So um let me start again with um what I was mentioning. So imagine you're in a region where you have seasonal wildfires, category three h four or five hurricanes or you're living in, in, in a state like Florida or in a region prone to earthquakes.Now, when any of these events occur, what is the first and foremost concern in your mind? I'm guessing, but I'm thinking right that you, that you want to take yourself and your loved ones to safety as soon as possible and make sure that any of the urgent needs are met like medical or food supplies or shelter. Now wouldn't be trend of tremendous relief if you could get an early warning that there is a, an earthquake coming your way or if there's a hurricane uh in your, in your path or if there is a fire kind of winding away to your home. Um So it would be great to have that kind of warning. And um you know, uh and, and so you would take that information and go seek shelter or go uh seek medical attention if you need to and so on. So I will today go over two use cases which illustrate how using A I, you can get your loved ones to safety in an aftermath of an earthquake and also help rescue agencies respond to your needs of food, shelter, medical aid and provide relief to your family.

Um So um hello, I'm Nasrin Kalla. Uh I'm a data science P MA technical and chapter lead at Om Dana and I'm also a women in data science ambassador. I'm an enthusiast for A I for good. And it is my passion to empower youth, women and underrepresented in my community with A I technology A I can do so much for disaster management. It can map, it can analyze, it can model disaster zones, it can provide updated advisories, it can mobilize, help find people, ensure that the disaster response team knows what resources are needed and, and much, much more while A I exceeds at suggesting appropriate incident response actions, there is a drawback that is very challenging to accurately train the models to interpret a disaster, the more accurate A I can predict these disaster outcomes, the more it's used.

And with that improved accuracy, we can issue warnings ahead of time. We can take preventative measures, we can prevent damage, loss of human life and so many other things. And however, to make this kind of predictions a reality, we need to move people, we need more people like you working on that problem to advance the research and technology in the use for the use of A I in disaster management. So this is this is where we can all chip in and help. So um climate change is leading to more disaster is very evident from some recent uh um cases around the world, right? There were wildfires in California, in Australia, flooding in Germany and China. And why is this happening? Um there why are more and more disasters evident because there is climate change. And I think we all are very familiar with the topic. It is human and technology development, which is uh creating this climate change because we are pumping more greenhouse gasses into our atmosphere, disturbing that delicate balance uh of uh of temperatures on the earth and, and ever so minutely increasing its surface temperature.

And it is it is a fact that even a rise of a half a degree of the surface temperature of the earth can lead to severe and frequent natural disasters and that's what's what's going on. And it's not unless we put, get this in control, we are going to see more and more um more and more of these events and more and more severe and this will lead to larger negative impact on human life and nature. So at times, um we are not prepared for these calamities. Um You may have noticed uh this recent pandemic as well as uh hurricane Katrina a few years ago. Uh that indicates uh and it has revealed to us, a lack of disaster preparation on our side and, and disasters are not um not cheap. Uh pandemics are not cheap. In addition to the loss of human life and hardship, they cost billions of dollars. There are many, uh this is just in 2021. These are the billion dollar uh disasters that occurred in the United States alone. So, um it's just a, a visually shocking image to see. Um So uh a question to ask ourselves is, are we prepared to handle more severe and frequent disasters? Uh wildfires in California, flooding in, in, in Europe and the recent COVID pandemic has revealed a lack of disaster preparedness on a global scale.

Um Disasters also don't, don't have an equal effect on everyone. Um uh especially marginalized communities, Children, women, elderly people, people with disabilities, es and and especially in lower income countries are often disproportionately affected by disasters. While these statements are alarming, many countries have made efforts to mitigate and better predict and respond to these natural disasters. How do they do that by prioritizing resilient infrastructure in cities? So they're building stronger foundations and cities um using Nature Based Solutions which is fighting nature with the nature uh protecting livelihoods by studying cyclones, uh snowstorms, uh heat waves and early warning systems. So there is a huge opportunity for A I to help in disaster management.

