My name is Thalma Chakravarthy. I'm from Bangladesh. With me I have Hasan Mustafa from Pakistan and Jordan Bates from the United States. This is one of the projects that we did in our class. We are all three students from the Erasmus Mundus Scholarship Program of the Masters in Geospatial Technologies.

This is one of the projects that we did completely based off open source data and open source tools to see the functionalities that we can use in real time and to combat some of the issues that are going on around us. Since I'm from Bangladesh and I know a bit about the crisis that is going on, it's the Rohingya crisis.

The Rohingyas are actually the minorities who are from Myanmar and they are exiled from their country because of various reasons. I don't want to get into that. Bangladesh, my country, has decided to take them in. Now the problem is that when you decide to take in refugees you need to have a plan prepared.

You need to be able to help them so that another new crisis does not become born of that. So this is our topic where we are trying to see the settlement trends and how if we can predict the settlement trends it will be a better opportunity for them to settle down and have a better life.

So these are the table of contents. We have the study area, we'll give you an overview of the situation, the goals, our problems that we had for data acquisition, the methodologies, the tools that we have used, classification, accuracy assessment, and last of all, but not least, the settlement trends and the futures that we are trying to save.

So just as a glance, this is one of the very small camps in Cox's Bazar in Bangladesh. So as you can see, it's quite crowded out there. So this is the study area in the southeast corners of Bangladesh. There is a place called Cox's Bazar and in that area there is a lot of areas.

So one of the main camps, the largest camps, it's called the Kuthu Palang and Balukhali area. So since it's the largest refugee camp, we focused our attention there. And just to show you guys an overview of how the situation has actually progressed, this is an image before August 2017.

And the next image that I'm going to show you is just three months later. Please concentrate on the small bubble that you're seeing in the middle. This is how it has expanded. Yes, that is the reaction I was hoping for. So we do need to come up with a plan or an action plan to combat such things and to do that we need data.

So we have a lot of data with us, but we'll talk about this a bit later. So as I was saying for the last three months, from August 2017 till after that in September, there have been a lot of people who have came. A total of about 700,000 people have just migrated into the country and they are looking for a place to stay.

But the land is not educated enough, is not stabilized enough for them to make settlements. So they are just staying in camps and Red Cross, UN, UNHCR, OCHA are all trying to gather in Pichin and trying to help. So our goal for this is to find out the expansion since January 2017.

The main expansion starts from August. Look for the settlement patterns and then consider the future expansions. So we use satellite images. Now whenever you work with satellite images, you have some kind of constraints that we need to address. So these are the four constraints that we have to address. First is the resolution.

All available data are not like five meter resolution data, so you cannot work very much with them. The best you can get is like 10 meter, 20 meter resolutions. There is also a factoring of clouds. So if you have clouds over your study area while you are trying to focus and you cannot get the image, the classification is done properly. Availability is related to price and the timing.

So we use Sentinel-2A and 2B satellites, which are a constellation of these two satellites. They have a 20 meter spatial resolution and a five day temporal resolution. And they are free. So these are the band selections that we worked with. We worked with bands 2, 3, 4, 8, 11, and 12. And we combined them into various classes so that we could get the natural color, the false color, and the urban false color.

So with the help of those, we are able to classify. And now Hasan will talk more about the algorithms that we used. Thank you. So for the algorithms, we did a bit of literature view. And we went through four of the main algorithms that are used for these kind of projects.

The first one was random forest, the neural networks, the support vector machines, and maximum likelihood. Random forest was the one that had the best accuracy and was the most used and the recommended algorithm for this particular use case. That was basically a land use change detection. And we used R script that was developed by the NASA R set training.

And it's an open script that we have already also shared. And it's available online. And it's free, it's open source, and it's configurable. And we used this in R. And we developed all of the scripts.

They are all open source, as Tanmay already said. We were focusing on developing everything open source. And the main advantage of the random forest algorithm is that there is no need for pruning, overfitting is not a problem. The outliers are not, they don't influence the outcome of the training. And there are some limitations, though, and the main limitation is that it can

only classify images that you, the spectral ranges that you input in the training data. So the training data needs to cover all the possible spectral ranges of each class that you're trying to classify. So here's an overview of the methodology. I don't know if you can see it properly there or not. But we started the data acquisition and do the pre-processing.

All these steps, the pre-processing steps, to clip the study area, to do the band combinations, to do the image enhancements. Everything was done in QGIS. The development of the training sites we did in QGIS. Only the algorithm itself, because it was based on R, we ran in R studio. And we got the output and we displayed the maps in QGIS and we used the base maps to see how the images were classified.

We did multiple iterations using different training sites trying to get the best results possible. And then we did a change classification and we exported the maps using QGIS. So here's a look of the three images that we focused on for this project.

The first image is from January 1st, 2017. The second is from 15th of December. And the third is from 30th of November, 2018. We were looking to get a gap of roughly one year. And we couldn't find the exact timelines because of the constraints that Tanveer already talked about.

And one of the main constraints in this particular case was that there is a lot of rainfall in Bangladesh. And that rainfall is mainly focused in the months from April to November. And we could only use images that were free of cloud cover. And we found them from January and December. And in November, we were lucky to find one.

But it's usually January and December where there is the least amount of cloud cover. So we could actually do some optical remote sensing. So these are the examples of the three band combinations that we used. The first one that you see is the natural color. The second is the false color composite. This is used to identify vegetation because natural vegetation has a distinctive red shade in it.

