 Okay then I started with this slide because I want to say thank you to Maria Laura. This is a seminar part of the Alumi Network as you can see there is presence from people from all over the world and yeah thank you Maria Laura for keeping up this important work for the Institute and for everyone. I think all of us as Alumi that studied the HR are very interested in keeping always up to date with different things. Just one fast question I just want to confirm that you are seeing my slides. Now what I have prepared for today is precisely this discussion about what is artificial intelligence, how artificial intelligence is making a change in water resources and I'm seeing from the perspective of hydrophromatics and hydrophromatics at IHE has been bringing this together in different ways. I have to acknowledge a lot of this work. Professor Dmitri Selomatin have been here for already for some years at IHE and he have been always pushing new technologies, implementing new areas, always something that can be applied and can be new. It's started, it's implemented, it's tested and therefore what you're going to see here goes like in that framework of ideas that have been developed at IHE. So where to start in all these things? This presentation is artificial intelligence and it could be interpreted as complex as if you go to the Stanford Encyclopedia of Philosophy Online. You will see that is the field devoted to building artificial animals or at least artificial creatures that in the suitable context appear to be like an animal. And for many artificial persons at least artificial creatures that is suitable context. So as you can see it's very very cumbersome definitions that well we really don't know how to understand that. But if you go into a simple not also technical description or maybe a proved one Wikipedia is mainly used for this type of finding information and descriptions and Wikipedia says that is intelligence demonstrated by machines then now it's more up to something that we can understand that machine that is an intelligence and unlike the natural intelligence displayed by animals and humans this is this is different is leading artificial intelligence nowadays talk about intelligent agents. And these agents are devices that receive its environment and take actions that maximize its chance of successfully achieving its goals. And this is this can be if you look at that sentence it can be just a simple sensor on the wall where you just pass and the light turns on you pass and the doors open you enter into a room and the temperature regulates you enter to a room and the light regulates according to the day such that you have a friendly environment. So these many small things can be interpreted as AI type of things and and this goes even with the concept of some people in electronics called Domotic that make intelligent houses. You clap your hands and the lights turn on or you say a voice and talk to some device and device opens the door close the doors and alarms and so on that's called intelligent houses. So indeed this definition is kind of another part of the definition it says the colloquial term artificial intelligence is often used to describe machines or computers that mimic cognitive functions that human associate with human mind and that's another point of view that in the same context computers are these machines right kind of develop analyze and provide some some responses in the context of what you are telling them. So artificial intelligence so far yeah it's kind of having some sort of machine that helps to respond in a certain way understanding what is around okay. Now where this this came from we can say that the computers came in from the first talks and the mind paper of 1950 from Alan Turing he argues the question kind of machine thing and six years after this question came from Alan Turing in this paper the concept of the term in Dormand conference in 1956 artificial intelligence born that was the first time many researchers start to talk about artificial intelligence. So from that perspective we have computers where it started the concept then programming on top of the computers then came internet was a change in paradigm and it started to still have a lot of things ongoing and the last part is artificial intelligence on top of all these things. So what I'm going to present in this following 20-25 minutes will cover the development and at the end I will show four examples of artificial intelligence now from the most modern concept point of view but I will show you different paradigms that have been growing and you are still probably people working with the results are working with some of these paradigms in day-to-day life. So the first thing was the computer the device and that device the computer every computer has some memory and you will see about the RAM memory when you're buying a laptop the memory long-term that your hard drive and you want to move it from one side to the other to your USB. You have every computer have a central processing unit so it will process all your information and every computer have a ALU which have Arimetic Logic Unit which will process the mathematical component of that and then you have peripherals that are the ones that allow you to interact with the machine. So kind of the computer already is kind of mimicking this concept of what is needed to be able to understand some artificial way to reproduce things. After we got the computer then we started to program on top of it so solve many tasks how can we make computer do more and more and more things and this was a boom from all type of areas of science then the main goal on this as you see here in the second part is automate or program activities. How can I reproduce a very complex problem that I solve in one place create a program and use it to solve another and another and another. The main principles of this automation in programming is has only two components the first one is just half a logical decision if you have some condition then do something that's the basic principle of programming and the second basic principle of solving is our loops so how can I do something many times iterated so for a condition do something until something happened. So you have these two components that allow you to do almost everything in program and that's I would say is the main let's say challenge to how to use these two components to build and build and build and build much more complex automation with a computer but that was the beginning and the