 Thank you very much Adrian and thank you for inviting me here as well to Trieste. It's very nice to be here So it's Adrian said I'm working at the European Center So you know all about the European Center now from Frederick from last week, right? So I will keep it short in terms of introduction So I will speak about what we've been doing for the last four or five years on drought forecasting tomorrow we'll talk about flood forecasting as well, but so This is the just a basic outline of the talk today. I will talk a bit about In the season seasonal forecast and then I will concentrate on Speaking to you about how we do this drought forecasting and a little bit more technical terms a little bit about monitoring and forecasting and How we do the merging of our different products as well to create this and I will talk just a little bit on on finals on A short case study of is done on in the Horn of Africa There is a lot more case studies that I have If you're interested in this I'll give you a literature list at the end as well if you want to go to look more into details because I'm going to talk quite generally now about how we do drought forecasting and monitoring But as Andy said in the morning everything has to be Adapted to the local scale if you want to have something useful like indices or threshold something like that So we have some some studies showing how we worked on on local adaptation of the of the forecast as well Just a recap probably covered this last week as well So the European Center we are doing quite a large number of different forecast products So we have the high-resolution forecast on the 10-day lead time We have the ensemble forecast which is issued twice daily 51 members up to 15 days So they they're really interesting thing to you start using now And I guess is what this course is about as well. It's looking at going beyond these two weeks into the monthly 46 two months and then going on to the seasonal Horizon as well, so we have the monthly forecast now up to 46 days which going into this perfectly into this S2S region and then of course we had a seasonal forecast and Then we also have the reanalysis. Did you talk about that? Yes, but last week was your offer of that Yeah, so we I'm going to using this talk focusing on the seasonal forecast and the reanalysis products Tomorrow I will mention a bit more on on how we're starting to use also the monthly forecast So we've had quite a few projects lately looking and Applications of our forecasts and this is just a Summarized table of the forecast we have been using or that the these kind of problems we have been tackling so floods Obviously, that's one of our main things we're doing now in the application group at ECMWF. We're using High-resolution ensemble and the reanalysis forecast, but we're also looking into monthly and seasonal. So those should actually be added to it So we're looking also at fire forecasts droughts We had a project with agent on malaria. So we are also malaria forecasting The project is finished now, but they're still available. We're also looking at other things like wind and cyclone. So we have a large range of Problems that we're trying to tackle on these timescales so droughts This is quite important thing to try to to monitor and forecast the spin as you all know Dr. Scott calls a lot of human disasters in terms of actual people dying from it or Immunitarian crisis in Horn and Africa later just a few years back and the US drought had a Quite costly drought and the the one in California now is supposedly one of the worst in They say in in recorded history, but I'm not sure if that's really it's still ongoing. So we'll see how how How much that will be the impact from it? So there's a lot of different Things that happen when you have a drought and this was a difficult to Sometimes classify the severity of a drought event because it is really in space and time and The impact is also depending on how well a country can cope with the effects of a drought as well So there's a lot of social Sites of this as well. I will not talk about the actual impacts I would talk more of the hazard hazard forecasting of droughts, but that's another very interesting topic as well How to deal with the effects of the drought? So the seasonal forecast then It's a summary of the weather events during a given season and The question here as well if we can use the seasonal forecast to provide an outlook from the evolution of a drought because when you're in a drought you've already Hopefully good monitoring facilities so you can look you can see what's happening But then you want to know when will this drought end or possibly also when if you're in a good position when can you expect a drought to start as well and we have some skill in our seasonal forecasts and on the Left-hand side here. We see some correlation anomaly correlation For three months forecast issued in January over the US So you see that on average we have the the redder or yellow well if you have read you have a very good skill If you have blue we have no skill or so we have a good correlation between the seasonal forecast Issued in time so they have a good correlation in some parts of the world and the right-hand side We have over Africa here. You can see that we have a good correlation in East Africa so we can say that we have some Some idea that we can use our forecast for for these purposes and other parts We we have very little skills so this is an important issue as well as and it mentioned as well We have no when can the systems to be used be useful