1. Introduction
Seasonal monsoon forecasts for Asian countries are undoubtedly important to society. This is a scientifically challenging endeavor, however, especially for precipitation. Operationally, observed precipitation is sometimes used for postprocessing of numerical forecasts. This is because precipitation is a localized phenomenon for which it is difficult to reproduce realistic patterns, even qualitatively. Quantitative estimates are difficult to attain in both short-term and long-term forecasts.
There have been many studies using multimodel ensembles to reduce model biases. The superensemble, which uses observation data to minimize bias of multiple model simulation results, has shown significant improvement of precipitation forecast scores (Krishnamurti et al. 1999). An accurate observed precipitation dataset is key for achieving the best forecast. Rain gauge–based, daily gridded precipitation data across India produced better short-term and seasonal forecasts in that country (Krishnamurti et al. 2009a,b; Chakraborty and Krishnamurti 2009).
The Asian Precipitation–Highly-Resolved Observational Data Integration Toward Evaluation of Water Resources (APHRODITE) project created daily grid precipitation data over the entire Asian domain (Yatagai et al. 2009, 2012). The initial purpose of APHRODITE was to formulate reliable, rain gauge–based, high-resolution products. These were for validating high-resolution climate model simulations (Yatagai et al. 2005) and for statistical downscaling of relatively course climate simulation outputs, to make localized precipitation forecasts according to future climate change resulting from the anthropogenic greenhouse effect. APHRODITE products are used for such purposes because they contain substantial rain gauge data and use an interpolation method that considers orographic effects (see next section). Hence, they represent a useful database for transforming biased model precipitation patterns into more realistic ones, especially in mountainous areas. Thus, APHRODITE can furnish training data in the superensemble method, as well as data to use in downscaling (Kumar and Krishnamurti 2012).
Satellite products are also used for superensembles. Satellite estimates usually have higher spatiotemporal sampling rates and fewer missing values than rain gauges, and are hence convenient for use as a downscaling tool. However, for surface precipitation, model and satellite estimates should be validated against rain gauge data (Chen et al. 2008). Above all, an advantage of satellite products over rain gauges is near-real-time delivery. Very few countries have dense gauge networks and can collect their data in a timely fashion and apply quality control and gridding in near–real time. It is also difficult to do so beyond national boundaries, although this may sometimes be necessary (e.g., to determine precipitation in the headwaters of an international river).
Hence, it is useful to adjust satellite products with APHRODITE data during periods of mutual data availability. The superensemble technique requires 1) at least four or five model results, 2) observation data for training, and 3) observation data for validation. The second and third items must be independent (usually different years). Because current APHRODITE products have temporal coverage through 2007 and because models should cover the same period as in items 2 and 3 except when applying climate superensemble (Yun et al. 2003), we did performance tests of the adjustment using currently available datasets. We compare the results of superensemble forecasting for summer monsoon precipitation using APHRODITE and Tropical Rainfall Measuring Mission (TRMM) 3B43 products.
2. Data
a. APHRODITE precipitation data
As reference rain gauge–based precipitation, we used a daily gridded precipitation dataset created by the APHRODITE project (Yatagai et al. 2012), with 0.25° resolution over the Asian monsoon region (APHRO_MA_V0902; APHRO hereafter). The data period is 1998–2004 (7 yr), and the coverage is over 0°–55°N, 60°–155°E. Maximum input data were from 1998, with a gradual decrease through 2004 (Yatagai et al. 2009). We used APHRO after formulating monthly total precipitation.
b. TRMM 3B43 precipitation data
We used TRMM 3B43, a 0.25° monthly precipitation product. It is a combined product of microwave satellite imagery, including that from the TRMM Microwave Imager (TMI), geostationary satellite infrared sensors, and rain gauges. For the latter data type in Asia, only data from the Global Telecommunication System (GTS) were used. Figure 1 shows the distribution of gauge stations used by APHRO and GTS. The APHRODITE project collected data from 2.3 to 5.5 times the number of rain gauges of GTS (this quantity varied with the year). There were substantially more input data around the Himalayas and Southeast Asia, which give precipitation estimates exceeding 1000 mm yr−1 (Yatagai et al. 2012).
Distribution of APHRODITE rain gauges across Asian monsoon region.
