Overview of the TAMSAT drought forecasting system The TAMSAT drought forecasting system produces probabilistic forecasts of drought by combining information on the contemporaneous condition of the land surface with historical climate data and seasonal forecasts. The system is run during the growing season to provide an evolving picture of the drought- related hazard faced by farmers. 1. System design The JULES land surface model [1,2] is used to translate time series of climatic variables into time series of soil moisture, and other land surface variables. These time series can be used to calculate metrics of hydrological and agricultural hazard, such as mean soil moisture or cumulative rainfall over the growing season. The time series are derived for the whole growing season: past and future. For the historical period, JULES is spun up at the beginning of the run and then driven with observations/reanalysis. Rainfall is taken either from TAMSAT[3,4] or from gauge observations and other driving variables are based either on station measurements (where available) or the NCEP reanalysis. For the future period, JULES is initialized on the present day land surface conditions and run forward in time, driven with multiple realizations of the historical climate (Figure 1). Figure 1: Daily time series of simulated soil moisture. The red dot is the day of the initialization of the forecast. The fine black lines are individual ensemble members. There is an ensemble member for each year included in the climatology.
For a metric, X, for an ensemble member for climatological year j, the ensemble member prediction P j is :!"#$%&'( P! = [X ]!"!#!$#!%"!"#!$%!"#!"#!$%!"#$" + [X! ]!"#$%&'(!"!#!#!$"!! where square brackets denote aggregation of the metric over the period indicated. If the metric represents a mean for the period (eg mean soil moisture), rather than an accumulation (eg total rainfall), the terms are weighted by the number of days. The ensemble mean and variance are: Ensemble mean =!"#$%&'"'() P! number of climatological years Ensemble variance =!"#$%&'"'() (P! P )! number of climatological years As the season progresses, and more historical information is incorporated into the prediction, the ensemble spread reduces (Figure 2). Figure 2: An example of a daily time series of ensemble spread, based on hindcasts initialized on every day of the year. The example given here is mean soil moisture for a single soil layer. The growing season is defined as the whole year. The ensemble spread (standard deviation) is represented by the grey polygon. The dashed lines are the climatological mean and +/1 climatological standard deviation. The red line represents an arbitrary drought threshold. The probability of drought is equivalent to the probability of a pre- defined threshold being breached. This is derived from ensemble mean and standard deviation. For the plot below, it is assumed that the aggregated metric has a Gaussian distribution. Figure 3 shows a time series of drought probability, based on hindcasts initialized on every day of the year. Figure 3: An example of a time series of drought probability (for an arbitrarily defined drought metric), based on hindcasts intialized on every day of the year.
2. Incorporation of meteorological seasonal predictions African meteorological agencies have access to probabilistic tercile forecast information. These forecasts are presented as the probabilities of total seasonal rainfall being in the lower (below average), middle (average) or upper (above average) tercile. Typically, tercile seasonal forecasts have a range of 1-3 months. The drought warning system can incorporate tercile forecasts by weighting the input climatology. The climatology weighting must be based on the meteorological quantity being forecast not the drought metric of interest. Thus, if total precipitation is forecast, the climatology is weighted on this. In some cases, the metrics of drought being forecast by the TAMSAT system are not strongly correlated with the meteorological quantity for which tercile seasonal forecast information is provided. For example, seasonal mean bottom layer soil moisture may be only weakly correlated with total rainfall over the next 90 days. In this case, the weighting of the climatology will have little effect on the prediction. The forecasting system thus implicitly accounts for mismatch between predictions of meteorological quantities and the likelihood of drought. There are several steps to the incorporation of seasonal forecast information. In the example below, the quantity being forecast is total precipitation: The prediction is run as usual Total precipitation for the part of the growing season in range of the seasonal forecast is calculated for each climatological year. The climatological years are ranked. The ensemble members are ranked by the climatological precipitation input. The output is weighted by the tercile forecast for the period of the growing season within range of the forecast. The output is aggregated, and the ensemble mean and standard deviation calculated, taking into account the weighting. 3. Presentation of operational forecasts In the experimental system, the following quantities are considered: Seasonal mean olumetric soil moisture in the following soil layers: 0-0.1m, 0.1-0.25m, 0.25m- 0.65m, 0.65m- 2m Seasonal mean total volumetric soil moisture from 0-2m. Seasonal mean soil moisture availability factor (beta) Seasonal mean net primary production (NPP) Formal definition of these variables are given in the published descriptions of JULES[1,2] Standard reports are disseminated every ten days for each station. Figure 4 shows examples of the plots produced (for beta):
