought forecasting system with impact models and application sectors for drought mitigation planning will maximize the effectiveness of the system and minimize losses. This system can also be applied to flood forecasting in the future to permit appropriate and timely flood precautions. Although this study established a seasonal hydrological forecasting system for drought forecasts over the upper Han River basin, many issues must be addressed in future research: (1) more realistic drought cases are required to demonstrate the stability of the results; (2) the estimated soil moisture is not completely representative of reality, and the observed soil moisture from in situ measurements, data assimilation and remote sensing can provide more reliable assessments of the system; (3) multiple hydrological models are necessary to reduce the uncertainties in drought predictions, and a hydrological post-processor is essential for a hydrological forecasting system, especially over river basins that are seriously influenced by human activities; and (4) for the headwater of the middle line of the Southto-North Water Diversion Project in the upper Han River basin, more effort should be given to improve the understanding of the influences of water resources management and human intervention on future hydrologic conditions and how to incorporate them into the seasonal hydrological forecasting system. The real-time monitoring and prediction of the drought seasonal hydrological forecasting system will provide critical information for drought preparation several months in advance and for water resources management and sustainable practices for the South-toNorth Water Diversion Project in the basin. Acknowledgements This study was supported by the Natural Science Foundation of China (No. 41475093), the Intergovernmental Key International S&T Innovation Cooperation Program (No. 2016YFE0102400) and the State Key Laboratory of Severe Weather Open Research Program (No. 2015LASW-A05).
References Ahmed KF (2011) Bias Correction and Downscaling of climate model outputs required for impact assessments of climate change in the US northeast. Master’s Theses pp212. http://digitalcommons.uconn.edu/gs_theses/212 Arnold JG, Williams JR, Srinivasan R (1997) Model theory of SWAT. USDA. Agricultural Research Service Grassland. Soil and Water Research Laboratory, USA
Seasonal drought ensemble predictions based on multiple climate models in the upper Han River… Becker E, van den Dool H, Zhang Q (2014) Predictability and forecast skill in NMME. J Climate 27:5891–5906. doi:10.1175/ JCLI-D-13-00597.1 Caffrey P, Farmer A (2014) A review of Downscaling methods for climate change projections. Tetra Tech ARD Day GN (1985) Extended streamflow forecasting using NWSRFS. J Water Resour Plann Manage Div Am Soc Civ Eng 111:157–170. doi:10.1061/(ASCE)0733-9496(1985)111:2(157) DelSole T, Nattala J, Tippett MK (2014) Skill improvement from increased ensemble size and model diversity. Geophys Res Lett 41:7331–7342. doi:10.1002/2014GL060133 Garen DC (1992) Improved techniques in regression-based streamflow volume forecasting. J Water Resour Plann Manage 118:654–670. doi:10.1061/(ASCE)0733-9496(1992) Huggins LF, Monke EJ (1966) The mathematical simulation of the hydrology of small watersheds. Technical Report No. 1. Purdue University Water Resource Research Center, West Lafayette Jin R, Guo H (1993) Water resources assessment in the water source areas of the Middle Route of the South to North Water Transfer Project and water quantity analysis in the Danjiangkou Reservoir. Yangzte River 24:7–12 Kirtman BP, Min D, Infanti JM, Kinter JL, Paulino DA, Zhang Q, van den Dool H, Saha S, Pena Mendez M, Becker E, Peng P, Tripp P, Huang J, DeWitt DG, Tippett MK, Barnston AG, Li S, Rosati A, Schubert SD, Rienecker M, Suarez M, Li ZE, Marshak J, Lim Y.