 So, hello everybody and welcome to the EGU General Assembly 2024. This is press conference number six of our seven press conferences that we are holding this week at the General Assembly. My name is Hazel Gibson, I'm the EGU's Head of Communications and with us today we have four excellent speakers here to present on the topic of today's press conference. This press conference is titled Preparing for our Hot and Windy Future and we have as I said four speakers speaking on a range of subjects connected to this theme, three of whom are joining us on site as you can see and one who is participating online. If you are a member of our online audience I would ask that you mute your microphones until the end of the presentations as we will hold all of our questions until that time and you'll be able to answer all your questions after all of the presentations have been given. So now I would like to introduce our speakers for today. First of all we have Javier Martinez Amaya from the Image Processing Laboratory in the University of Valencia in Spain, Dominic Schumacher from the Institute for Atmospheric and Climate Science ETH Zurich in Switzerland, Yulia Momkin from the Institute of Meteorology and Climate Research at Karlsruhe Institute of Technology in Germany and joining us online we have Philippa Glasleitner, the Climate Change Research Centre University of New South Wales in Australia. So we will start with our first speaker of the session who is Javier Martinez Amaya. Thank you. Okay, hi everyone, thank you for this opportunity to talk about my work here. I'm Javier, a last year PhD student from the University of Valencia and today I'm going to talk about Medicaid tracking and forecasting using Artificial Intelligence, a work that I developed along with Veronica Nieves and Jordi Muñoz. This research is co-founded by the European Space Agency and the Regional Ministry of Science and Education. First of all I want to show you this image of these three extreme cycling events, Hurricane Florence, Hurricane Linda and Medical Janus. And I want to highlight the similarities of the three cases. As you can compare the eye, the egg wall and also the shape are similar between them. And that's what we have used to also construct a forecasting model for for Medicaid, adapting a methodology previously used for tropical cyclones. With that, I want to pose a question which represents our primary challenge, which is how can AI contribute to extreme cyclone forecasting? For that, well, as it is known, Medicaid and hurricanes are natural hazards whose intensity is projected to increase as it seems to indicate the last IPCC report. And for that, what we propose is to use a hybrid framework of AI to forecast the intensification of extreme cyclones, especially we are focusing on the forecasting of rapid intensification cases, which traditional models are limited in when they try to predict them. For that, we prove that machine learning could be a plausible solution in this field. I'm showing here the main developed regions of the different cyclones. In the upper figure, we can see tropical cyclones where they can develop. And in the figure below, we can see where the medicans, the Mediterranean hurricanes, tend to develop. In any case, tropical cyclones can be grouped in different categories depending on the maximum wind speed value, ranging from tropical depressions to category five. And as it is typically done in extreme weather event predictions, we face this problem using a binary classification to distinguish between non-extreme events, which in this case are tropical storms, and extreme events. In this case are from category three to category five. So we face this problem as a binary classification. To predict these extreme events, there are different approaches, such as the traditional models or the AI classification models. In our case, we have to use an AI model, which captures the non-linearities without using atmospheric or numerical equations. And as they don't need these equations, we can have construct models that need a low computational time to be used. And also, AI models have the capability of forecasting wrap-intensification events. So what we have done is to construct a tool that can rapidly assess and predict wrap-intensification cases, which are the most extreme. So what are the processes that we have followed? First, we have pre-processed data. We obtained brightness temperature images from radiant images of satellite images, using also cyclone tracks. With these images, we have extracted different features, which are the brightness temperature difference between the inner and the outer part of the storm, the site of the cloud, and also some morphology traits. Moreover, we have added high-level special features obtained from a convolutional neural network model used as a feature structure. With all this information, what we have done is to construct a random forest model in order to predict extreme cyclone events between two, three days in advance. For that, we have explored different sampling strategies, such as under-assembling, penalizing, and combining it with augmentation techniques. With that, we have used a trying test split. We have separated our data set between train and test, using a separation of 80% to 20%, combining this with a scaffold iterator. Once the model is constructed, the outcomes of the model can be or non-extreme or extreme events. And what we have done to assess the quality of our results is to use Cohen's Kappa and precision value over the extreme class. We have chosen these metrics among a lot of metrics that we could have chosen, but these metrics represent, well, our goal represents what we want to achieve, which is how many cases we have predicted correctly out of all the cases that the model has predicted as an extreme event. With that, I'm showing here the results of our studies. First of all, for the combined Atlantic and Pacific regions, we have