And it is pretty useful in this case, particularly because we're trying to see the change from tree to settlements. And the false and urban color is also pretty useful in that case. This is an example of the classification mistakes that appear on the left side.

It's not very clear in the image, but it appears to be a forest vegetation area because the water appears green in the natural color. But it is actually water. And you can see the difference pretty easily when you use a false color composite on this.

So this is the first classification that we did. The main area here is vegetation, as you can see. The spread is mostly vegetation and there's like a very small amount of settlement on the top right side of the image. Then the vegetation is decreased a lot and the settlement expands exponentially.

And you can see that this is the result of the early influx that is mainly just three months of the refugees coming in. And this is the final one from 30th of November 2018. And the expansion is pretty huge in this scenario.

And you can see that almost the whole area is covered by settlements. And I don't think they were expecting them to expand so fast and there's not really much of a plan going on. Okay, so the accuracy assessment methodology we use is a pretty common one, but it is also very effective and straightforward.

So we produced 200 random points using QGIS within the extent of the classification images. Then we vectorized those classifications and did a spatial join to give those attributes of the classifications to those random points. And then we took those points and imported it into Google Earth and used the higher resolution archive data in Google Earth

to then compare those points of vegetation and barren land, those attributes on those points, to see if they're verified and correct. Then we also used UAV data from UN that you can find available in Open Humanitarian Exchange.

So that's actually the image you see on the very top layer being compared to the random points there. The only problem we had with this is that the most recent imagery, both Google Earth and the join imagery, was that it was four months or six months behind the most recent classification that we have performed. And so we'll talk a little bit more about that here in a second.

So here are the error matrices or the confusion matrices that we created for the accuracy assessment. The first date imagery was the 94.5% accuracy. The second date was for the 94% accuracy. And the very last one was 93.5% accuracy, which would probably be due to the fact that we didn't have very current imagery

to compare it to. And then so here we're going back to the fact that the high resolution imagery we had was four months old compared to the most recent classification imagery that was done. Here we had one of the random dots that is supposed to be a settlement, but you can see here it's barren land when we're scanning over it.

But you can see that this is probably most likely being construction being done here, so that you could probably say that this is currently actually settlements by now. And so what we did for fun was just outline all the construction sites, there's many of them on the very west side of the camp, and then overlaid that on our classification map.

And so you can see that probably the accuracy is better than what we actually calculated earlier. And so we then took all those classification of the settlements and overlaid them with each other so you can see the expansion over time. And we also used the raster calculator to also calculate the areas of each one of those settlement areas. And so you can see in the first image going from the second one,

there's a 4.9 square kilometer increase in that date. And then on the second period and to the third, you see a four square kilometer increase. Okay, and so we went further with this and we did a mean coordinates on each one of those settlement areas for each date.

And so you can see that with the lighter colored dot to the darker colored one, how it goes more north and then west. And so that also supports that evidence we have there that's probably correct is that we got another map from the open humanitarian exchange that shows that there are camps being built there on the western side

and also population density is increasing there. So this is probably correct with the analysis we did with the mean coordinates. So what we did next, we're trying to figure out why is it leading that way instead of connecting the southern camp to the northern camp. We're really curious about that. And so we couldn't find any data on it or any information or journals on it.

So we took other data and tried to see what we could find for the reasoning for this. And so you have a DEM here on the left from Astor. Maybe it's the terrain in the southern camp, maybe make it hard for them to traverse down to that area. Or we also looked at some of the UAV imagery that we obtained also from the UN on the open humanitarian exchange. And you will notice that in between the south and north camp

that there is a lot of crops in those areas. So maybe that's why they're not trying to build any camps there. But on the western side, it's a lot of barren land that's not being used. So to wrap it up, we kind of just add a bunch of things that we've talked about and outlined each of the classification dates and settlement increases.

And you can see how it's trending. And if it keeps up that trend over the next year and considering the land use around the areas, this is most likely what you might see in the next year of settlements. And so with this open source methodology of remote sensing combined with open source software like QGIS, you can create effective ways to monitor settlement trends

and document these as well. And so we'll open up to discussion or questions. And if not now, we'll be around a little bit later. And all this, the R script and everything should be handed over so you guys should have it available here soon. Alright, thank you.

Any questions about this?

Actually, I did mention in the beginning that they were from Myanmar. They were coming in. And to check whether the accuracy assessment of the trends that we did were actually accurate or not

because just a couple of weeks back, 600,000 more refugees have already arrived in the country. So as soon as we get images for that, because of the cloud cover now, we cannot. So as soon as we get with that, then we can actually do the accuracy assessment and convert the images according to that and then lay over it and then we'll be able to say

whether we have done a good job or not. So we are waiting for the images to come. Thank you. I think this will be better to an expert.

We actually classified all the... We did the classification for four classes. It was vegetation, urban area, build-up area, water and barren land. So we are only focusing here on the changes that were most apparent, but we did the classification for all the classes.

So we did... So you're doing segmentation? Yes, yes, yeah, yeah. And what is studies? Which one is it and which universities are involved? We are...

All three of us are in the Masters of Geospatial Technologies program. It's a collaboration between three universities, UNL in Lisbon, Universidad Nova de Lisboa, University of Münster in Germany and... UJI. UJI in Spain, Castelho. Okay, thank you.

Any more questions?