-K, Tribbia J, Pegion K, Merryfield WJ, Denis B, Wood EF (2014) The North American Multi-Model Ensemble (NMME): Phase-1 seasonal to interannual prediction, phase-2 toward developing intra-seasonal prediction. Bull Am Meteorol Soc 95:585–601. doi:10.1175/BAMS-D-12-00050.1 Li S, Zhang Q (2012) Basin ecosystem management in the Upper Han River for the South to North Water Division Project, China. J Environ Anal Toxicol S3:002. doi:10.4172/2161-0525.S3-002 Li S, Gu S, Liu W, Han H, Zhang Q (2008) Water quality in relation to land use and land cover in the upper Han River Basin, China. Catena 75:216–222 Li S, Liu W, Gu S, Cheng X, Xu Z, Zhang Q (2009) Spatio-temporal dynamics of nutrients in the upper Han River basin, China. J Hazard Mat 162(2–3):1340–1346. doi:10.1016/j. jhazmat.2008.06.059 Luo L, Wood EF (2007) Monitoring and predicting the 2007 US drought. J Geophys Res 34:L22702. doi:10.1029/2007GL031673 Luo L, Wood EF (2008) Use of Bayesian merging techniques in a multimodel seasonal hydrologic ensemble prediction system for the eastern United States. J Hydrometeor 9:866–884. doi:10.117 5/2008JHM980.1 Luo L, Wood EF, Pan M (2007) Bayesian merging of multiple climate model forecasts for seasonal hydrological predictions. J Geophys Res 112:D10102. doi:10.1029/2006JD007655 Ma F, Yuan X, Ye A (2015) Seasonal drought predictability and forecast skill over China. J Geophys Res 120(16):8264–8275. doi:10 .1002/2015JD023185 Ma F, Ye A, Deng X, Zhou Z, Liu X, Duan Q, Xu J, Miao C, Di Z, Gong W (2016) Evaluating the skill of NMME seasonal precipitation ensemble predictions for 17 hydroclimatic regions in continental China. Int J Climatol 26:132–144. doi:10.1002/joc.4333 Meng C, Yin SY (2012) Study on precipitation change and drought prediction in upper reaches of Hanjiang River during the last 50 years. Res Agri Modern 33(1):125–128 Mo K, Lettenmaier DP (2014) Hydrologic prediction over the conterminous United States using the National Multi-Model Ensemble. J Hydrometeor 15:1457–1472. doi:10.1175/JHM-D-13-0197.1 Pagano T, Garen D, Sorooshian S (2004) Evaluation of official western U.S. seasonal water supply outlooks, 1922–2002. J Hydrometeor 5:896–909. doi:10.1175/1525-7541(2004)005<0896:EOO WUS>2.0.CO;2
Pereira L, Pereira A, Allen R, Alves I (1999) Evapotranspiration: concepts and future trend. J Irrig Drain Eng 4:45–51. doi:10.1061/(ASCE)0733-9437(1999) Ren L, Yin S, Peng W (2013) Statistics and causes of historical drought disasters in upper reaches of Hanjiang River. Bull Soil Water Conserv 33(1):129–371 Sheffield J, Goteti G, Wen F, Wood EF (2004) A simulated soil moisture based drought analysis for the United States. J Geophys Res 109:D24108. doi:10.1029/2004JD005182 Sheffield J, Wood EF, Chaney N, Guan K, Sadri S, Yuan X, Olang L, Amani A, Ali A, Demuth S, Ogallo L (2014) A drought monitoring and forecasting system for sub-Sahara African water resources and food security. Bull Amer Meteor Soc 95:861–882. doi:10.1175/BAMS-D-12-00124.1 Shen D, Liu C (1998) Effects of different scales of MR-SNWTP on the down stream of the Danjiang Kou reservoir. Acta Geographica Sinica 53:341–348 Shukla S, Lettenmaier DP (2011) Seasonal hydrologic prediction in the United States: Understanding the role of initial hydrologic conditions and seasonal climate forecast skill. Hydrol Earth Syst Sci 15:3529–3538. doi:10.5194/hess-15-3529-2011 Shukla S, Sheffield J, Wood EF, Lettenmaier DP (2013) On the sources of global land surface hydrologic predictability. Hydrol Earth Syst Sci 17:2781–2796. doi:10.5194/hess-17-2781-2013 Su B, Kundzewicz ZW, Jiang T (2008) Simulation of extreme precipitation over the Yangtze River Basin using Wakeby distribution. Theor Appl Climatol 96:209–219. doi:10.1007/ s00704-008-0025-5 Tao X, Chen H, Xu C (2015) Characteristics of drought variations in Hanjiang Basin in 1961–2014 based on SPI/SPEI. Journal of Water Resources Research 4(5):404–415. doi:10.12677/ JWRR.2015.45050 Van den Hurk B, Bouwer LM, Buontempo C, Döscher R, Ercin E, Hananel C, Hunink JE, Kjellström E, Klein B, Manez M, Pappenberger F, Pouget L, Ramos M-H, Ward PJ, Weerts AH, Wijngaard JB (2016) Improving predictions and management of hydrological extremes through climate services. Climate Services 1: 6–11. doi:10.1016/j.cliser.2016.01.001 Wang A, Lettenmaier DP, Sheffield J (2011) Soil moisture drought in China, 1950–2006. J Climate 24:3257–3271. doi:10.1175/2 011JCLI3733.1 Wilks DS (2011) Statistical methods in the atmospheric sciences. 