obtained a mean average value of almost 85%, using the best strategy that we did. And for the case of the Mediterranean cyclones, for the medicans, we obtained a relatively low precision due to the scarce number of events and also due to the lack of a tracking method, in this case. So, to sum up, we have constructed CNN Random Forest Method, a hybrid AI method, in order to forecast extreme cyclone events. And with this model, we have obtained accurate predictions up to two or three days in advance. And moreover, this tool has let us obtain some promising forecasting results regarding rapid intensification events. You can find more information in the AI for Asian website, where we have also developed an interactive tool where you can see all our results. And you can also see our results in our studies. Finally, what's the future world, what's next now? In our case, what we are thinking about doing is to add some intermediate intensity scenarios. That would imply the use of other related variables such as the ocean state. And we are, the next research line is this. And other research lines could be the use, the study of additional regions, such as the Indo-Pacific, to test also the generalizability of our models. That's all for me. And I will be happy to answer any questions later. Thank you very much. We will now move to our next speaker. That is Dominic Schumacher. Give us a couple of minutes just to change the slides. And thank you very much. Okay. Yes, why do climate models struggle to represent the strong summer warming in Western Europe? Let us take a look at this summer warming first. On the left-hand side, you can see the observed warming between 1980 and 2022. And on the right-hand side, you can see the difference of this warming and the warming simulated by climate models. And there is lots of blue. So we really have a widespread warming under estimation. In the following, I would like to focus on Western Europe as suggested by the title. And since this is a bit ambiguous, I show the region that I work with. It's indicated by this black contour. But before I get to show you some results, I would like to explain that regional warming has actually two potential contributions. And I think only one of them is fairly obvious, or it might be the one that you're already thinking of. The primary contributor that we would expect is, of course, that we have emitted lots of greenhouse gases, which essentially causes heat to be trapped in our earth system and it also triggers climate feedbacks. And to put this in really simple terms, it's just that the air temperature increases, the air is warmed up. That's one of those contributions. But if you think in a very regional perspective, instead of only heating up the air, you can also change the circulation just ever so slightly that, for example, you would bring air from another source region. So for example, if the winds shift and they blow a bit more from the south than in Western Europe, we would expect that to also contribute to the warming. And I will get back to this very soon. For now, I would like to show to you just how much Western Europe in summer has warmed across 1988 to 2022. On the left-hand side, you see a black marker. Those are our observations. The warming amounts to about 2.3 degrees Celsius. And I show you global models as well as regional models. And I hope you can see that these models generally really do not manage to capture the strong observed warming. What I haven't said yet is that the key difference between these global and regional climate models is their resolution. The global models here, CMIB-6 models, coupled model, Intercom-Marism Project Phase 6, those contributed or were used for the latest IPCC assessment report, they generally have a spatial resolution of about 100 kilometers. Whereas the regional climate models have a much higher resolution. And so each pixel is about 10 kilometers wide. And that is why these regional models are so important for countries such as Austria, for example, where we have mountains that you can imagine are not very well represented when we have a pixel size of 100 kilometers. And this is why the regional models are used in many countries to provide climate services and really serve as the foundation to inform on future changes. So these regional models really, ideally, should do a good job at capturing the observed warming. But as you can see, this doesn't seem to be the case. Now, I will get back to the circulation changes that I mentioned before. And actually, it's a lot easier if we essentially remove their effect, because then we get a more clear idea of the human-induced warming. And this is really what we want these models to simulate. And the global models actually do a pretty good job. And there isn't much of a problem there. But for the regional models, there is still quite the systematic underestimation. So in the following, I would like to zoom in and look at only these regional climate models. Specifically, I now show you the same thing as before, but I essentially added a dimension. I expressed this human-induced warming as a function of sunlight intensity change. And I think you will find this fairly intuitive if the sunlight becomes more intense than this warming is stronger. And that's what you see in here. The purple markers, by the way, are our regional climate model simulations. The black one is our reference. But why is there an increase in sunlight intensity? Well, the primary reason is that back in 1980, the air over Europe was a lot more polluted than it is now. It still is, but less so. And so I hope it's visible. I tried to visualize this with a dim sun, little bit similar to not so long ago when we had dust from the Sahara, but the key difference is that one, at least the dust itself from the Sahara, that is natural, but the air pollution isn't. But the effect is the same, the sun is dimmed. And so now, since the air is less polluted, we have essentially, we've