 connection and we are up. Okay so welcome back everybody. It was my great pleasure to introduce Anne Jones. Now you see her, now you don't. So in our next series together that we run ICDP together as you know most of you know with the University of Trenton. So this week I've got great pleasure of introducing Anne Jones as I said and you can see already from a heading slide that she's now currently working at IBM. She was actually previously at the University of Liverpool where she did her PhD and worked for a number of years after her PhD and she was a great expert there on basically working on climate applications and so she worked in a lot of areas of applications including epidemiology and also hydrology which was actually one aspect of your past I didn't know actually. I obviously know a lot about your work on epidemiology and looking at climate impacts on disease and disease transmission risk and then Anne actually a while ago actually it was further back than I realised it shows you how time flies moved from the University of Liverpool to a new opportunity at IBM that she's going to be telling us about at the beginning of her talk which is basically a new endeavor a whole new project which has been set up during all of the chaotic period of the pandemic so I think they've done really well to get this going I'm really looking forward to hearing about it where in this project they're looking at ways of applying artificial intelligence to geospatial problems and Anne was just telling me that she's a sub theme leader in the aspect of that project that deals with basically AI applied to climate so with no further ado I'd just like to thank you again for joining us Anne and I will hand the floor over to you just to remind everybody before we start because I always forget if you wish to ask a question please post your question to me in the chat so that I know that you're in the queue and at the end I will invite you to unmute yourself to ask the question or you can specify that you want me to ask the question for you if you so wish so hand the floor to you thank you again. Thank you Adrian and yeah and thank you for the invitation to come and talk to you today so yeah my background is as as Adrian said is I'm my climate impact scientist by training and working for for many years on vector-borne disease modelling linked to climate and then about two and a half years ago I joined IBM research focusing on geospatial data science applied to industry problems and then recently for the past year I've been working on this new activity called future of climate so scoping this out and then now leading some research in this area so I'm going to talk about that today so just firstly to introduce IBM research so we are a community of around 3000 scientists across IBM researchers global labs we work on innovations in science technology from artificial intelligence to cloud computing quantum computing and we work on both new foundational exploratory science science that's that could impact IBM's clients and IBM's technology and also science applied to global challenges such as healthcare and climate change so over the past year we've been scoping out this new research program called future of climate and this has four areas of focus and I will talk in more detail today about the area I'm working on which is called AI for climate risk and impact just quickly the other areas that we have activity on are a sustainable hybrid cloud so this is carbon accounting and reducing carbon in data centres and we have activity on carbon footprint optimization focused on supply chains to reduce emissions we have some work on materials discovery for carbon capture as well and then we have the theme that I'm working on which is focusing on using AI in various ways to better quantify climate risk and also integrate that information into business processes to build resilience to climate change and so this is concerned with climate change adaptation and you know IBM has clients across multiple sectors from supply chain financial services energy and utilities who are all being impacted by climate change and need better insights and tools to build resilience to climate change here it also says accelerated discovery for climate impacts which means making scientific progress by developing new technologies and capabilities and doing this quickly and at scale so today I will talk about what we're doing first some background on what we're doing around AI and what we mean by climate impacts and then I'm going to talk about three specific examples focusing specifically on machine learning how machine learning can improve quantification of climate impacts and throughout these examples I go through the first one is more kind of straight out the box so some algorithms we can take ready and apply to this area and then the other examples are maybe less straightforward and more emerging in development so to start on this then from the basics what do we mean by artificial intelligence so in you know popular usage artificial intelligence just refers to the ability of a computer or machine to mimic the capabilities of the human mind so here we're using AI as a general term to encompass any computing technology which has these characteristics irrespective of the method used machine learning then is a subset of AI where the model governing the logic of the decision making or the concepts that I'll learn from data using variety of techniques including your own networks okay sorry my video disappeared then okay so deep learning models then are a subset of machine learning models where these models typically require less human input to train because the models themselves are learning the features of the data that are important for a particular application and you know the disadvantage then of this is for example that you can end up with a complex black box model where you're not clear exactly what it's learned and also that these these models can need a lot of data to train so today I'm going to focus on how machine learning and in some cases deep learning can help predict model and predict climate impacts okay so as a first step how can machine learning how is it relevant to climate change so of course data driven modeling has a long history of being applied to specific tasks in climate science but there's been a particular acceleration over the last couple of years due to you know the explosion in machine learning being used across commercial applications and academic applications as well and