Citation: Journal of Climate 27, 3; 10.1175/JCLI-D-13-00332.1
Considering the potential application of this study to operational seasonal forecasting, an observation dataset and multiple model results are required. Hence, optimizing the use of satellite-based data is also expected. Therefore, we prepared three training datasets: 1) APHRO; 2) TRMM 3B43 (TRMM hereafter); and 3) modified TRMM 3B43 with APHRO (m-TRMM hereafter). The time (7 yr) and space coverage is the same as described above. We used all monthly data, but show only results for summer (June–August).
c. Model data
The multiple superensemble requires at least four model datasets. Thus, we used four coupled general circulation model (CGCM) outputs those are the part of Climate Prediction and Its Application to Society (CliPAS) project (Wang et al. 2009). These are from the Geophysical Fluid Dynamics Laboratory (GFDL) of the National Oceanic and Atmosphere Administration (NOAA; Delworth et al. 2006), the National Centers for Environmental Prediction (NCEP) and NOAA (Saha et al. 2005), Seoul National University (SNU) in Korea (Kug et al. 2005), and the University of Hawaii (Fu and Wang 2001). Each model consists of different sets of atmospheric or oceanic GCMs. Forecast precipitation data from the four GCMs were converted to monthly precipitation for the seven years.
CGCM results do not reflect the dynamic field in a year. Hence, the climate superensemble technique (Yun et al. 2003, 2005) uses empirical orthogonal function (EOF) analysis to compare the 7-yr summer model data with observation data prior to the construction of a multimodel superensemble.
3. Method of analysis
a. Preparation of m-TRMM
Pattern of a (slope) and b (intercept), defined in Eq. (1).
Citation: Journal of Climate 27, 3; 10.1175/JCLI-D-13-00332.1
Figure 3 shows time series of the coefficient of spatial correlation between TRMM (m-TRMM) and APHRODITE and root-mean-square error (RMSE) between TRMM (m-TRMM) and APHRODITE, which was calculated monthly over the entire domain described above (Fig. 2). There was large error in summer relative to other seasons because summer is the rainy season across most of the domain. Correlation coefficients improved with m-TRMM compared with TRMM itself, because of the related adjustment.
Time series of (top) correlation coefficients and (bottom) RMSE, between APHRO and TRMM (black) and between APHRO and m-TRMM (green).
Citation: Journal of Climate 27, 3; 10.1175/JCLI-D-13-00332.1
b. Preparation of model data and downscaling
The regression coefficients contain information on model biases (Kumar and Krishnamurti 2012), but this information is not shown here. We defined parameter sets c and d for the three types of observation (APHRO, m-TRMM, and TRMM) and for the four models described in section 2c. The observation data used for downscaling [Eq. (2)] and for teacher data are the same.
c. Synthetic superensemble
The multimodel superensemble approach (Krishnamurti et al. 1999) yields high skill relative to that of participating member models. The approach consists of two phases: “training” and “forecast.” In the former, a parameter matrix is defined so that error between multiple models and observation is minimized. The parameter is defined grid by grid and month by month. For climate forecasts, a slightly modified superensemble method called the synthetic superensemble (SSE) is constructed (Yun et al. 2003; Chakraborty and Krishnamurti 2009) because a CGCM does not represent the circulation field in a year. In this method, expansion of the forecast and observation fields in time is done using principal components (PCs) and spatial EOFs. From the 7-yr dataset (1998–2004), one year of data is separated for use in forecasting and validation.
We computed the correlation (COR), bias, and equitable threat score (ETS) between forecast precipitation of the SSE and independent dataset of the training phase.
d. Study flow
The flowchart of downscaling, SSE training phase, and forecast and evaluation is shown in Fig. 3 of Kumar and Krishnamurti (2012). Here, we summarize the steps taken in the present work.
1) Coarse-resolution precipitation data from four coupled climate models (resolution 2.5°) were bilinearly interpolated to the 0.25° grid.
2) Regression coefficients c and d were obtained by a least squares linear fit of model-interpolated precipitation with that of the high-resolution observational datasets (separately for APHRO, m-TRMM, and TRMM).
3) A cross-validation technique was adopted, in which the year to be forecast was excluded from calculations of the regression coefficients. Coefficients varied spatially and monthly during the study years.
4) These regression coefficients were applied to the forecast year to obtain downscaled model forecasts for that year.
5) The above steps (1–4) were repeated for each year of 1998–2004 to obtain downscaled forecasts of individual models.
6) The final outcome was monthly precipitation forecasts on 0.25° grids across the greater monsoon region, from the four coupled models.
7) We also computed ensemble means of seasonal forecasts from our suite of multiple models, which did not use TRMM or gauge datasets. The seasonal precipitation forecasts were validated against the dense gauge-based seasonal datasets over the greater monsoon region. Metrics for forecast validation included the standard RMS errors, ETS, and bias against APHRO.
8) The same metric was applied for the three observation data, and validation is cross checked with other observation data.
4. Results
a. Precipitation pattern and time series
Figure 4 shows precipitation patterns of APHRO and m-TRMM, along with results of forecasts for summer 1998 (June–August). The superensemble and other forecast experiments were executed monthly, and the figure shows the three-month sum. The correlation between APHRO and m-TRMM is 0.96, and that of RMSE with APHRO is 0.43 (Fig. 4b).