roofed some of this artificial cooling effect, which to some extent managed to mask the warming that we caused with greenhouse gases. And the key reason why these regional climate models failed to reproduce this human-induced warming is that most of them, and those are the blue markers in here, do not represent this effect. They essentially assume that the air pollution is constant. And as you can see, the ones that do not make this assumption are a lot closer to our black marker, which is our reference, the observation or the observation-based estimate. Now, yes, I also added this to try and provide an estimate that this amounts to about 0.5 degrees Celsius. So it's not negligible. And the thing is, we expect that the air pollution in the future will decrease further, even if we keep emitting lots of greenhouse gases. So just to clarify, those are separate things. And as a consequence, we also remove this artificial cooling effect more and more as we progress into the future. And I have to emphasize here that if there is one thing that you should not take away from this, it's that air pollution is good. It isn't. The WHO estimates that 7 million people die every single year because of air pollution. I also checked the literature. This is clearly not my expertise, but the higher estimates seem to be close to 10 million. So these are really high numbers that I cannot even wrap my head around. But so this is not to say that we should keep polluting the air. What I am saying is that what is actually happening, the line in here that is more realistic is the sort of orange or, yeah, let's say orange one, the air pollution is changing. It is decreasing and we expect it to decrease further. And the issue is that a lot of the regional climate model simulations that we currently have, this should no longer be an issue for future simulations because this will be taken into account. But these simulations really underestimate both the historic warming as well as most likely the future warming. Of course, I cannot show your reference here, but based on our physical understanding, there is little doubt about this. And with that, I would like to thank you and pass on this remote. Thank you very much. Our next speaker will be Julia Mumpkin. Just give us a minute to change the slides. Thank you. Yeah, then also welcome from my side and thanks for the opportunity to present my work here. It's actually a quite recently published paper I would like to present where we analyzed windstorm losses in Europe in different types of data sets. As you probably know, windstorms are one of the major natural hazards that regularly affect Western and Central Europe. And a single event can lead to millions of euros of losses. For example, by uprooting trees or disrupting the power supply or the transportation services, but also by building damage. And the information on the loss of windstorm causes is also important to assess the windstorm risk and to develop adaptation or mitigation strategies. A windstorm can be described from different perspectives. So for example, purely from the meteorological side by analyzing its core pressure or the associated maximum wind gusts. But of course, you can also describe a windstorm from the loss side. And this is what our study focused on. So we focus only on the impact side, but also the impact side can be assessed from different perspectives. So what we did in our study, we analyzed five different data sets which come from different types, let's say. So we had insurance data, we used a disaster database and we used three different meteorological indices, which kind of combine meteorological variables and insurance aspects. And the two questions we wanted to answer is how comparable are these data sets and what can we learn from the information they provide? So what we first did for the common period of the data sets from 1999 to 2022, we collected all storms that were reported in the data sets and ended up with 94 storms. However, we quickly realized that only a few of them are common in all data sets. It was only 11 storms of them, so less than 15%. Whereas a large majority of over 60% were only reported in single data sets of the five. This is also seen when we look into the storm numbers per winter. You see quite large differences. If you focus, for example, on the winter of 2007 or 2008, there were not a single storm reported in the insurance data, but we have, for example, five storm events in one of the meteorological data sets. And there are different other years where they agree on the number of storms, but what I also have to highlight here is even if they agree on the number of storms, it doesn't necessarily mean that they have the same storm reported. It could also be different events. The different storm numbers are also reflected if you look into the average storm frequency for winter. This can also largely differ between the data sets. Typically, we see more storms in the meteorological data sets than we see in the insurance data sets. What we did after that, we kind of took the loss values, the different data sets provided, and ranked the storm events based on the loss values. And you can do this for all the storms. What you see here is an example of storm Sabine or internationally also called Chiara from February 2022. And we have a ranking for each country that is reported in the data sets. And on the first glance, you can already see that the ranking is quite different in the data sets. We also have data sets like here in the middle. For example, where the storm is not reported in Sweden, although you can see from the black line, which is the storm track that the storm definitely passed by this country. So we don't really know why it is not in the data set. What you can also see is that the differences are a bit reduced when we focus only on common storms in the data set. So when we redo the ranking and focusing only on those storms that appear in all of