so there's been interest from the machine learning community and from the client science climate science community in how how we can use machine learning so this is an example this paper was from 2019 by David Roanick and some very high profile authors from the machine learning community looking at the whole of various issues around climate change and matching up different machine learning algorithms for a huge variety of different applications from forecasting energy supply to precision agriculture for example you know and they they they highlighted that it seemed like there were various different kind of algorithms that had potential to be applied across multiple areas so this is obviously there's great potential but to move on from this we need to understand then deeply how we can leverage this so we need to understand exactly what machine learning can do here and understanding from the climate side and what the requirements are and also for this to move beyond machine learning to AI we want to try and do this in some kind of scalable generalizable way to avoid the need for significant manual steps in collecting data and devising models for very specific application areas and situations you know which is a traditional way statistical models will be applied so that's that's that's the approach that we're taking in general so now I'd just like to focus on climate impacts as opposed to climate science in general and just sort of be clear on some definitions here so by climate impacts what I'm referring to are the effects or the consequences to natural or socioeconomic systems of climate change so this might be impacts to geophysical systems such as wildfire floods droughts or downstream impacts to human health or disruption to supply chains for example so specifically what I'm interested in is models that can quantify predict those impacts in order to build resilience to climate change I'm sorry I'm having difficulty with my video okay Adrian can you still hear me okay yes absolutely okay right I'll just keep going if the video keeps misbehaving okay right sorry about that right so yes so these models of climate impacts require climate drivers to be combined with non-climate data sets and the other term we hear a lot is risk so what do we mean by that so risk of climate for in the context of climate impacts here is a combination of the magnitude of the climate related hazards so from this could be something like flooding with vulnerability and exposure so by exposure this would mean something like where are the assets that you're interested in that could be at risk of flooding and vulnerability is what is the damage that can be caused to them by that particular hazard so in order to quantify all of this we need to combine this kind of information together so also to clarify on time scales the approach we're taking is to say that to address climate change adaptation and resilience we need to build better tools across prediction time scales so this may be you know from short-term extreme event forecasting flood forecasting linking that to applications like power outage prediction and through seasonal forecasting for vector-borne diseases droughts crops seasonal flood risk and wildfire risk right out to looking at decades ahead for more on the kind of policymaking side and and in terms of impact modeling often the same impact models can be used across multiple timescales because they're linked to and they're actually ingest weather information so we can use this the same the same models across multiple timescales which means we can we want to build scalable tools to do this we can potentially take advantage of that so specifically what we're doing in our IBM research program is looking across these time scales so IBM already has some capabilities at short time scales so the weather company is part of IBM their issue operational forecasts and some of that data is ingested by IBM applications which help IBM customers with short-term planning as a research focus we're particularly interested not only improving those predictions but extending them out to longer time scales and the research activities that we've got are along across this end-to-end model looking at improving the climate the weather and climate predictions themselves building impact models based on numerical simulations combined with machine learning and then taking that through to risk analytics and resilience planning for specific industry applications so next I'd just like to highlight some particular challenges with impact modeling and then see if we can link those to solutions and improvements that machine learning type models can offer so some technical challenges modeling climate impacts include quantifying the climate drivers so often the drivers of these impacts are things like extreme rainfall for example which may not be particularly well represented in forecasts that we have to have good data coming into our models we need to translate often large-scale global datasets into local information because the impacts are experienced locally they need to be locally accurate and precise we're often faced with a huge variety and size of different relevant datasets which we need to consume if we have numerical simulation models we're faced with a challenge of these models being computationally intensive and expensive to run and then often these systems in reality consist of a lot of different interactions and there may be compound effects so we may want to understand the probability of different hazards occurring so different types of flooding occurring at the same time and they are actually linked and are independent so this system is getting larger in terms of what we actually need to try and model. Another big issue is uncertainty so as we move through the modeling pipeline of taking data and adding more datasets and adding more models to get to some impacts every step of this is adding another layer of uncertainty and we ideally want to somehow quantify this in our predictions and include these uncertainties to make sure they're represented adequately and then finally want to do robust validation and this can be a challenge if we don't have we may not have particularly extensive continuous set of validation data available to do that so there's a huge array of challenges that we're facing here so then what I've done is had a look at for some of these how can we maybe address these using AI techniques of machine learning so these three at the top are the examples that I'm going