beginning growth into programming whatever sources and this was I would say hydra informatics that programming and that idea of being able to use these tools who came with the numerical methods to solve for physical equations and we had in those days hydra informatics Michael Abbott IHE Delft in the Netherlands they started to think about how can we build models that can represent fluid dynamics from the basic equations so they started to move from what was before a conceptual model a model physical model into much more complex representations where we solve the fluid equations we can evaluate cross sections in a river see where they are going to inundated we can build hydrological models that flow that can show the flow that will happen in certain moment inundation areas water depth and so on so that's kind of an example on this part of programming but it didn't stop there and that when you finish that that concept on on programming you move into internet and the cloud and how internet and the cloud started well you had these machines we had these models you you just saw them in water resources at least these examples in water resources is how they can communicate how can one office that has one information in one place and another has another how the computers can communicate and then it moved we need to communicate faster so they started to create the networks and computers but then we need to share information and then the networks start to grow and grow capacity distance and so on and reach a level where the communication and information was not constrained was open so everyone could move and start to add the information and this ended in the new social world of digital citizens everyone has its profile everyone has its place where they are sharing what they are happening there are different types of of spaces where you can contribute and you can say what you what you're thinking what you're doing so that that created a kind of unexpected space in this technological development but as well there was a competition in one side from all the internet if you have a large technology that covers all the world you have a lot of information I want to search along it how to solve this problem of searching through hundreds of thousands of of data and that was the different internet searchers we used to be digital Alta Vista there was Yahoo there was a MSN hotmail still there are many many search engines but Google took the lead it seems that it was able to respond to what the people wanted if you want to know about something then the most accurate one was Google and then they say he became company became rich would be largest in the world in that moment only because of the search engine but to complete the story of what these have providers it end up in now this search engine became like a kind of a game a kind of a artificial intelligent space where you can nowadays ask what's the weather in any country in Latin America what is the currency conversion right now what is happening in Australia so you right now you can even say how can I cook something what where can I find a place to replace my things so as you can see this starts to become a kind of intelligent responses to everything that you want and it's only based on searching through databases only looking what is there is nothing really intelligent behind but it's answering all the questions that you have that paradigm and what the resources end up in what we have right now on what the related data information about the weather sensors in real time that we can have right now the devices information about extreme events this is also an important thing because in the past something could be happening in Africa in America in America or whatever and no one knew about it or it was not so well informed but now everything is connected so it would appear that now you you can understand what more what is happening worldwide before was not possible and the concept of remote modeling what resources this was something very interesting in the year 2002 they started a program called hydro Europe and this I mentioned this because it's the concept where I started to see for first time remote modeling how can people access models complex systems through internet and hydro Europe in 2003 in fact this is my batch on the master of science in IHE I am there somewhere with some glasses and also professor Joana Popescu which was the leader on this program through Europe and there were four or five countries and what they wanted to do with this program is that we all in different countries were accessing the same model but each one working in some characteristics and problem and we had to build a project out of all these very complex information system and at the end we went all into France and East and we met the people we met the we went along the the the region understand the things that were happening but before that we were three months working online with remote systems so this was an interesting change in paradigm everyone started to think about how to do models and to be able to share them now reaching the the the area of more advanced things in water resources we can say that this end up in now we have more information there is a boom of how to deal with extreme events you you can see that due to the fact that the news about events are happening everywhere we we are not more aware so now we we want to deal with extremes we want to see how the other people are dealing with the streams we we want to be more informed and knowledgeable in the rest of risk how much risk right now I'm having right now we are more prone for experience like which type of models are they using what type of infrastructure have they implemented what sustainable development interventions can I do all this information is there now the last thing that you can see in this boom of the internet is that now you have social media in water if some expert in water wants to find a job that's water jobs if there are there are many many groups about risk risking extremes in floods in droughts in water management diplomacy and so on there are government awareness organizations social media even Facebook for for for cities for regions for awareness of an extreme events and so on but this is very interesting that it came into that but now that I move the last day I think I counted myself 10 minutes to reach here I don't know if I took 10 or less or more but this is artificial intelligence now and what you are seeing in these slides is the the concept