them. This is also highlighted when you kind of look into the correlation. So between the data sets, so the higher the correlation, the better the agreement between the data sets. When you focus on the top row, so all storms, you see that the correlation is quite low. In some cases, you also have an anti-correlation between the data sets. The correlation generally gets higher if you focus only on common storm events. And we also see that there are some countries where the correlation, for example, in the UK is often lower than for the continent. And with this, I already come to my conclusions or the key findings of our study. So what we saw is that the data sets provide kind of different views on windstorm impacts and we almost had no redundant information. What we think that these differences in the data sets could in future studies, for example, be used to assign an uncertainty range to windstorm losses. So to not say, okay, this is my data set, this is the ground truth, but to rather say, okay, I know there are different perspectives and maybe my real loss ranks somewhere in between. It could also probably be used in insurance windstorm models to test which features are really relevant for calculating correct impacts. And finally, we would say that only a combination of different data sets can really provide a representative picture of windstorm impacts. If you want to have the full details, the paper was published, I guess, two weeks ago. So you'll find it there and also feel free to contact me by email if you have further questions. Thank you. Thank you. We will now move to our last speaker, Philip Arglas Leitner. Please give us a couple of minutes to switch the slides. Thank you very much. Can you hear me? Yes, we can. Okay. Good afternoon. Thank you very much for this opportunity. Today, I would like to present a study on projected heat waves in the context of recent heat extremes. My name is Philip Arglas Leitner. I'm from the Potsdam Institute for Climate Impact Research and from the research group of Sarah Perkins Kirkpatrick, who previously was at the University of New South Wales and has recently joined the Australian National University. This research has been conducting in collaboration with my two supervisors, Sarah Perkins Kirkpatrick and Dolly Stone and our collaborator, Lopez Brunner, from the University of Vienna. Next slide, please. We live in a time where we are routinely confronted with extreme heat events and record-shattering heat waves. This inevitably brings up questions about what the future holds and what future heat waves might look like. And of course, what this in turn means for the human population. So we set out on this scientific journey to get closer to answering these types of questions. Therefore, the overarching question of our study is how do future projections of individual heat waves compare to recent heat extremes? To this end, we selected 25 relatively recent heat waves in approximately 22 different regions across the globe that happened between 2010 and 2023. Here on the next slide, you can see a very rough sketch of our regions that are affected by the selected recent heat waves. In addition, we selected our heat waves mainly based on three parameters. First of all, the events had to occur during the warmest six months of the respective region's yearly cycle. In other words, the focus was on extended summer heat waves. The second was data availability. And finally, we primarily looked for record-breaking and or high-impact events. The rationale being that anchoring future heat waves to record-breaking and or high-impact events in the lived experience is essential for effective adaptation and mitigation and successful communication of how devastating heat waves might become in the future. For the analysis, we used the two most recent generations of the coupled modeling comparison project, in short, CMAP, which is for those who are not familiar with climate model, LINGO, a set of multiple different climate models. And for analyzing the selected recent heat waves, we utilized two independent global reference datasets, namely Berkeley Earth and Euro5. Next slide, please. And now, before we get to the results section, I would like to give you a brief introduction to our heat wave analysis framework. So it will be easier for you to interpret the results. Our framework is based on four parameters and additional subparameters where applicable. The first one is heat wave duration in days. The second one is heat wave severity, which is an intensity index enabling interpreting excess heat relative to the respective regions' climatology. The third one is cumulative heat, which shows the cumulative temperature above the heat, above the heat wave threshold for an individual heat wave. And the fourth one is percentage of locally affected area. This is as the name already indicates, an area-based parameter providing information on how much of the region's area is affected by extreme heat during the course of an individual heat wave in percent of the respective region's overall area. So in other words, the higher the percentage, the bigger the heat wave geographically speaking. Next slide, please. Now I would like to show you some of our results. I would like to know, as this study is soon to be submitted for publication, we would very much appreciate if the following figures were not distributed without our approval. Thank you very much. So here you can see one of the figures of our parameter correlation analysis for our 25 heat waves. As this figure contains a lot of information, I will walk you through it. Each of the dots represents an individual heat wave found in this case in the CMAP-6 multi-model ensemble, which is the latest generation of this set of climate models from 2015 to 2100. On the y-axis, you can see the heat wave length in days, whereas the x-axis shows the percentage of locally affected area, in percent of the respective region's overall