to show today so the first one is concerned with getting better validation data for our models which we can do using satellite data for example so I'll show an example for that we then if we have impact models which are based on numerical simulations we want to do some calibration of that and do uncertainty quantification uncertainty propagation in order to address the computational intensity of complex simulation models we can try creating emulators of simulations using machine learning so this is a statistical machine learning model of a simulation model which then is much cheaper to run and finally although this is not really in the scope of what I want to talk about today I just want to highlight that all of these things are improved by having better tools to do our modeling so we need a modeling platform that helps us do all this and on that note I always wanted to mention that the way that we're doing this in IBM research is using IBM pairs which is a geospatial data platform that ingests lot climate data and other relevant geospatial data sets so this is actually what we're using as the foundation of the models that we're building and what pairs does that helps us is it just makes it a lot easier to deal with a variety of different data sets that we need to ingest into our models different projections different resolutions so it makes that process much easier so that we can focus on the modeling okay so now I'd like to go on to show some examples for these examples focusing around flooding so we are developing different types of different hazard models flooding is the number one for us and is what we've been focusing on first just because it comes out top across no different industry domains and different societal you know risks and also different locations and so that has it been important that we address that as the first one and we're considering three types of flooding so flooding from surface water flooding from rivers and coastal flooding as well so to take an example in order to model surface water flooding here we're initially focusing on background risk so climatological risk from flooding but also looking at seasonal forecasting of flood risk as well and we can use the same approach to create the flood risk maps across those timescales just changing the data that is being used to drive the model and then we're using AI machine learning in various ways to improve the modeling so if we go into one example ground truth detection so numerical simulation models like flood models contain parameterizations of physical processes so the setting of a model usually requires some tuning of those parameters to match observations and of course validation of the model's ability to simulate past events also requires observations as well so for the particular case of flooding recorded extents of flood polygons are available from a number of different sources for past impactful flood events but they're not always in a consistent format and the coverage is patchy it's not always clear for example exact time step time stamp that's corresponded to the flooded area and the method that was used to create it if all flooded areas were recorded for example or just ones that were particularly impactful so we found in developing these models that we need a more consistent way to generate that ground truth data to use alongside our models so this is an example fortunately where machine learning can really help us in a form of being able to take advantage of earth observation data so Sentinel-1 satellite data has become the go-to source for detecting floods and whereas Sentinel-2 data is optical imagery which just can be interrupted by cloud cover Sentinel-1 uses synthetic aperture radar so that's weather independent and it detects water essentially by detecting the difference in the roughness of the surface between water covered areas and other types of land cover so they're starting to put on our work here to sort of generate for the sake of our modeling want to be able to generate on-demand ground truth for flooding is this paper published by Bonneville and offers on the Sentinel-1 flood 11 data set so this in addition to an application of machine learning to detect floods also provides a data set which we can use to train more advanced models so what we're doing here is under the category of computer vision so computer vision problems consist of a variety of different types from classifying the classifying images the contents of images to detecting particular features within images to semantic segmentation which is labelling an image by a category and then instant segmentation which would be detecting different instances of a category within an image so within this established kind of recognized field of computer vision flooding flood detection fits in neatly as image segmentation so what we want to do is take a satellite image and label it is this for each pixel is this water or not water so the approach that we're taking here is to use a convolutional neural network and conventionally so CNN's convolutional neural networks have three types of layers and the convolutional layer which extracts image features such as edges and corners and then a pooling layer which does dimensionality reduction on that you can then have multiple sets of those layers which allow you to progressively extract more and more complex features so going from edges to faces and full objects as you stack up the layers so finally once you've done that you then have a final fully let's call the fully connected layer which does the final classification that gives us for each pixel on the image a category whether this in this case corresponds to flooding or not flooding so in this case we're using a CNN architecture called UNET so this was developed actually in medical imaging and was designed to address the complex subtle features that are present in in medical images so this was something that we could directly take and apply to to flood detection so here are some examples of what this looks like so this is the UNET model trained on the CNN on flood 11 dataset and then used to detect flood water for some events that were interested in so on the left we have an example of a flooding event from heavy rainfall occurred in Orden France in October 2018 so this is a particular interest for testing our models because we have high resolution rainfall data available for the same time and we also have flood polygons available from Copernicus so we can test our whole modeling pipeline so