Global Seminar | Forecasting district-wide pedestrian volumes in multi-level networks

Hello everyone. I think we can start the introductions for today as more people are joining us slowly.

My name is Panos, I will be your host today. I am senior researcher at the Futures ITS Laboratory and I work for the Architectural Cognition in Practice group.

It is my pleasure to introduce today our speaker Michael Van Egermont who will give a lecture on forecasting District wide Pedestrian volumes in Multi Level Networks in High Density the Mixed Use areas, which is the result of a multi year project that Michael and I worked together in the past and it is really great to have the opportunity to listen to the results of this work today.

Michael.

Was trained in transfer planning and obtained his master's and PhD from ETH Zurich working here at the Singapore TEHA center where he was previously a postdoctoral researcher and a project coordinator at the Engaging Mobility Group and a few years now already he is Senior researcher at the University of Applied Sciences in Northwestern Switzerland.

With now without further ado, Michael, it's our pleasure to have you with us today and we're really forward to hearing about your your work.

Great, thank you.

One more point for our listeners. Please add your comments in the chat. You can also ask them at the end of the presentation or otherwise as Michael prefers. It would be really great if you can speak out loud as we can have a more jovial conversation at the end. Thank you. And Michael, again, the floor is yours.

Cool. Thank you Panels for this introduction. As mentioned, I would like to talk about work that we did in Singapore already several years back in a project together with panels and several others, Alex Serrat, Vera Helle, Pablo Archabilo and Xuchen Shu among others. But I also would like to go into.

Forecasting pedestrian flows in general.

So first I was not too sure about the audience and given that this is online, it makes it even harder to gauge the audience. And.

So really, why do we want to forecast pedestrian flows or why is it of relevance?

First of all, pedestrian mobility is one of the most, if not the most.

Environmental friendly mode of transport. There is almost no emissions involved. There's almost no space or less space involved in other modes of transport. It's healthy, it reduces congestion and pollution. If more people walk, there's less people. Using other modes of transportation in an environment that is comfortable and interesting for pedestrians results in these vibrant centers. And here's a picture from Orchard Road. We took this already a long time ago. I'm a bit scared now. I think of the year seven years back.

That we took this picture of a juggler in Orchard Road and you saw all these people wondering and sitting and enjoying their walk. And this positive feedback loop is something that we would like to understand.

At the same time, there is no, there are not so many methods that can help us in understand the forecast pedestrian flows. This is mainly to do with the advance of motorized transport and public transport over the years. And much of the modeling efforts and focus and understanding and funds have resulted in models for vehicular traffic and to a much lesser extent to models for pedestrians. And that gives us a lack of tools to forecast the usage of pedestrian infrastructure. And.

This is something that I've learned in my time in Singapore is that numbers count. So if you can put a number on the amount of pedestrians walking somewhere or the travel time gains multiplied by the number of pedestrians walking somewhere, then you have a number that you can use in cost benefit analysis. And then you have a number that you can communicate to somebody. And that's something that you can fund policy on. Just to give you an idea. And I drew up some figures for Switzerland last night.

How much do people walk? That's difficult. You have a mode share that you sometimes hear here. I presented a daily distance. So 8% of the daily distance in Switzerland is by foot and to a lesser aesthetic, bicycle, but let's say by foot.

But at the same time people spend 40% of their time walking. I'm not going to give you this trick question that we give during lectures. Why is this? It's because walking is a much slower mode of transport as compared to other modes. But 40% of the time spent is walking and 40% of the trip stages, so that includes a walk to a public transport stop.

Is by foot. Which gives you just an indication how relevant.

Going by foot is as a mode of transport. Specifically in Asia and Singapore.

There are district level urban developments and.

Some pictures for the viewers from.

Outside of Asia. You see here the same street that I just showed you, but then different images of the pedestrian network underground. Almost all the pictures are taken on Orchard Road and you see that there's this fusion of indoor outdoor. So there's this indoor network which is really a network by itself, but it's also part of the network. There's different levels where people move upon, there's people move underground, people move at grade, people move elevated between buildings. And in these images you see different ownerships. Top left you see two shopping malls where the ownership changes.

You see public, which is this transit stop that you might see on the bottom right. And these are all interconnected which gives very difficult challenges to planners. And what more is that These developments are connected to transit and these transit lines here on Orchard Road, there's 4,000 people coming per.

Two to four minutes. They're arriving at a train station and they need to disperse and they need to go somewhere.

Now. Changes in a street layout are very difficult to do ex ante. It's almost impossible to. Once you've built, it's there.

And especially if it's underground or above ground, it's even more difficult to change a repurposed space because the width is already provided for. Whereas above ground, you might play around with the placing of the trees and how many lanes you dedicate to buses and how many lanes you dedicate to cars and to pedestrians and cyclists. Once you construct something underground or below, above ground, it's very difficult to change this. So an understanding of pedestrian flows can help planners in setting guidelines.

And this is also what motivated some of the research that I showed.

So which tools now do we have at hand to forecast pedestrian flows?