when can they give some information that it's actually skillful so monitoring is also very important because that's really gives us the the baseline where we are and There's a problem with monitoring often that we don't have enough good data and the data quality is We have to rely on maybe global products if you have local products Precipitation for example or temperature you you can use that but on a global scale We have to rely on on what's available and that's usually not a very very good quality So I will also show an example of how we can overcome this by using our own Weather forecast as a sort of proxy monitoring tool for forecast as well There's quite a lot of Systems as well already in place for for monitoring droughts There's some examples here We have a US drought monitor on the left to be have something called the African drought observatory on the right Which is run by JRC and they typically give this kind of products where you have an outlook or a monitoring product for for droughts and On the bottom here, we have a regional climate outlook for for Eastern Africa as well the Greater Horn Africa and I will come back to this case study as well where we did actually try to mimic the seasonal outlooks that did the That this Region climate outlook are doing as well So that's a good example where we can try to take our own global models and something useful on on the more of a local scale or a regional scale in this case So drought indices Maybe you've heard this already, so I just recap a bit on why we're using indices So the indices a drought is represents an anomaly in terms of the past climate and that can be the observations or models so we're trying to Make something that's easy to understand and comparable in space and time an anomaly is Not the same everywhere. So we're trying to Go away from the anomaly into more of a standardized format. So For example, if you talk about metrolocal drought, which is one that happens when you have a lack of rainfall We have something called the standardized precipitation index and I will explain to you in detail how we do that How many have heard of that by the way? So you're quite familiar with it already. Okay, it would be For you, so I haven't heard it, but I'll just explain a bit about how they works and You can also use something looking at drying spells on on the thresholds for example number of rainy days or Maybe precipitation amounts that are important for vegetation can also be used Now on we also have had radical drought Which means that you have a lot lack of water available for for river discharge and for for watering for irrigation And there's a lot of indices have been developed as well. You have the palma drought survey to index You have a standardized run of index soil moisture anomalies And then you can also look at actual cultural drought when you actually have not enough moisture for a vegetation. So there's a lot of different ways of Sort of standardizing droughts to looking at depending on what the problem you're looking at I will talk about the the SPI today Because we don't have time to cover all these things, but SPI is quite useful because you can use it for for many many purposes as well and it's recommended by WMO to use as sort of the standard drought index So it's quite known and many weather service and stakeholders already know about it So that's that's good things value until they explain it in detail It's only based on monthly precipitation, which also is quite important to stress here as well So we're not taking into consideration ever for transpiration. There are other indices. There will be developed later We should also look at sort of the the ever berateive Effect as well But the good thing is very quite simple to use if there precipitation you can calculate the SPI You you can look at different accumulation periods You can look at three months Six months nine months 12 months depending on what you're interested in as well So what you need is a monthly series of precipitation It can be local observations and or it can be modeled it can be a region average or you can go go into a point average as well or a point Observation as well. What we do need is long and homogeneous series. So at least 30 years of data we need to have And then of course you need to select your your accumulation time So you do sort of a time filter of the of the your time series It's just a submission of running it's running average and then you normalize this into a precipitation distribution So you transform your your anomalies into standard normal distribution With the subdivision of one so We do this standardization for each calendar month typically So all the January's are pulled together or the three months that ends in January for example So when I talk about a January SPI on a three-month basis, this means the precipitation of a December November December January which is issued in January That means that we are taking away the seasonal seasonality of the of the forecast as well or the the monitoring And this is important to remember that some you can have an SPI in a dry season but it doesn't really mean anything because Dry season is not the important season important season is a wet season So you have to also be careful on one when you're looking at it what where you have a good skill you want to capture that the Important months for your for your territory for your organism for your region So the the principle is that you we fit Distribution function you can use a non parametric function for this, but you we are using