 Thank you. I want to take the opportunity to congratulate organizers in particular Carlos because it's been a great conference and thank you everyone for being here. So now I'm going to talk about this project about estimating the human development index at very high resolution using satellite images. This is a project that we are doing with colleagues from different centers but this is mostly around the global policy lab in UC Berkeley. So probably all of you are familiar with the human development index but I think it's important to remind you about what this is about where it's a multidimensional indicator of capabilities. It captures the health measures using life expectancy at birth. It captures education measuring the expected years of schooling that is associated with achievements from kids up also mean years of schooling reflecting the education level of adults and also a decent standard of living which is captured using the GNI per capita. So I'm not going to go into the definition but you just need to get the general idea and this is probably the main indicator to assess multidimensional development that we have is very popular and it was a sign to measure things at the country level. Now we release every year the country estimates and I have to stress that this is a very political thing. We are part of the UN and this index is used to eventually rank countries that are the member state, these are our bosses. So the reason why we are allowed to do that is because there is a general assembly resolution that gives us the mandate or the possibility of doing some independent research but we need to follow some certain rules and one of the main rules is that this has to be based on solid evidence and here the issue of comparability becomes very important. So it's very hard to have good data that is comparable across country. So this means that in general we have faced many years restrictions in order to advance towards to more designery level of data. There are many experiences of countries that they have the freedom to develop their own human development index but we don't have a comprehensive measure across country. Except for recently this paper by Iñaki Permanja and Smith and they but this is an independent project of researchers that were able to develop estimates of human development index at the province level for the entire world. So the question that we try to answer is whether we can use satellite images to make predictions about development indicators in a consistent way. Why this becomes important for us? First because now we have a lot of new information that is available that we can use and there is a lot of evidence that this can be done. And second and this is important for our purposes is that this data follows certain standards. So it gives us the opportunity to measure things in a way that cannot be questioned and also that will allow us to make progress in tracking what is happening with development at a very granular level. So this is a very broad agenda but here I'm going to concentrate only on the emphasis on the HDI and I have to emphasize that this is experimental and so far we are not sharing the data but I think it's just to open a conversation. This is not official and we can do many things and we are going to be talking today about downscaling. So what is downscaling is that eventually we have data at the country level and how we go from that to provincial level and how we go to that to municipal level. So just to give you an idea of how this works we have 193 member states but typically we publish information at least for this project we have data for 189 countries. In terms of provinces we are working with 1,700 units and if we go to municipalities we can eventually cover more than 60,000 units. So the methodology I will try to make things simple because actually the methodology is not very complicated. It takes in a very flexible way satellite data using daytime images. We also complement this with something that has been tried before quite a bit which is nighttime lights but the main innovation is here and what happens is that we have images that through a process that is based on some algorithms of random convolutional features that captures some of the principles of convolutional neural networks but in a much simpler way. So we transform an image in a vector of numbers. We don't know the meaning of those numbers but it's just a vector of 4,000 numbers for each image. So then we have information nested there and then we use that information in order to understand something. In this case the human development index that if for a certain area we have a value of the human development index we can train the model to understand what of the features that are coming from here are relevant to predict the human development index and we find the coefficients. In order to do that we cannot use a very simple or less regression which allows us to reduce in practice allows us to reduce the number of active coefficients. So now the results. So here it's going, this is an exercise that goes from province level to province level. So I mentioned that there is this paper by Permanger and Smith's that computes HDI at the level of provinces. So in this exercise what we do is to try to take some part of the sample like 80% and we run these regressions we learn and then we test this with the rest of the observations. And these are the results. So the feed is very good and the area square is 78%. So this gives us an idea of the potential for instance what happens if we have information for certain areas of the planet and not for the others so we can use that information that we have to learn how to identify levels of human development index and then predict that over the other areas. Now this exercise is about on scaling. So here is what happens if we have information only at the country level and we want to now try to find information at the province level. We can test this because we have the data from Permanger and Smith. So here what we do is first run a model that is a model in differences that captures the within country differences of the different provinces and then using the country level we recover the final number and the results are very impressive. So you can see so these are the final results and we have that the feed is excellent with an R square of 96%. So this is very good news because it tells us that eventually even if we don't have very segregated level and we just have aggregated level national level data and we have access to these numbers from satellites we can get pretty accurate predictions at this level, at the state level or province level which covers like 1700 areas in the plan. Now this is the result of this exercise. So in a way it's not very different from what we can find in the Permanger and Smith's paper but the interesting thing is that we were able to generate it just using country level data with satellite images. Now do you have a little water? No, water. No, I'm sorry. I have one mouth. So now but the question is can we move to the next level? From province level to municipality level the problem we have is that we don't have information of the HDI at the province level that we can use in order to actually test the validity of the exercise. However we have other variables for which we can identify the values at the municipal level. Now one of them is night light because precisely this is something that has been used as a proxy of development and we can see that if we do this exercise jumping from province level to municipal level the fit is very good and it's 78%. So it's not as good as in the other case but it is still very good. We also have information about the international well index sorry this is coming from mostly from DHS surveys and also we can using the same procedure find a very good fit with an R square of 0.75 which is again not as good as the 96% we found before but very good. So based on this we're encouraged to go to the next level using the human development index and this is what we find. So these are estimates of the human development index that jump from the province level to the municipal level using satellite data and these are over 60,000 small areas. So I think we're sorry very encouraged for these results now let's try to connect this with the so why we are doing this well it's going to be very important in order to improve policymaking so this is just an exercise of what we are gaining by doing this exercise here we have quintiles based on municipal level HDI and here we have the bars reflecting the previous classification the quintiles just taking into account provinces and we can see how we can uncover things that are hidden if we just use the province level exercise for instance here we have that 17% of the population are labeled using province level information in the 13th tile what in fact they are part of the top of the distribution and of course we can improve targeting in the other part of the distribution but the most important thing for us is that is the space that this research opens in order to understand what is happening in the world today and what is expected to happen tomorrow so many of the problems we are facing as part of the Anthropocene context are very difficult are very complex and we are trying to find information that can give us indication of what is going to happen because of climate change on human development in different parts of the world so we are going to be launching in the next few weeks a platform that is going to be giving us this information information about mortality information about what is going to happen to workers probably next year we are going to release information about what is happening with agriculture coming from climate change and all of this information is going to come in a very granular way so I think we are going to need some tools in order to make sense of this information and use it and cross this information with other variables so we think that in order to do that to go to the next level in this type of analysis it is critical that we can desegregate variables that we know are important in the analysis one of them is the human development index but I think for us this is just the start so I'm going to close with the conclusion so what we have shown is that there are promising results desegregating development variables in this case the human development index first it is possible to do this in a horizontal way province to province also it is in our view very promising the ability to have to downscale variables from natural level data which is what we have which is more common to the next level that is province level data and also we see that there is a very interesting space to go to the next level we are still doing more research on this to get that with validation but I think we are on track on that as I mentioned now we are trying to push this agenda to include more variables also to to have more clarity how we can validate to understand what works what doesn't work and also we are working in making these resources more accessible to researchers because I think this is very powerful and it would be important that can be used by in different parts of the world so hopefully we are going to be able to construct some platforms so things like that that would make this easy for other researchers to use and of course this is very important for a better formulation of policies and also for pushing the next generation of research so I'm going to stop here, thank you