on the left you can see in the light areas water that was detected by the AI algorithm before the event occurred and then on the right you can see that flooded areas detected overlaid with the the the polygons from Copernicus and the accuracy here is very high 97.8 so this is an example of the model the flood detection model working very well and detecting the flooded areas so of course the model doesn't always work well this is just a starting point so on the right as an example of when it doesn't work so well this is for flooding in the aftermath of Hurricane Harvey in Texas in 2017 this is an urban area and in this case the model is not able to distinguish between water and highways so we're not successfully retrieving the flooded areas here so the next step for this work is to then try and bring in additional geospatial datasets to improve performance in situations like this where the baseline model doesn't work so well okay so this next example that I'd like to move on to is concerned with model calibration and uncertainty quantification so in the context of surface water flooding for example we're aiming to predict water levels by a simulation that takes rainfall as input and along with other geospatial datasets so the model has two different components that need parameters to be tuned so it has soil infiltration which depends on the soil properties for given location and overland flow which depends on the surface roughness so we take this information on the categories of soil and land surface from geospatial data in pairs but then these are passed to the model via some parameters so these need to be tuned to match past flood events and to get the most accurate model for a different location but because there's a large number of parameters to tune this is challenging to do so really we want an automated process that will enable us to efficiently search parameter space and get not only the optimal parameters but actually some sense of parameter uncertainty as well which we can then propagate through to our predictions so one approach that's particularly suitable for this is Bayesian optimization so Bayesian optimization is actually used in machine learning so it's a method that can be used to help fit machine learning models to tune their hyper parameters but in itself is also machine learning and it can be applied to all sorts of other situations as well so what this method does is it learns a representation of parameter space so that it learns the representation of the simulation model's parameter space usually using a process called Gaussian process regression so it's essentially learning a surrogate of the simulation model a statistical model of the simulation model and the Gaussian process is chosen because it's a particularly flexible way to represent any function so the method works by building up this model of parameter space that's then used to guide the search process to try and find the best parameters so it's a bit like the machine learning model is building up a mental model of that parameter space and using that to choose which next set of parameters to test and the way it does that is called an using this acquisition function so an example of this is the expected improvement so it's the expected improvement based on the Gaussian process model of the model that tells us how the improvement that we think we will get by testing that the next candidate set of parameters and so Bayesian optimization is said to balance exploration so exploring the uncertainty reducing the uncertainty around the parameter values and exploitation which is narrowing down on the particular refining the best current solution and the advantage of doing it this way is that it converges quickly compared to conventional sampling methods and another key advantage for us for interested and uncertainty quantification is that this is inherently represented in the Gaussian process so for example we can use this to derive a set of best parameter samples that correspond to a specific range in the target that we've calibrated against so how does this look like what's like an action so this is an example for another flood event in Palgar in India so this is one of the case studies that we've got from July 2018 and because this is early work this is actually a synthetic example so we're not here calibrating to flood detected by the AI model because we have to connect that part in here we're just calibrated to a single simulated snapshot to test the algorithm to test if it can find the correct parameter set so calibration then is performed over the whole grid matching the mean flood depth for a single time stamp so that's the similar to the situation where you might just have one single flood event that you're wanting to calibrate against and the time series here just shows the examples of how the calibration process iterating for a one specific grid point one specific location so you can see as the calibration proceeds the model predictions in black start to tune in to match the ground truth target in orange till it matches very very well of course because this was just an artificial example and then the output of that is a set of parameter distributions which we can then take and use for our model when we're making predictions so here's an example of that once we have a calibrated model then we have this parameter set which we can propagate so that means that we run our simulation model multiple times with these different parameter sets that have been chosen to represent the uncertainty in the parameter space so this is just an example for the same location of time series running over 10 years so there's the rainfall time series for 10 years and then the simulated floods for each of those years as well and the two lines that you see for each of for each of those years correspond to the parameter uncertainty so actually what we're seeing in fact is the default parameters and the calibrated parameters propagated through for all of those years and then it's interesting to look at how the uncertainty across the parameter values compares to the uncertainty and the difference that we get from year to year and we can then take that ensemble that we've generated for multiple years and multiple parameter settings to generate a risk map by considering the probability of floods for example above a given threshold by counting the number of the ensemble members above that threshold so that's this is an example on the right of what that might look like and then we can take