a gamma function so here you have your Your precipitation Okay, and the the red dots here is are your cumulative precipitation for your point or for your region The black line here is the fitted gamma distribution to those points and then for each As a top button, okay Okay, perfect. Thank you. So The first step is to fit again in a distribution to the to your observations And gamma is gamma distribution is a quite useful used for this and it's Probably the best one if you're looking at the parametric distributions You can also use your just empirical distribution function as well And then you look for each value You find a probability on the fitted the gamma distribution here so you find your your probability of an event happening and Then you look at the inverse normal function to go back and find The normal the normal distribution what that would be in the normal distribution So then you move from from a normal is into a normal distribution And from there we can categorize this From normal distribution this this indices into normed conditions or drought conditions Usually drought conditions are below minus one is one standard deviation, but you can choose basically which Treshold you want to use but minus one is typically and dry drought So as I said before there's no rule really what what accumulation period you want to use it really depends on your your Application if you're looking at the three months of the accumulation. We are looking more interest maybe in soil moisture It could be crop production in rain for the areas If we go into six and twelve months horizon We are maybe more interested in water reservoirs river discharge groundwater things like that So it depends on what your application is really and this is an example where we have looked at the the upper Niger Basin So here you have SPI for three months accumulation and 12 months accumulation and on this plot here I've just pictured the the correlation between the river discharge in September and different accumulation times And you can see that if you go to longer timescales the more red you have the more the better correlation is you have as well So if you want to look at river discharge, you should choose a Index with a higher timescale than just one or two or three months But you can also see here that we are if you're looking at July-August we have no No correlation with with river discharge that can happen that in this case because we have too much Inundated areas and we're not really looking at a hydrological model here. We're just looking at precipitation So here we have something else some other processes that are disturbing the discharge So they can't really capture that the correlation with rain at this time So you should also be careful that this is not the hydrological model This is not this is just looking at the precipitation itself. This is the entire basin. Yeah, but this is just an example This is not to really just showing that you have to be careful when you when you're Yeah, we can be basically we have no correlation between precipitation Accumulation over a longer time period and river discharge But this doesn't mean that we don't have a correlation between precipitation and discharge But those timescales are much smaller or much shorter than we can capture with it with a SPI in this case So this is also just to show that you have to look at You have to see where you have a good skill of your of your SPI use so I mentioned in the beginning here that we Don't always have good data. So I'm going to show some examples of how we can overcome that by what we call a probabilistic monitoring and The reason why we're doing this is that this is a picture of the data series called the GPCC Which is the global precipitation climatology Center and they are gathering precipitation data globally and creating a gridded outlook and or gridded product and here you can see are the number of Stations in this climatology going from 19 1980 up to 2010 and it goes back further as well But this is an example but as you can see here There's been a quite a dramatic drop in the number of stations reported because of many reasons either the stations are not maintained anymore Or they're not reported to be back in anymore. So when you're using a product like a global precipitation or Observational product you have to be very careful to see what actually goes into that product because it can be quite large Deviation and you don't have a homogeneous time series often And we see that when we compare with our own reanalysis product as well That they can be quite a bad correlation between certain climatology products or observational products So we have this huge uncertainty in what we think are the observations and this will affect our Product or our monitoring product or because if we try to build the climatology over 30 years, we will we will get into a problem here So the idea here was to try to see if we can use our own ensemble short-range forecast to generate a probabilistic monthly means of precipitation anomalies Because in our monthly forecast, we it's not perfect. It's not it's still a climb. It's still a model output It's not observations, but we are trying to fact focus on the uncertainty So using the EPS we automatically have uncertainty in the precipitation So here here's what we do normally if we try to do a monitoring Using this GPCC products. This is for a six-month period. So if we here's our our starting date and you would usually have four months of quite quality check data and then they will release sort of a short