Precipitation pattern (mm day−1) in summer (June–August) 1998 from (a) APHRODITE and (b) m-TRMM. (c) Ensemble mean of summer 1998 precipitation forecast by the four models; (d) summer 1998 precipitation simulated by best performing of the four models; and SSE results with (e) m-TRMM and (f) APHRO used for training.
Citation: Journal of Climate 27, 3; 10.1175/JCLI-D-13-00332.1
Figures 4c–f describe the results of an ensemble mean, a model achieving best performance, SSE with m-TRMM, and SSE with APHRO, respectively. Statistics (spatial correlation and RMSE) shown in the figures are versus APHRO (Fig. 4a).
Downscaling was applied in Figs. 4c–f, so even Figs. 4c and 4d show orographic rainfall along the Himalayas and over South and Southeast Asia realistically, compared with the model climatology/seasonal patterns [precipitation patterns simulated by CGCMs, including the four models used here, were displayed in Kumar and Krishnamurti (2012) and Krishnamurti and Kumar (2012)].
Spatial correlation in the areas shown in Fig. 4 reveals that the SSE with APHRO had the best correlation, followed by the SSE with m-TRMM. In terms of RMSE, the SSE with APHRO had the least error. RMSE time series for the four results are shown in Fig. 5a. This shows results consistent with Fig. 4, the example plot, throughout the years. The SSE with APHRO performed best, followed by SSE with m-TRMM. In the summer monsoon months, the difference among the four [ensemble mean (EM), best model, SSE with m-TRMM, SSE with APHRO] was large compared with winter. In summer, the EM and best model performed similarly; SSE with m-TRMM showed better performance, with about half the error of the best model and SSE with APHRO.
(a) Time series of RMSE computed against benchmark APHRODITE data for ensemble mean (green), best model (blue), SSE by m-TRMM (purple), SSE by APHRO (red). (b) Bias and (c) ETS computed against APHRO. Colors in (b),(c) are as in (a).
Citation: Journal of Climate 27, 3; 10.1175/JCLI-D-13-00332.1
Figures 5b and 5c show bias and ETS according to rainfall intensity threshold, respectively. In every rainfall class, SSE with APHRO gave better performance, followed by m-TRMM, best model, and EM. In terms of bias (Fig. 5b), SSE with APHRO and SSE with m-TRMM were almost the same, but the other two showed large biases for weak (2 mm day−1) and heavy (>20 mm day−1) rainfall.
b. Metrics against other observations
The above three metrics are relative to benchmark APHRODITE data. It may be that forecast precipitation looks like training data, not real values. To evaluate this effect, we derived statistics versus m-TRMM, and TRMM versus APHRO. Results are listed in Table 1.
Evaluation of score statistics against the references (APHRO, m-TRMM, and TRMM). Values in italics denote the forecast was evaluated with benchmark data used for training. Boldface values indicate the best score statistics.
It is clear that correlation coefficients are higher in the case in which the forecast was evaluated with benchmark data used for training (shown in italics). The same was true for RMSE and ETS. RMSEs are lower in the case of evaluation against the data used for training, and ETS was larger in the couple that is evaluated against the same with the training data.
Interestingly, compared with the value itself, the correlation of APHRO (0.91) is the highest among the same pairs (m-TRMM had 0.88 and TRMM had 0.82). It may be expected that APHRO had a low correlation because it has very precise patterns that may produce bias. The RMSE revealed excellent performance, similar to the use of APHRO. The RMSE of APHRO was the smallest at 0.64; m-TRMM yielded 0.80 and TRMM gave 1.0. The result of ETS was a slightly different. The TRMM pair showed the best value (0.65) followed by APHRO (0.62), but this is a slight difference.
These results demonstrate that high-resolution precipitation data from a dense network of rain gauges are essential for improving seasonal rainfall estimation over the Asian monsoon region.
5. Conclusions
Using APHRODITE precipitation data (APHRO), the TRMM 3B43 product, and four CGCMs, we performed an experiment combining an SSE with downscaling. The results were compared with ensemble forecasts lacking observation data and singular (best) model outputs. This revealed the following:
The dense rain gauge network dataset (APHRODITE) considerably improved seasonal summer monsoon precipitation forecasts. An m-TRMM product, which was created by modifying TRMM 3B43 using APHRO data, gave similar results. Skill scores of these two experiments were much superior to the best single model and ensemble forecast lacking observation data.
Availability of a dense rain gauge network is imperative to success of the seasonal forecast. It is confirmed by cross-evaluation of superensemble with using other observation data than the teacher data used in the training phase.