And in the last few years A I has become even more powerful by because of the availability of more computing power, availability of data and innovations such as GP US that process these A I algorithms very efficiently. So in this big, in this era of big data and artificial intelligence, we have new tools to protect our society and manage the damage of such natural disasters and pandemics. Um Just a few examples. Um So there is opportunity for A I to help in disaster resilience and it's vast. Uh There are many, many things we can do, it can reduce the time it takes to assess damage to buildings such as schools from weeks to hours. Imagine that kind of powerful information, you know that there's a damage to a school, you can immediately respond and have rescue teams out there immediately in in within a few hours. Uh monitoring social media such as Facebook and tweets. Um There have been examples of that to guide relief efforts quickly and ensure better evacuations, uh accelerating the delivery of aid to millions of people uh by using geospatial weather and previous disaster data to predict how many people will be displaced by a hurricane or an earthquake and where will they most likely migrate to, right for shelters, parks or or certain areas and how much aid like food, water and medical uh medical care they will need and where to send it.

While there are many challenges to overcome. There is a bright future and a huge opportunity for A I and disaster management and it's clearly within our reach with the right level of coordinations, partnership and, and, and activity from all of us. So let's get to the two use cases. I'm gonna talk about today. The first one is A I and earthquake response. Um uh They addressed the SDG sustainable Development goal. The previous speaker uh alluded to that which is set by the UN for 2030 that we have to meet these goals uh as a, as a, as a, as a global uh team, let's say. Uh and uh the disaster reduction relates to target 13.1 and uh which is to make more resilient cities and, and, and be more resilient to climate change hazards. And then um STG 11 which is to make cities and human settlements more inclusive, safe and resilient and sustainable.

So the first use case was with a I Wonder girls team, a global team of uh women data scientists uh for which they won the Aws hackathon um award. And the second one was with the um Dana Silicon Valley chapter which is devising safe paths in the aftermath of an earthquake. Um So going to the first um uh use case. So the focus of this project is on disaster response and earthquake is a disaster type that we cannot control or even have the capability to mitigate. It happens regularly. It happens all across the world and uh does not discriminate of course and affects millions of people, we can only build more resiliency and be prepared to minimize its effects. So when an earthquake strikes humanitarian agencies and local communities, they need to quickly deliver assistance to affected populations through relief packages that include food, shelter items and maybe some medicines and essential daily items. So our disaster response assistant helps relief agencies estimate how much food and non food items the uh are necessary for an earthquake emergency. Um So this was the pipeline.

Um So it of course starts with data collection and preprocessing phase which is followed by which was followed by training of the machine learning models and Amazon sage maker. Um And this was used to estimate the earthquake um the affected populations in an earthquake. Using this information, we calculated the number of emergency relief supplies that we would need. Um There are also we also added on some N LP modules and natural language processing modules for Chatbot assistance and a news module. Uh The um the app also includes an earthquake related analytics module. So, through continuous integration with our open source, um this is an open source github repository. Uh all these modules are then deployed to a web uh streamlet app. Uh So this entire pipeline is an end to end A I BASED solution. Maybe we are leveraging data from reliable sources um to enhance the disaster response. And the application is web and mobile based for easy access for most users. You know, all of us have a, have either a laptop or a mobile with us and it can be localized to suit any country and the app can expand to other disasters as well such as hurricanes and floods. So there is a potential to expand it. So the data for the model was aggregated from established databases like us, Geological Society, em dat, uh World Bank and so on. Uh what was what uh then merging, we merged information about earthquake disaster management, geolocation, socio-economic.

Uh um So uh factors for various regions and then we used uh desires to management guidelines given by the United Nations for calculating the relief packages. So after merging and cleaning all the relevant data, we tried some machine learning models to predict the affected populations and some of the factors we considered in the models were the location of the earthquake, the population density that was affected, the soil characteristics, the depth of the earthquake and, and several other socio-economic factors.

So the app, I'm gonna go over the app right now. Um So the app was built using an open source framework called Streamline and we have several do drop down menus as you can see on the left. And the, the, the first one I'm gonna go through is the earthquake information tab. It provides general information on earthquakes based on data from, from us GS Geological Society and various other disaster management databases. The user can explore the number of earthquakes by country by year, you can visualize the earthquakes by the magnitudes and selecting even the years. I think we, we built a sliding bar um to do that. So we can see uh in this example that I've captured, we see a quakes from 1983 to 2014 and, and with a magnitude between six and 6.4. Uh on the diagram on your right and um the second larger tab on your right hand side that, that helps the relief agencies estimate the amount of food and non food supplies based on the country, the magnitude of the earthquake and the total number of affected casualties. So we estimated the total number of hectic casualties based on an XG boost regress. So I think that was the final model. We used to estimate the total number of affected population in an earthquake.