period of two months where they haven't really quality check the data so you'd have a mixture of two different data sources to create your your initial states and There will be quite uncertain So instead we are using we're taking the what we think is the best guess of the Previous climate, which is this monitoring of the GPCC and there we are concatenating that with our own EPS probabilistic in the probabilistic sense So we're using a short-range forecast with 51 ensemble members and The average over there of a particular amount to create 51 monthly means So we're trying to to shortcut this this uncertainty in the monitoring with our own on the forecast data instead For the GPCC in this case by using the observation the monitoring Yeah, okay. Sorry. That's a good point. So in for this case, we can use a handcast data set Did you talk about handcast last time as well? so for the handcast we have a 20 year handcast of the PS as well as we can use So to create a climatology we use the handcast of the operational product And the handcast is the same model as the operational but run 20 years back in time So we have sort of the climatology of the operational product But that's a very important point. So we need to have the the climatology of the operational product. Yes Absolutely not no, but we what we can do since we have the handcasts we can calibrate the model with the observations as well So we we we have looked at both using our era interim and then the seasonal forecast, but we do get a better Skill if we start off with the mixing of the observations with the forecast So you can what you can do is you can you can calibrate your your I think you will talk me about calibration as well this afternoon So you can you can look at the handcast and then calibrate the errors of the mean of the biases of the of the forecast using the observations Yeah, because What we're doing here is we're looking at a time in in the forecast DPS where we have a spread Which is high enough to capture the the uncertainty in the forecast. So we just taking the 48 The short-range forecast here Just to look at as a proxy for observations. So this is not really a forecast It's just like trying to capture the absurd uncertainty in the In the short in the monitoring part Yes, because the reason for doing this is that we we're still in the sort of monitoring phase We're not really going over to the the forecast phase yet So to capture the uncertainty in the monitoring we are using our own forecast as a proxy for that spread So the spread in the EPS should be similar to the spread in the observations as well or the uncertainty in the Yes, there's there's papers on this as well, which you can look for more details I've I don't have time to go into all the technical use here But we do get since we have the forecast we can look at Is this 18 years back in time and for this particular case we had five ensemble members So we have quite a lot of samples from the from the climatology to use. It's constant It's in this case. We have well this case. We have to use a 20 last year In this case because we only have the the or actually 18 years, but it's not now has been increased to 20 years But you should keep your reference period constant yes Otherwise you introduce if I well We are running a hind cast for the EPS So that's that's something done at our at the incident we have so every Thursday We run 20 years back in time with exactly same depression model and it's been now be extended to 46 days as well, so it's A good product to use. I don't know if you if for research purposes. I'm sure you can Try to use this as well Because it's something that that gives you a climatology back in time as well I said this is this is why we we use this probabilistic monitoring and this is an example from the Horn of Africa So we are comparing here now The spread from the EPS, which is the black line here And then we also had to do sort of an inflation of the actual spread as well. So we did just an variation inflation This is the gray one here the red one is the gpcc products and Yeah, if you compare them you can see that we are most of the time within the range of the of the gpcc It's not always the case we are but but in this case when we have this kind of probabilistic monitoring We can use that also in a probabilistic forecasting mode as well So we can combine the probabilistic monitoring with probabilistic forecasting So then you get into an example here from from from Africa where we can See if I have the Yes here. So here you have Here you have your monitoring period and then you combine that with a seasonal forecast. So you have to If you want to look at a 12 month of accumulation You have a mixture observation and forecast and if you have and the longer you go to Well, it depends on the lead time as well. So if you're looking at a 12-month lead time On a certain verification date You will have a mixture of observation and forecast depending on how long you are as well And here we are doing a bias correction of the seasonal forecast as well So we are looking at the the climatology of the feet of the Of the of the seasonal forecast to correct for the biases. Yes. So this We can you can either do quite simple bias correction just of the mean, but we also have done a Quantile matching of their of their forecast as well And that that does improve this field quite quite quite likely policies in the forecast Yeah No, this is not not here, but we have done that as well and I'll give you some examples of that as well I'm not gonna Because I didn't