exactly the same approach with a seasonal forecast so we can use a seasonal forecast of rainfall to drive the model and combine that with some parameter uncertainty to provide seasonal flood risk maps so the final example I'd like to show you is on simulation emulators so complex models like the flood model that I showed before are expensive to run so that means that we're limited in the geographic the size of the geographic domain that we can run for spatial resolution and the number of ensemble members that we can use quantify uncertainty so one area that's active in research at the moment across multiple scientific and scientific disciplines is exploring how these complex numerical simulation models could be entirely replaced by a machine learning model which is trained on the simulation data set so this is called a simulation an emulator or a surrogate model so it's worth noting also you don't actually need to have the simulation model to do this if you just have the output of the model you can potentially train an emulator which is another advantage so impact models need to both represent complex systems and also consume multiple drivers and depend on multiple different types of data so our work on using emulators and impact modelling is still in its emerging stage so I just wanted to show a few illustrative examples here of what's possible but we don't have a complete emulator of that flood model yet so this is an example from my colleagues in research who developed an emulator of water circulation in lake george in new york state so here they were emulating simulations from a 3d hydrodynamic model of water circulation driven by winds and the way they did this is to take the simulation output from the model to convert that into snapshots at different time periods this was then decomposed into spatial modes and each using pca and each of these spatial modes then had a time series associated with it so this was a step that's done first then they trained a neural network to predict the future evolution of these time series and at the final step then in prediction mode is to take those predictions and then combine them back with each of the spatial modes to provide a prediction of the future state of the model and just to illustrate a kind of efficiency savings that you can obtain with this kind of technique so whereas the full simulation with a complex model would take one and a half hours that forecast then just takes two seconds so as I say in order to if we save our computational time like this it means we can then potentially look at well what else we can do we can run a higher resolution over larger areas and better explore uncertainty so another example from another colleague of mine at louisa is an example where this time the neural network is doing both the dimensionality reduction that the principal component analysis did in the previous example and also the dynamic evolution as well so this uses an example technique an algorithm called auto encoders is a neural network technique so this can perform intensive dimensionality reduction because unlike pca they're not limited to linear transformations and this production is achieved by the neural network by encoding inputs into a smaller latent space which is then decoded in to create the prediction the disadvantage of this approach is that the neural network is is capable of doing such complex transformations that you can overfit to the data which is why variational auto encoders were introduced to reduce this where they have a particular constraints on the the neurons within the latent space to make sure that they're independent so this particular example based on this model called signer is interesting I think for impact modeling because it takes this concept of variational auto encoders but adds another thing which is to say that maybe we don't need to completely replicate all of the simulations we don't need to represent the state of the system all the way through but what if we're just interested in certain things so for example for flood risk we might only be interested in maximum flood depth over a certain period rather than the flood depth every hour or maybe we're only interested in certain locations not everywhere so can we take advantage of that to kind of to to compress the model that's needed to to represent the system and a further feature to note here is I mentioned earlier that deep learning models can be difficult to understand so you end up with a black box where you're not quite sure what's happening inside it so the other question that Eloise was looking at with this work was to say can we make this more explainable and more understandable and in fact this signet model was created by physicists who were exploring if machine learning model could learn physical laws from experimental data so in this very simple example Eloise took this model and looked at applying it to the problem of prediction of maximum water height in a simplified domain so let's see if I can get that to play so this is just a regular domain that has an infinitely high infinite height levy here that the wave is breaking against and the shape of that structure is just controlled by two parameters and then she looked at whether the emulator was able to reproduce the dynamics of this system as she varied those parameters and then she looked at what if we just want to predict the maximum water height at certain location can we encode that using this signet approach so she was able to create an accurate emulator of this system and also do so in such a way that when she looked at this inside this latent space which ends up with just two neurons inside it to represent the dynamics of the system you can see them they're plotted here the behavior of the two neurons as the as the predictions are made she was able to get a structure out of that that represented the degrees of freedom of the system so there was some intuition then that this model had managed to capture the dynamics but as I said this is a simplified model you know so it remains to be seen if these kind of approaches can be successful for more complex real world impact models with high dimensionality external drivers for example okay so some conclusions on this and thinking about where this is heading next so I think hopefully I've illustrated that machine learning can help address some of the key challenges in modeling climate impacts and you know just to highlight that this is such an active area there's a huge growing range of AI software