And then we input that into, into this app and uh using the guidelines of um from the United Nations, um food and nutrition for emergencies, we came up with uh with a certain numbers. So for example, in, in this case, I've taken an example of Nepal with a magnitude six earthquake and say, let's say an example of affected population being about 10,000. So it gives you an estimate of number of food and relief items that, that an agency would need and then they would take this, this number, um and, and dispatch it to the right region for uh for relief purposes. The next part of the disaster response assistant was a chatbot help. Uh Again, it was trained on N LP models based on curated lists of reliable sources about first aid and an earthquake. Now you can extend it to other um this kind of kind of assistance. Also, we chose first aid and and earthquake information during uh for for in earthquake emergencies. Uh You can also get news, you know, you want to know what's the latest, what's going on around you by selecting a date range. So you can pick a date range that you're interested in the next um day or two or next five days, what has been going on?

You can select a date range and then the number of articles of newspapers that you wanna, you want the aggregated news from or the number of tweets, you can also access social media and, and get that information. So this is uh the first part of our um of the use case. And um I'd like to acknowledge the A I wanna girls team. It's, as I mentioned, it was an all women's uh global team with a background in data science and A I and who are very motivated and very passionate about building solutions for social impact. So uh I've also input in the info about their github, our github as well as on de Post uh where we entered the hackathon. We have a lot more information, there's a video, there is uh all other links that that you might want to take a look at. So the next use case um is uh by the S um Dana Silicon Valley Chapter and this is devising safe paths in the aftermath of an earthquake. What we do in this um here is to use A I to map analyze and model risks, um uh model risk zones to provide safe travel paths.

This helps citizens to move to safer locations in, in, in an aftermath of an earthquake or during an earthquake, ensuring that um and also uh give, you know, allow disaster response team to know where, where to find people, right? If they've gone to some shelters, they know where to go and find them and where to move them from um from one location to another. Um So uh what um so the San Andrea's fault uh crosses California from top to bottom and is responsible for many earthquakes in California in the United States. We chose a small county uh L A county in the, in, in the state of California as an example region to illustrate this, this use case. So we considered three neighborhoods in the San Fernando Valley in, in L A county, Northridge, uh Chatsworth and North Hills. And uh what people uh can travel to safe destinations uh such as parks or shelters or even custom destinations if you wanna go from A to B during our afternoon earthquake in this region considering um the risk path. So what we did was we built paths and these paths had five risk factors involved in it. And those risk factors uh were density of buildings.

So our density of the building from the road, the width of the road, the speed of the roads and the soil liquefaction factor, which kind of um uh is, is a proxy for the damage to buildings that it can do. So we considered these five factors for our paths uh for risk on our paths. So this is the safe paths pipeline. So using satellite images of the city, uh we extracted from them street and building masks using image segmentation models. Then next week, what we did was we made these five risk factor calculations on them. And then, and then uh the example that I've shown here in, in uh in, in stage three is for distance um uh transformation. So the distance transformation was done on the segment in mass which outputs an a a roster map and then each pixel um represents uh the distance from a nearby building or how open the space is. So the more open the space is the less risky that path would be right. So that's kind of the intuition behind it. So we calculated and combine of course other factors as well. I've just shown an example here and that's shown in, in, in, in, in stage four and then we um uh undertook several fi uh looking at several uh pathfinding approaches and gave it various objectives such as minimizing the path length, minimizing the time, minimizing combinations of time and path lengths as well.

And then lastly, we integrated the risk map with the paths to get you an optimum path to um to go to safety. So, um and then we deployed all this to a, a web and a mobile app as well. So in the interest of time, I'm not going to go through all the risk factors um that we consider I'm gonna just share one in detail. So how did we build the building risk factor? So the intuition behind this is that the higher the density of building on a given road or a neighborhood, it would pose a higher risk um to travel during an earthquake or aftermath of an earthquake. Uh We use the OS mnx, a Python package that lets you download geospatial data from openstreetmap. So what we, what was done was the neighbor was taken, it was divided into 20 by 20 grids and a total number from the grids is extracted using this omnx API. And then we apply the total building value on all the edges in that grid. So uh for the ONX package, the nodes and the edges are stored as a geodata frame. So what this allows you to do it, it allows you to manipulate the underlying graph and you can also include new values or new attributes on this graph for later use. So that's, that's a cool thing about it.