on on on different specialized websites about how the artificial intelligence developed which basically is what I I said this last five ten minutes it's okay the concept that started but it was a little bit vague everyone was thinking about what can be and what is and there was always fairs of what can be reached then 1980s and 2010 there was machine learning moving and starting and from 2010 there was a shift in paradigm again on a competition an international competition on on any major pattern recognitions where an algorithm that used deep learning made a breakthrough it managed to improve the performance of the of the of the classifier to identify things with with an impressive accuracy and relatively very fast and the nowadays yeah you any one of you can install in five minutes this network of 23 layers and you can identify many objects there is a test where you put the camera and you put your cup your copycuff and it will tell you if that's a copycuff and it is really simple that's I encourage you to explore these type of things it's really simple and so on but okay what is the context of artificial intelligence and deep learning or machine learning I would say the machine learning in general it's more wider and that's why here I will talk of most of the things about machine learning it's the ability to learn without being explicitly programmed and that just that sentence makes a very important shift in paradigm so nowadays there are many people saying you don't need to learn how to program now you just put a machine learning that will program in between packets program the thing for you what's the meaning you have inputs and outputs and use in the past you had to learn what were all the process and you had to program all the processes to be able with one input obtain one output now not anymore now you start to think about oh hey I have this data from the input and I want to obtain these outputs every time that I have this combination I need to have this type of responses and then if you build a table that allow you to know what is happening in the action and what is the consequence you can build a neural network without having to program thousands of lines of code and that's that's one of the things that I want to show you here how it works so in this concept artificial intelligence is building from knowledge from mathematical equations it's a framework let's say now we are here talking in more detail on machine learning and it's built from different areas of science so terminology sometimes in different papers it starts to keep coming to be cumbersome so experts in finance talk about artificial intelligence and machine learning precisely in their own terms and they develop their own algorithms people that are working in image recognition another one people that are working in databases and big large databases talking in other contexts so that's why having all these areas of development requires some some time to be understood and know what is the best to do is not magic you cannot just select one thing and we'll solve the problem you need to really go a bit deep into it I don't know I somehow I lost the presentation I don't know if hello can you hear me yeah would you like me to bring your presentation back yes please sorry I somehow I press to change not this yep thank you yep do you see it now yes thank you okay now continue with the concept of machine learning and now I want to bring more more ideas and bring that's now down to earth so you understand what is machine learning and this will open a bit the scope for what you and Jose will present in a few minutes so what is machine learning so the machine learning is described that they use and development of computer systems that are able to follow in instructions okay and then here you have inputs and outputs on the on the right you have a real system and what you want to represent is that when something happened in the system you have some output that you are measuring you're observing and then you want your machine learning to learn from that experience and receive the same input know what is the output and he will produce an output and you need to differentiate the if the predicted output should be equal to your observed output and this is trying to learn is minimized the difference between both so how can you change the the parameters inside the machine learning data driven model to make the best prediction on on what would happen in real life and this this this slide is from the lecture from Dimitri Solomatin in IH in 2009 and where he most of this concept of machine learning has been explained as data driven models so you will find a lot of literature publications where machine learning has been described in this area of science as the data driven models whoops again somehow if I press a space it jumps here sorry Abraham I will try not to press a space okay now to to understand a bit more these systems and and how they became so so so how they work and why they start to become complex once in a while when you start to grow and grow the problem is this thing like a transition from one neuron what is one neuron in these concepts of machine learning and I'm giving the example of the one of the most well-known algorithms which is the MLP multilayer perceptron network so everyone that thinks about machine learning identifies it with neural networks although it's not like that there are many many types of networks and there are many type of algorithms that feel that are part of machine learning that are not neural networks but in this case I wanted to mention this at the core because well some of the things that is going to are going to present Jose and and some of you is are in this area of work so one neuron is simply like what you're seeing in the screen is like a kind of linear regression you have x1 x2 x3 multiplied by a weight 1 weight 2 weight 3 and if that combination of values of your input multiplied by a w w is any number that you need to fit such that your value of a will tell you that is one if it's one class and zero is if it's then the other class they call this activation so it activates and says for example it's above the threshold it's going to be a float is below the threshold then it's not a not a flood in normal machine learning or when you are looking in other areas of science it says oh that's a dog not a dog it's a cat not a cat you know it's it's black and white and this was more of the most simple way to see this activation functions the the next level is