have time to go into that in detail here, but we have done different kinds of Bias corrections to look at how that can impact the results as well And here's here's what we look like in real time when you're doing a forecast so here's for the for a limpopo basin here in the South Africa and Here's just an example of a drought that developed in or in 91 to 92 the The gray here is the climatology or sorry that the the black lines here Is the actual climatology forecast just taking The pre historical observations using them as the forecast that is sort of like our benchmark forecast the blue one here is the seasonal forecast and the red one here is the verification of the This is a gpcp and the purple one is our own seasonal or reanalysis they are into him And you can see here just by looking at the drought of evaluation on this two Observational products that we're going from a fork drought the start sort of in October 91 and then deepened into February in March 92 and if we're using our our seasonal forecast, which is the blue one We would actually miss the uncertainty this case We wouldn't we would be not with a lead time of three to four months. We would not be able to capture this event With the climatology we would be in part with the climatology But moving closer to the event we can actually pick up some signal that the drought is emerging So we did we can't pick up the signal with a three monthly time And moving in forward into the time here We can see that we can actually capture the depth of the of the drought better than the climatology So we do have some skill here, but it's not perfect as well It's a lot of things that can be done to improve the forecast, but just an example of how you can use the the forecast for Drought monitoring and forecasting as well and I'll skip this one and you go to An example of a regional Project that we did in the Horn of Africa they produce before the rainy season they produce Consensus climate outlook for September to December and this is based on observations and a bit of model output as well So they can they Try to do a situation where they can say whether you're above normal or if you you expect to be near normal or below normal So they produce this kind of outlook maps and for our forecast we can actually See that we can We can do precipitation forecast for this period as well. So this is our forecast going into time and If we try to do sort of an SPI type of product We can create a similar kind of probabilities of being above normal or normal or below normal so we have defined here sort of same probability Quant or quantizers they have in the outlook form and in the September December precipitation In this in 2010 it was associated with the La Nina event. So we can actually say that We can't forecast the can this kind of events for the for this rainy season We're quite long lead time as well and here's another example where we try to forecast the March and May forecast in 2011 and for this particular event We didn't really do really well because we we can't really capture that the March to May forecast In our product we have a much better skill in the in the autumn rains in the summer rains or spring rains so for Horn of Africa for this case study we found that we we have a Quite good skill in October December season with a connection with El Nino the forecast for March to May season really doesn't have any skill at all. So We need to evaluate the monitoring in the forecasting for each each region really to say something about the usefulness of the forecast But also knowing that there's no skill is also quite useful because we can try to understand why what are we missing and To avoid that these are used just off the shelf. We need to know exactly when they are useful or not So just some fine remarks here that the forecasting is strongly depend on good quality monitoring in seasonal forecast products As a for monitoring local observations should always be used because we found that that's the really the best product if we use error interim you have to be very very careful to do proper validation of it and It's a short an example here. We can use the short range forecast to generate the probabilistic monitoring tool as well and We can you take advantage of this past forecast the handcast of the seasonal forecast So you can do a robust verification and you can also apply this bias correction method so for to tailor your application for a specific region and The as always the forecast is not better than of SPI is not better than the forecast of precipitation So we have to know whether our presentation forecast or are useful for a specific region as well and I'm not going into other fields like your surface temperature and soil moisture, but this can also be used of course as well In your in your applications, so I'm just focused on on precipitation for this specific case This this talk here. So This is going to be give to you as well after the lecture. This is a Some publications which from this is drawn as well and there's a lot more information on on the case studies There's one case study on them only important. We look at bias correction. We also have some papers we're looking at Seasonal focus in the climate change perspective as well and these papers are found in in hydrology in Earth system sciences. There's a special issue on droughts We should reason 2013. So if you look on Hess online, you can find all this information But I'll give you to you after the lecture as well. So That was a very brief overview of the drought. So welcome to questions