tools and algorithms and new applications being developed across so many different scientific and commercial domains it's a really valuable area to be involved with some challenges of using AI include the need to have this expertise to understand which for a given problem we need to find the appropriate solution and architecture and also inevitably doing this work we need to work with very large datasets so we need tools to help us do that some exciting future areas and machine learning I think are things like knowledge representation which I've not I've not talked about at all so this is moving beyond just this numerical prediction to understand and to build models of what we actually understand about a system explainable AI which I've touched on which is a scenario of active research at the moment to to get some insight from what the models see and then automating this more fully so that we can move from just machine learning to AI and take out some of the human tuning of what we're doing for a particular application and then finally just a note and what we've found is that we need to do careful planning and have many iterative discussions really to start to understand how we can translate what we're doing in research into operations and then on to delivering business and societal value what we're doing which is the real target of all of this in the end okay so I'd just like to mention if I could I think it looks like we have time if you're interested in collaboration with IBM we have multiple collaborations with universities there's the IBM global universities program which welcome to look up we have research projects with universities we also have internships so research and internships usually for PhD students around usually three months which are advertised on our careers website and then finally if you may have heard of the call for code which IBM is a key sponsor for so the call for code for this year is now open and the subject is climate change so please get involved in that if you're interested so thank you very much be happy to answer any questions and for you to get in touch with me at this address back to you Adrian thank you very much and there's one problem with online videos is you can't hear everybody raptiously at home to the end of your talk but thank you very much that was really really interesting so I'm waiting for a question in the chat but just to get things moving actually I had a couple of questions myself I want to ask about so one was coming back to this application to the flood forecasting and you mentioned at the moment that you haven't yet applied it to seasonal forecasts but you wanted to do that I was wondering when you apply to a seasonal forecast do you have to go through the whole training again because the forecast model that you're using to drive will have its own bias characteristics so do you need to then go through the whole learning process again to account for those so in a way you're kind of building a bias correction into the system by accounting for the characteristics of the seasonal forecast or do you just use the training that you've already done from for example observations or reanalysis and just apply it straight out of the box or maybe do it just a traditional kind of bias correction that you apply to the seasonal forecast model before you pass it through I was just wondering if you might perhaps talk a little bit about the strategy that you're thinking of following there when you move on towards forecasting yeah that's a good point so so what I didn't mention is which seasonal forecast we're intending to use so we have seasonal forecasts from the weather company which have already gone through a bias correction process so we're ideally going to be using forecasts that've already had those corrections applied I think we would prefer to do it that way because it eliminates something else from the interpretation stage of what's coming out of the model at the other side and of course if there's thresholds within the model like they would you know they're often are you actually need the input data to be realistic so you want to correct all that first before you drive the model and so we also have some work on the team that are focusing more on the climate side on one weather generation as well to to further kind of improve the inputs to the to the models right because you could actually apply the technique separately to the forecast models themselves we actually use these similar techniques to actually apply the bias correction rather than the kind of the old-fashioned quantile matching and so on where you're only really looking at low order moments you could actually use some of these techniques I presume to have a much higher order bias correction technique now when you're actually applying it to these forecasts is that right yeah I mean so I think we still need to do some experiments with a seasonal forecast and see and and see how it comes through and how it would affect the tuning that we do okay okay so and also with this training because I'm going to resume when you're looking at flood events I mean by definition you're looking at kind of quite low return times quite rare events is do you ever find an issue with seasonality that you for example the model is trained and you're focusing on floods that may be occurred at a certain time of year and then if you get something that pops up out of season then have you looked at the seasonality of the quality of the should we say the fits and the training that if is there any seasonality that comes into the should we say the performance assessment so I suppose what we're focusing on in terms of calibration of a flood model is oh you're sort of calibrating the physical process of responding to the rainfall that we get so we're not we're not really trying to correct any of anything else that's coming out of seasonality for example but it might it might affect the performance of the calibration um yeah so so we would for a particular location um look at specific flood events but of course then we're just limited to to those so it's quite a challenge when you don't have properly continuous data to do this okay I've got a question coming in now so let me but but but but but but but bear with me one second because it's just a period so I just find the here we go and I will pass the floor ask to unmute at least say you should be out on mute yeah yes okay so uh thank you for a nice presentation so I want to ask suppose if you are predicting or if