to see it in this two-dimensional so this regression in fact is aligned and what you are trying to find is this slope and the place where the line should go such that you classify all your points in according to what they are so you already know that the blue points on the right are one class and the crosses are another so you need to find the slope of this red line that's that slow is this double you and the bias parameter of course to displace it and if you find that that exactly divide both then you have a classified everything above the line is across below the line is a dot that's the first concept of one new which is this one but what if you combine that with another with another with another with another and with another and finally with another and you have what is called a multilayer so each one of these rows for these columns of of sorry it's one of these rows is one one layer and that's what they call this multilayer and this is this was one of the the beginnings of neural networks and they say that this is a universal approximator the challenge here is that you have so many nodes and every one of these lines that you see there is a weight and to be able to learn and that the learning is identifying those weights they will how can I find the weights that will perform the task of classifying or being able to let me know that certain combination of inputs will have some result and as you can see that's that's that's the challenge learning and learning is an optimization problem how can optimize the weights such that they will produce what I want so where these machine learnings are applied so you will find the experts in finance time series that works with things like this or even in voice recognition but in water resources we use it to to forecast or to simulate a railroad runoff options pattern recognition you can find it like this in other areas of science in water resources we try to look at remote sensing images to try to identify patterns of floods or patterns of crops or whatever in other science a large colliding hydrant in physics in water resources we have big data in atmospheric science every second in every place and there every minute every hour every model simulated in a scenario of climate change you have big data there in text mining we have in in general in internet Google systems and so on in in water resources well we are just right now starting to understand how can we convert this qualitative data into quantitative data now again going deeper where where is this AI in water resources how can and you can find many applications nowadays that use machine learning for simulation of events from forecasting rain or discharge for early warning systems and even building surrogate models that replace a operational system so like real-time controlled situations where you want them to reproduce or represent something that you want to achieve certain performance so this is this first paradigm in water resources completes only time series and there is most of the time no spatial information the second one is the image classification remote sensing so there are many algorithms and you go to Google earth and gene and you will find classification programs in remote sensing images mainly about land use flood classification and so on so how can we learn that some pixel in a survey was identified as a certain type of crop or certain type of grass or whatever then you go into the image and you look at the place where you did the survey and then you make your remote sensing image to learn this this result in every pixel and then you say okay now if now that I know that that combination of bands in my remote sensing image give this this this value of colors that colors in fact is a class which is a land use then I will use it for the whole image and then you reproduce what is happening so mainly this this type of problems are a spatial or geo reference information as you see they do not work in time it's only one layer or it's a it's a spatial representation and you want to identify what is happening everywhere and if a new image comes you also would like to try to explore the same use of things but it's not explicitly in time the third area of application of artificial intelligence is in decision-making and I would say that the last 10 years there have been approach of using text mining and recently I would say in the last two or three years there are people working with natural language processing that has advanced quite a lot to take information from social qualitative information this information is highly abstract and dynamic is changing every day the news the people tweeting Instagram and so on so this all this information how can you take this information to obtain okay you look at this information and obtain something that is useful for either for your models or for taking a decision and number four of the ones that I know I'm aware of all these some of these applications is the spatial temporal models so let's say agent models cellular automata models pattern tracking so how spatial temporal information of dynamic yeah nature can can be seen and can be used for replicating what is happening in real life so let's see the context these are the four examples which are with this four examples I will finish my presentation I hope not taking too much time because we started late and the first one is flow forecasting I mentioned it before if someone wants to make a full flow forecasting with physics you could go with a very complex physical based model most of the time flow forecasting systems work with conceptual simplified representations even with the statistical and this is one conceptual model and although it's one of the simplest models have many many parameters what has been with applications of neural networks is to try to say okay what are the inputs and outputs of that situation so or either you go to the conceptual model and you program every equation and say if this is happening then do this and if there is a saturation of the flow then do that and if there is a level of the groundwater then put some percolation and change it and put some runoff and so on so you can put all these type of complex equations or two say okay let's build a machine learning model in this case a neural network model what are the the previous time steps which are in fact the inputs of the normal conceptual model what is the PT precipitation in time minus one minus two minus three so what's the lag precipitation what is your lag discharge and then