The method using satellites can be effective in real-time application.
Acknowledgments
This paper is a contribution to the APHRODITE project, supported by the Global Environment Research Fund of the Ministry of the Environment, Japan. This work was also supported by NASA GPM Grant Number NNX13AF75G. Part of this work was supported by a research project on Human Life, Aging, and Disease in High-Altitude Environments administrated in the Research Institute for Humanity and Nature and a Research Collaboration with Disaster Prevention Research Institute, Kyoto University.
REFERENCES
Chakraborty, A., and T. N. Krishnamurti, 2009: Improving global model precipitation forecasts over India using downscaling and the FSU superensemble. Part II: Seasonal climate. Mon. Wea. Rev., 137, 2736–2757.
Chen, M., W. Shi, P. Xie, V. B. S. Silva, V. E. Kousky, R. W. Higgins, and J. E. Janowiak, 2008: Assessing objective techniques for gauge-based analyses of global daily precipitation. J. Geophys. Res., 113, D04110, doi:10.1029/2007JD009132.
Delworth, T. L., and Coauthors, 2006: GFDL’s CM2 global coupled climate models. Part I: Formulation and simulation characteristics. J. Climate, 19, 643–674.
Fu, X., and B. Wang, 2001: A coupled modeling study of the seasonal cycle of the Pacific cold tongue. Part I: Simulation and sensitivity experiments. J. Climate, 14, 765–779.
Krishnamurti, T. N., and V. Kumar, 2012: Improved seasonal precipitation forecasts for the Asian monsoon using 16 atmosphere–ocean coupled models: Anomaly. J. Climate, 25, 65–88.
Krishnamurti, T. N., C. M. Kishtawal, T. E. LaRow, D. R. Bachiochi, Z. Zhang, C. E. Williford, S. Gadgil, and S. Surendran, 1999: Improved weather and seasonal climate forecasts from multimodel superensemble. Science, 285, 1548–1550.
Krishnamurti, T. N., A. K. Mishra, A. Chakraborty, and M. Rajeevan, 2009a: Improving global model precipitation forecasts over India using downscaling and the FSU superensemble. Part I: 1–5-day forecasts. Mon. Wea. Rev., 137, 2713–2735.
Krishnamurti, T. N., A. K. Mishra, A. Simon, and A. Yatagai, 2009b: Use of a dense gauge network over India for improving blended TRMM products and downscaled weather models. J. Meteor. Soc. Japan, 87, 395–416.
Kug, J.-S., S.-I. An, F.-F. Jin, and I.-S. Kang, 2005: Preconditions for El Niño and La Niña onsets and their relation to the Indian Ocean. Geophys. Res. Lett., 32, L05706, doi:10.1029/2004GL021674.
Kumar, V., and T. N. Krishnamurti, 2012: Improved seasonal precipitation forecasts for the Asian monsoon using 16 atmosphere–ocean coupled models: Climatology. J. Climate, 25, 39–64.
Saha, S., and Coauthors, 2005: The NCEP Climate Forecast System Reanalysis. Bull. Amer. Meteor. Soc., 91, 1015–1057.
Wang, B., and Coauthors, 2009: Advance and prospectus of seasonal prediction: Assessment of the APCC/CliPAS 14-model ensemble retroperspective seasonal prediction (1980–2004). Climate Dyn., 33, 93–117.
Yatagai, A., P. Xie, and A. Kitoh, 2005: Utilization of a new gauge-based daily preciptiation dataset over monsoon Asia for validation of the daily precipitation climatology simulated by the MRI/JMA 20-km-mesh AGCM. SOLA, 1, 193–196, doi:10.2151/sola.2005-050.
Yatagai, A., O. Arakawa, K. Kamiguchi, H. Kawamoto, M. I. Nodzu, and A. Hamada, 2009: A 44-year daily gridded precipitation dataset for Asia based on a dense network of rain gauges. SOLA, 5, 137–140, doi:10.2151/sola.2009-035.
Yatagai, A., K. Kamiguchi, O. Arakawa, A. Hamada, N. Yasutomi, and A. Kitoh, 2012: APHRODITE: Constructing a long-term daily gridded precipitation dataset for Asia based on a dense network of rain gauges. Bull. Amer. Meteor. Soc.,93, 1401–1415.
Yun, W.-T., L. Stefanova, and T. N. Krishnamurti, 2003: Improvement of the multimodel superensemble technique for seasonal forecasts. J. Climate, 16, 3834–3840.
Yun, W.-T., L. Stefanova, A. K. Mitra, T. S. V. Vijayakumar, W. Dewar, and T. N. Krishnamurti, 2005: A multi-model superensemble algorithm for seasonal climate prediction using DEMETER forecasts. Tellus, 57A, 280–289.