Um So as I said, I won't go through all of these, but uh these are available for you later on. So this is the for the road width, this is for the distance, this is for the distance transformation. And then finally, after building all the risk factors on, on the map, we combine the five risk factors on one map. And then um and then all the three neighborhoods uh were combined as well. Uh So next, what we did was since we have all these risk factors, we want to see how we are, how do we want to weigh these risk factors? Do we want to weigh, weigh one more over the other? So we looked at the correlation plot which you can see up there uh to see how we can weigh these factors. Uh For example, for here, you can see that there is a high um 0.84 value for uh maximum speed and growth road with. So they're highly correlated. So what we did was we give that that risk score half weighted um and so on. So similarly, we did uh waiting for other factors as well. Uh So finally, in the, in the right, but you see the red image uh over there, mostly red image over there. That's the graph of combined risk forces from all the three neighborhoods. So the blue lines that you see in there are, are, are the safe paths and the darker red are the riskiest and then the colors in between are varying degrees of risks.

So uh what are the kind of path finding algorithms we explore? So we explore Dykstra, we explored a star search and hierarchical aar algorithms. So what we found was the hierarchical A star search was faster, but of course, the, the um uh as expected, but the risk and the length were not optimal. So what that algorithm does, what the hierarchical a store does is it prec calculates exit and entry points in, in, in blocks of regions. So let's say if you've given a large reason, it'll, it'll chop up the region into small blocks and then prec calculates the exit and entry points and then uses those when you specify a start and an end destination. So this may not be optimal uh for all sources and destination combinations, right. So, and we've also um in the bar charts, you can see the comparison of all the algorithms in terms of total time, uh the risks and as well as the total length. And what we found is the Aar and Dykstra are performing well and gave the least risk and, and the best lamp. So the next is the ri risk heat map uh with the three derived paths from the AAR algorithm uh given a start and an end destination.

So we give it a start and end destination. And then we shared um we use the AAR algorithm to come up with the, with the different paths. So the red one that is indicated here is the shortest path. The blue one is the safest path and the magenta is the one that is optimal. So it kind of reduces the path distance as well as uh manages the the risk score. So taking, if you take a closer look, you can see the numbers on, on, on the top right hand side is that the risk and the lengths that are associated with the optimal path are between the shortest and the safest path as, as you know, you would expect intuitively. So next thing is we deployed all this onto a streamlet uh web app. Um We give it a current location and we can search for the nearest shelter, we can search for the nearest pack or um get a custom destination based on our risk factors and, and a store algorithm. Uh Also we included a heat map with all the parks and shelters in, in the regions. And then it's, it's super imposed with the risk factors. So you know that it's risky to head to this park.

I want to go to a park which is a little more safer path and so on. So that was included as well. And then we also deployed this uh for uh for mobile applications where you can give it a current location and search for the nearest shelter or park. And then uh the technology used was a PQD five. So uh that's all I have. Um uh I wanna thank my uh chapter uh earthquake team who did a fantastic job on on the safe paths project. Um We have seen this deploy to two other countries such as Japan and Turkey as well and hope to many more. Uh So we're really excited to see this uh work spread. Uh The uh as I mentioned, all the source um that I've presented here is open, open source. It can be found on our github uh repositories, the disaster response assistance, uh details on Depot as well. Um And we also, I also have a detailed presentation on on this entire uh safe path which is much longer and goes through in detail about each and every aspect of it. So make sure you go visit and, and see if you're interested and if you want to follow my chapter on linkedin for um for interest in, in, in participating in future projects, feel free to contact me.

My contact information is there or if you have any questions I'm happy to answer um that as well. So thank you for your time. I know we went a little bit over. I appreciate you all being here and um yeah, have a great day, everyone. So yeah. Uh I don't know if I have any time for any questions but, um, uh, happy to answer if they are.