what is the the the output of that what is the discharge at the next time step or what is the discharge at the following forecast horizon and this is how the network learns learns the the real system and then when you use the real an operation it will reproduce this limitations is that it is fixed time step so whatever you give him is what he's learning he and most of the time you read only the the measurements from your device that makes it well there is no no knowledge representation and is fixed on on your delta T in a conceptual model or a physical based model you are representing the physics so you can play with time in in this type of models not a second example is the optimal reservoir operation and this is a Dominican Republic the the MSc research of Carlos Tommy from Escuela Colombian engineering he just finished in fact this week and was very interesting okay so that's why I am bringing here this is in the river tuna and that was a war at the which is one of the most important in this river basin and it has a there have been many many floods in this region the reservoir is meant for producing energy it also helps on on the irrigation district and also has been discussed oh sorry again space I cannot hear you somehow is mute about now yes now I can hear you yes okay I just wanted to ask you to please start pointing up they are running a bit behind yeah it's only two slides more yeah okay thank you that's sorry for I was thinking in this 30 minutes but yes thank you okay so this is the case study of the tuna river and let's imagine that in fact this is not good but okay let's go backwards let's imagine that you want that we have the last 12 years of of of operation of this reservoir and there have been many floods and there have been many situations where there was not enough water for the for the crops in the region and it was in fact operated trying to to satisfy everyone but mainly energy production and that if we look in detail why we don't try to learn from what they have and this is a relatively recent concept you learn on on the past you built a model that represents what is happening but after you learn that model then you say okay now I have a representation and I want to in the future to solve the the challenges of the of the system I need an optimal energy production and I need a reduction in the terms of flood situation and I need enough water for microbes so here we build a neural network model where the three objectives this was first was built to reproduce what was there happening to reproduce what you can see in the blue line the rainfall is on the upper part that's an average frame of the of the system and the lower line is the green that is the energy production and the artificial intelligence there were many many models built and finally we got one that managed to reduce the the floods and you see the blue and the red line on the on the middle graph and all the events yeah maybe 80 percent of the events were were really solved by making an intelligent machine learning to take the decisions and this this was built on past experience and then optimizing what will be the maximum reduction that you can have in flood with the maximum production of energy with the maximum satisfaction of of agricultural irrigation okay and then this is third example as I said it's only four examples of how I'm almost going to finish sorry just mentioned the PD studies of vitality as he's the one that working in these special droughts it's a case study in eastern India this region is very famous for the production of rice and there is a situation where if you have a drought in this region well you will impact in your crop yield but the region is so big that to be able to make a survey and to analyze the impact it will take you like two three months to be able to to understand how much was the impact so we look at past situations and how how much the crop yield was impacted and we made it learn and by looking at the spatial temporal variations of the of the droughts there is an animation here that I see it's not working okay what is supposed to happen is that these areas starts to change by learning the changes in the areas we could determine how much drought how much crop yield will be affected now the this work was extended and what you can see here is the whole India and what we are finding the variations in the space and time of each one of the events along a certain number of years and this was 15 years and then we look at how the the patterns and we made a tracking of all the large events in the lower part that you can see the vectors that show how are the the droughts appearing and by doing this you can identify when a drought is going to happen how it's going to happen what is the synchronicity of the drought so this is very very interesting and this is all looking at the pattern tracking emotions and the last one is the example from the news social media this is also very very new it was in a competition of artificial intelligence in Colombia with 22 universities and the the concept is how to to bring the qualitative information from the network and what was proposed by Santiago Duarte the student from Colombia was to create a system that was scrapping we navigated through the internet started to extract information from all the newspapers more than 4090 news articles and after extracting he passes processes with a natural processing language extracting intelligently the the information such that you can have an index that will tell you that is good or bad and he created this app that you can see in the address HTTP Santiago Duarte 09 users earthengine.app and what is going to show you something like this this is one of the six apps that you will see there and what it will show you is the feelings of the people and how this has an index of the help of the rivers how how they are feeling that the river is in a bad condition a good condition since this analysis is done historically it was mapped with decisions made by people in the regions and we were able to see what wrong decisions led to decay in the information that was from the news from the people and from the social media as well when there was success in a decision we also felt the feeling of the of the of the people around the system and that's it that's my presentation so that's so what you think will be the next AI application in water resources what it is just to spark ideas and I hope the ones of Theo and Jose can complement and they for sure will go deeper in many things