Figure 4: Left: Daily time series of the forecast variable for two years, commensing on January 1 st of the year of the start of the season. Vertical green lines denote the growing season; the vertical red line is the day of initiation of the forecast. Middle: Kernal density plot for the climatology (grey polygon) and the forecast variable (green line). Right: Cumulative quintile plot - i.e. the likelihood that the variable in question falls in a given climatological quintile or lower. Note that the quintiles are calculated using by fitting an empirical cumulative distribution to the data i.e. the data are not assumed to be Gaussian. 4. Potential predictability Direct evaluation of drought forecasts against observations is not possible in Africa, because such observations are rare. An alternative is to assume that the metrics derived by the system using observations/reanalysis are accurate and to ask: How early in the season can we be reasonably certain that there will/will not be a drought? This question can be asked of meteorological drought (cumulative rainfall deficit) or of agricultural drought (seasonal soil moisture deficit). It is expected that the memory of the land surface will result in some predictability, and therefore that it will be possible to warn of an impending agricultural drought significantly before we can be certain of a large rainfall deficit. In essence, if the soil is dry and the season is well advanced, even if it rains as much as it has in any past year, a drought is inevitable. The converse is true for non- drought years. Once the soil is wet and the season is well advanced, even a very dry end to the season will not result in agricultural drought. The point in time at which we can say that a drought is likely, depends on the distribution of rainfall within the season and on the dryness of the soil. At the beginning of the season, for example, even if the soil is dry, a drought will not ensue if the seasonal rainfall is well above average. This concept is illustrated by Figure 5, which compares drought probabilities inferred using beta and using precipitation during a drought year in northern Ghana. It can
be seen that during a 170- day season, the drought prediction based on beta indicates an impending drought significantly earlier than the drought prediction using precipitation. Figure 5: Time series of hindcast drought probabilities derived using Beta (brown line) and Precipitation (blue line). In this case, drought is defined as <25 th percentile of the climatology. Figure 6 shows the day of the year on which concrete information (probability > 0.75) is available on whether or not there will be a drought. Figure 6: Day of the year (DoY) on which it can be said that there is >75% chance of a drought or >75% chance of no drought, based on precipitation and based on beta. A more formal evaluation can be carried out using the Brier score (Figure 7). It can be seen for most of the season, the Brier score is lower (indicating higher skill) for beta, indicating higher skill.
Figure 7: Brier skill score skill in predicting drought based on hindcasts initiated on each day of the season (DoY 128 DoY 288) It should be noted, however, that for metrics based on the output of JULES, such as NPP and beta, the reliability of the forecast will be affected by the ability of JULES to simulate variability in these metrics. If rainfall proxies, such as TAMSAT, are used in place of station data to drive the prediction model, errors in the rainfall add significant uncertainty to the soil moisture predictions. 5. Conclusions The TAMSAT drought warning system is designed to help farmers and agricultural extension agencies to interpret risk, in the light of seasonal forecast information and observations. The framework can accommodate any probabilistic quantile forecast and can predict any metric, which can be either output from JULES or derived from observations. Further developments will be to develop the method by which NWP seasonal predictions can be incorporated and to constrain the soil moisture and other variables more tightly using data assimilation techniques. The drought forecasting system was developed by Emily Black, Matt Brown, Fred Otu- Larbi and Tristan Quaife with the support of the Walker Institute and the NERC project SatWIN- Scale. 1. Clark, D. B.; Mercado, L. M.; Sitch, S.; Jones, C. D.; Gedney, N.; Best, M. J.; Pryor, M.; Rooney, G. G.; Essery, R. L. H.; Blyth, E.; Boucher, O.; Harding, R. J.; Huntingford, C.; Cox, P. M. The Joint UK Land Environment Simulator (JULES), model description Part 2: Carbon fluxes and vegetation dynamics. Geosci. Model Dev. 2011, 4, 701 722. 2. Best, M. J.; Pryor, M.; Clark, D. B.; Rooney, G. G.; Essery, R.. L. H.; Ménard, C. B.; Edwards, J. M.; Hendry, M. a.; Porson, a.; Gedney, N.; Mercado, L. M.; Sitch, S.; Blyth, E.; Boucher, O.; Cox, P. M.; Grimmond, C. S. B.; Harding, R. J. The Joint UK Land Environment Simulator (JULES), model description Part 1: Energy and water fluxes. Geosci. Model Dev. 2011, 4, 677 699.
3. Maidment, R. I.; Grimes, D.; Allan, R. P.; Tarnavsky, E.; Stringer, M.; Hewison, T.; Roebeling, R.; Black, E. The 30 year TAMSAT African Rainfall Climatology And Time series (TARCAT) data set. J. Geophys. Res. Atmos. 2014, 119, 10,610 619,644. 4. Tarnavsky, E.; Grimes, D.; Maidment, R.; Black, E.; Allan, R. P.; Stringer, M.; Chadwick, R.; Kayitakire, F. Extension of the TAMSAT Satellite- Based Rainfall Monitoring over Africa and from 1983 to Present. J. Appl. Meteorol. Climatol. 2014, 53, 2805 2822.