you are evaluating the model for the flood variation so do you need to go to train your model by each factor which can affect your model like a eda probe temperature and eda topographic locations and the other factors do you need to train your model by using all of those factors or not yeah so just let me skip back to that so so in fact what we're doing is not is basing everything on a simulation a physical simulation so the simulation model itself ingests all of that information um about like about precipitation but also um deaspatial characteristics of the location um and then if we were going to create an emulator of this we would also take these data sets as input as well so we would we we would use that as training data everything that the simulation model consumes to simulate physical processes would would then go into an emulation of the model okay now we have a question from Sebastian hello uh thanks for the very nice presentation um I'd just like to know whether IBM is also working on the weather or climate models themselves or say only on the climate impact models so hydrological or flood models yeah so the weather company does some weather modeling um the focus for research activity on the climate side is on AI enhanced techniques so not running simulation models themselves but focusing on on AI methods um ingesting um outputs from um from models all right and so that would be actually weather impact models like um not the weather model itself but you post process the output or you put the output of the weather model into your um artificial intelligence model yeah so it might be trying to for example improve the um the accuracy of a seasonal forecast um using some seasonal forecast data but also some other um some other data as well so or improving the prediction of extreme rainfall for example using some observations and which will then have you know be something that we can consume with our impact models and produce better better inputs to the impact models okay thanks and just another question uh is there cooperation with universities or public research institutes yeah so we do have different collaborative projects um IBM research has a number number of different collaborative projects with research institutes so that's definitely something we do um and there's also IBM has academic licenses for various products so pairs for example um you know so so yeah that's definitely something we're we're actively interested in doing cool thanks okay I had a question that I was asked to ask you from Lorenzo Bertramme who asked if you could provide a little bit more detail about some of the agencies and organizations that provide your raw data so you mentioned Sentinel for example as for a satellite source he was just wanting to know perhaps a little bit more detail about some of the agencies that are providing data or maybe the networks within the projects that you already have established um yeah so we'd use a lot of open data sets basically so you know datasets available on Copernicus um we analysis datasets like era five these are all datasets that we're consuming and pulling into pairs um to you know to set up with our modeling pipelines okay um last uh I don't see any other questions in the list I just wanted to ask actually because obviously you've got this long extensive experience on the uh Vector Bourne disease uh will you be branching out into that area it's one of your applications so I mean I can see a lot of potential for some of these hydrology products in fact uh temporary flooding temporary water bodies you have this analysis at very high spatial resolutions um so is that an area that you think IBM might be interested to branch out into what drives the you can obviously see a very strong potential uh in the in the private sector for the the flood forecasting there's a lot of demand but what's kind of driving the uh should we say the end products that you're actually choosing within this this program so I mean what we're interested in as research is working on things that have the the significant impact so we we you know we need to work on projects that can have significant economic impacts but also societal impact as well so we've been interested in doing that if the you know those kind of um those kind of collaborations that we could make would facilitate it I would say that definitely all the tools that we're building as generic as possible so you know we can see all the commonalities across different predicting different hazards using impact models a lot of it is the same that you need to do you need you need to pull data you know and you need to do calibration you need so so wherever it's possible to build this in generic way that's what we're doing so the idea is then if um you know if there's some other model that we want to use within the framework that we've built we can put it in we can start using those AI tools with it quickly and that that that you know that's my vision for this um you know we can then address something new very quickly by just uh just getting the model in there and and and going with the data that's already um you know available and pairs ready to ready to run yeah I mean the the biggest bottom there always with the the applications data is that these time series are often quite short now if you look at something like a malaria data set we know from experience that you're lucky to have you know six seven eight years of good data and that must restrict some of the applications though they're just the lack of available data in the area of the sector that you're interested in is that right yeah I mean I would say that having worked in um malaria and vector ball and disease is good um it was a good background for this because it's most maybe one of the most difficult areas in terms of getting that ground truth data um but yeah we are limited we are limited by the the data that we can use yeah okay brilliant okay so um I think that's all our questions and we're we're at the one hour mark so it just remains for me to thank once again Anne Joans for a very interesting talk on the the applications of artificial intelligence in climate thanks once again for your time and I look forward to seeing everybody here again next week on next Thursday where we have another applications directed talk topic where of basically climate and health so I think this this material this week would lead very nicely into next week's talk topic so thank you once again and see you all soon bye bye thank you very much bye bye