3rd edn. Academic Press, Salt Lake City Wood AW, Maurer EP, Kumar A, Lettenmaier DP (2002) Long-range experimental hydrologic forecasting for the eastern United States. J Geophys Res 107(D20):6–15. doi:10.1029/2001JD000659(ACL 6–1–ACL) Wood AW, Lettenmaier DP (2006) A test bed for new seasonal hydrologic forecasting approaches in the western United States. Bull Am Meteor Soc 87:1699–1712. doi:10.1175/ BAMS-87-12-1699 Wood AW, Leung LR, Sridhar V, Lettenmaier DP (2004) Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs. Clim Change 62(1– 3):189–216. doi:10.1023/B:CLIM.0000013685.99609.9e Wood EF, et al. (2011), Hyperresolution global land surface modeling: Meeting a grand challenge for monitoring Earth’s terrestrial water. Water Resour Res 47:W05301, doi:10.1029/20 10WR010090 Xia J, Wang G, Lv A (2003) A research on distributed time variant gain modeling. Acta Geographica Sinica 58(5):789–796. doi:10.11821/xb200305019 Xia J, Wang G, Tan G, Huang GH (2005) Development of distributed time-variant gain model for nonlinear hydrological systems. Sci China Series D: Earth Sci 48(6):713–723. doi:10.1360/03yd0183
13
Xiao M, Zhang Q, Singh VP, Chen X (2016) Probabilistic forecasting of seasonal drought behaviors in the Huai River basin, China. Theore Appl Climatol. doi:10.1007/s00704-016-1733-x Xu Y (1998) An analysis of climatic cause for dry and rainless Han River basin during 1997. Hubei Meteorological (2):12–14 (in Chinese) Yang Y, Zhou N, Guo X, Hu Q (1997) The hydrology characteristics analysis of HanJiang up-streams. Hydrology 2:54–56 Ye A, Xia J, Wang GS, Wang XN (2005) Drainage network extraction and subcatchment delineation based on digital elevation model. J Hydraul Eng 36(5):531–537 (Chinese) Ye A, Duan Q, Zeng H, Li L, Wang C (2010) A distributed time-variant gain hydrological model based on remote sensing. J Resour Ecol 1(3):222–230. doi:10.3969/j.issn.1674-764x.2010.03.005 Ye A, Duan Q, Xu J (2014) A review of hydrological ensemble forecast and case study. China Water Power Press, Beijing (Chinese) Yuan X (2016) An experimental seasonal hydrological forecasting system over the Yellow River basin-Part II: The added value from climate forecast models. Hydrol Earth Syst Sci 20:2453– 2466. doi:10.5194/hess-20-2453-2016 Yuan X, Wood EF (2012) On the clustering of climate models in ensemble seasonal forecasting. Geophys Res Lett 39:L18701. doi :10.1029/2012GL052735
13
F. Ma et al. Yuan X, Wood EF, Roundy JK, Pan M (2013) CFSv2-based seasonal hydroclimatic forecasts over conterminous United States. J Climate 26:4828–4847. doi:10.1175/JCLI-D-12-00683.1 Yuan X, Roundy JK, Wood EF, Sheffield J (2015a) Seasonal forecasting of global hydrologic extremes: system development and evaluation over GEWEX Basins. Bull Am Meteor Soc 96:1895– 1912. doi:10.1175/BAMS-D-14-00003.1 Yuan X, Wood EF, Ma Z (2015b) A review on climate-model-based seasonal hydrologic forecasting: physical understanding and system development. Wiley Interdisciplinary Reviews: Water 2(5): 523–536.doi:10.1002/wat2.1088 Yuan X, Ma F, Wang L, Zheng Z, Ma Z, Ye A, Peng S (2016) An experimental seasonal hydrological forecasting system over the Yellow River basin-Part I: Understanding the role of initial hydrological conditions. Hydrol Earth Syst Sci 20:2437–2451. doi:10.5194/hess-20-2437-2016 Zhai J, Su B, Krysanova V, Vetter T, Gao C, Jiang T (2010) Spatial variation and trends in PDSI and SPI indices and their relation to streamflow in 10 large regions of China. J Climate 23:649–663. doi:10.1175/2009JCLI2968.1 Zhang Q (2005) The South-to-North Water Diversion (SNWD) Project. Front Ecol Environ 3(2):75–76. doi:10.2307/3868512