area. In this case of this particular figure, it shows the median percentage values. The dot size represents the cumulative heat in degrees Celsius, and the color intensity indicates the severity which you can see on the color bar on the right. In this case, it also shows the median severity of each event. The observed reference event is indicated by the tiny axis at the end of the arrows, with assumed in-window at the other end. And in addition, we introduced a second y- and x-axis, showing the distribution of the individual heat waves along the respective parameter axes. This particular figure shows the results for SSP12.6, which is the second most ambitious mitigation scenario in CMAP-6. As you can see, even in the case of enormous mitigation ambitions, it is still very plausible to exceed the already extreme values of our reference heat waves. As we are currently at EGU, let's for example take a closer look at the case of Europe. Next slide, please. Here we can expect heat waves that are substantially longer than the 2018 European reference heat wave at the end of the arrows. Some, especially towards the end of the century, might come very closer, to even hit our theoretical six-month limit. In addition, we can see that even in this scenario, heat waves might become significantly more severe, both of which feed into the increase in cumulative heat indicated by the size of the arrows. And in terms of the affected area, we can see heat waves that are substantially bigger by a factor of approximately 1.5 or even two of the 2018 reference heat waves, which was already among the spatially biggest heat waves in Europe. Next slide, please. As a comparison, I would like to show you the same type of figure for a climate model scenario, which is closer to where our emissions pathway is currently heading. As expected, here we can see a substantially different picture with a lot more severe and bigger heat waves. When we again take a closer look at Europe, for example, it shows a significant increase in the number and proportion of heat waves approaching or even hitting the six-month limits. And even in the case of the median affected area, we can see heat waves reaching values of close to 90%. In other words, in those cases, the heat wave will on average nearly affect the entire continent. Next slide, please. So to conclude, in summary, we can see that recent extreme heat waves are no match to the projected 21st century count parts. Without even moderate reduction in greenhouse gases, the recurrence probability and exceedance of recent extreme reference values is significantly increasing. And on top of that, the exceedance is still plausible under aggressive emission reduction scenarios. So we argue that, of course, ambitious mitigation is urgently needed. However, it is also necessary to consider heat waves well outside the lived experience for effective adaptation measures. Next slide, please. And finally, I would like to thank my supervisors and collaborators for their help and support, namely Sarah Perpinsk and Patrick, Dahlia Stone and Lukas Brunner. And would like to thank you very much for your attention. Thank you very much, Philip. We will now move to the question and answer portion of today's press conference. If you are joining us online and you would like to ask a question, please either type it into the chat and I will see it and read it on your behalf, or you can use the raise hand feature in Zoom and I will call on you and you can unmute your microphone and ask the question yourself. If you are in the room with us today, please raise your hand and I'll bring the microphone to you. So does anybody have any questions for us? Anyone online? Oh, thank you. Hi. So a question about this underestimation of sort of future heat. Can you just sort of go into more detail into why does it matter if these regional models are underestimating the heat in the future? Why is that so important? Yes, thank you very much for the question. It's a great one because I didn't explain all of the implications. So one thing is that this doesn't just affect the mean summer warming, but actually these effects are even more pronounced at the time scale of heat waves because you can imagine that if we have a heat wave, we have clear skies, we have lots of the sunshine is very intense. And so if models don't take air pollution changes into account, they will underestimate intensity increases of future heat waves even more than they struggle with the mean summer warming. And it's problematic because a lot of European countries strongly rely on these simulations or these projections to plan ahead for the future. And so essentially this just shows that for a given scenario, and I should say what I showed was for a pessimistic scenario, but just to illustrate, even if we keep emitting lots of greenhouse gases, we still expect to reduce the air pollution. That is simply something that the majority of these models does not capture. And so consequently, if we rely on an average of these models, then it is in this sense too optimistic. And can I also just check that graph you showed, was that still stripping out the effects of circulation changes? I mean, the one that extended into the future? No, no, there was no, the circulation is also in there. So that does not rely on any sort of additional assumptions, that was just the model output averaged and visualized. Yeah, thank you. Any additional questions either here in the room or online? Okay, so all that remains to say is to thank everyone for joining us for the press conference today. As I said, this was press conference number six of our seven press conferences that we are holding this week. The last of our press conferences is tomorrow at 2 p.m. and is titled PC7, Life in Space, Habitability in Our Solar System and Beyond. So all that remains to say today is to ask you to join me in thanking our speakers for their excellent presentations. Thank you very much.