The first, and this is a bit. My background, where I'm educated in, is that we have a transport demand modeling process which is called the four step model, you might have heard of it. And this is a very, very rigorous process. It's well documented, it's well understood. Over the last 40 years, maybe now 40 to 50 years, where there are four steps in modeling, you generate trips. So each zone that's first important to understand, there's a zone which is a larger district that produces trips. These trips have to go somewhere to other zones. Once, you know, once you know these flows, you can say, well, X people might go by car, X people might go by public transport. And the other x percent of people, we forget because we don't forecast those, those are pedestrians. And then we assign this to a network.

Now, this process is.

Not used to the extent for pedestrians. There's several reasons for that. Because it's very computational intensive to compute for Switzerland, which has maybe 4,000 zones for vehicles, and Singapore had over 1,000 model zones in the transport model.

To also include this resolution or this level of granularity that exists for pedestrians. Pedestrians have different networks, pedestrians have different preferences. And just out of computational purposes, you exclude a part of the process. Nevertheless, it's good to keep these steps into mind because we're going to contrast this with an other approach in the next slide.

So what you can also do is to forecast the travel times per link, or forecast travel, or the number of people walking along a link, or the number of vehicles driving along a link, is to use something which is called A direct demand model. And direct demand models, they have also been known since the 70s, but they've been a bit forgotten in favor of this more advanced process that we have. And here is a formula. The only formula that I show today, don't worry, is.

Where you have a regression model and a pedestrian count Y or a pedestrian flow Y, which you aim to forecast by a series of variables dubbed X. And we know how much each variable contributes to the dependent variable. And we say that these flows are then a function of. I'll just open up.

Are a function of social demographic characteristics along a link or in an area, in an area with a high employment density, we might expect more people walking build environment characteristics. If there is a high commercial land usage or if there's a school nearby a link, we might expect more people. And the role of a link in a network. If a link is more central, we might expect more people.

This approach has been used to forecast pedestrian flows. I'm not the first, or we're not the first here, but it also has been used speeds for vehicles. And it's a well understood approach where you circumvent some of these steps of the.

Four step model. So you circumvent, for instance, the distribution or whatever, you circumvent these steps and you're just interested in the final outcome, which is the dependent variable Y that we have.

Now. Let me just open my pointer.

Good.

So now if we have this family of direct demand models, we have two schools of thought and again, one comes from transport and the other comes from.

More urban planning. I would say from the space syntax school of Bob.

In transport, what people did is they looked at the accessibility of a link and what that means is they count the number of amenities in proximity to a link and then use that to correlate that or estimate a model of pedestrian flow. In transport, this relies heavily on using metric distance, something that I will get back to later.

Now an other school of thought is more driven by network topology. Space influences what we do.

And there's two measures that come here. There's one is called something which is called between, a centrality, which tells us how centrally located a link is in a network. And what is different is that these methods use something called steps as distance metric. How many steps is a link or a segment?

How far is that from me? The more steps.

The less influence it will have and the less central it is. Or you can think of this something called angular distance. Something that I will show in a bit is that the more turns I would make Is how far something is expected to be, how much, how further something is expected to be.

Now, important to the difference between these two approaches is that one includes land use.

So there is something which land use drives transport flow, Whereas this school of thought says the network drives flow and the presence of a network simply indicates that there is land use. And I know that this is a controversial topic, but. And I'm happy to discuss that afterwards, if people would like to discuss this. Now, you can also combine these ideas and you can include land use into your network topology. So you use a weight. You say that my link has a certain weight, and this weight represents something that is happening along a link. The most basic approach is to say that longer links, so the length of a link is used, drive more or create more pedestrians, create more activity. But you can also say that commercial land use drives more activity. And what you can do, and this is very exciting, is that you can combine different origin and destination weights.

And you can include distance decay, something much more advanced.

Into these network topology measures. And this is something that I will show in the remainder of this talk.

Now, to make it a bit more complicated is that we have this notion of generalized travel costs. Transport planners are very well aware of this concept, that we know that not every minute is equal. A minute spent with friends is much more fun than a minute spent at work, maybe for other people, maybe the other way around. And the same is that a minute spent in a crowded public transport vehicle might be perceived different than in a train where I have a seat with nobody next to me or waiting on a platform. And this is something that we use in cost benefit analysis all the time.

Is that we have different perceptions of travel time dependent on where I am and what I'm doing.

And this is very well documented. Just think of yourself booking your next train or flight and how you consider whether you want to transfer. Yes or no.

Now, also for pedestrians is that a meter is not a meter.

So you might say that a meter along a comfortable route might be perceived as much shorter than a meter walking along a very busy road with lots of noise. And this is something that has been shown in different studies, mainly in transport studies, where we see that greenery shortens travel time, commercial activity along a link shortens perceived distance.

But also if we need to use an escalator or a lift or whatever, or a case of stairs, this increases our costs.

For instance, Transport of London uses figures that say a staircase is equal to 50 meters.

Now, if you translate again to costs, because 50 meters would be now this is from the top of my head would be something of like 30 seconds perceived, whereas. And that's only 6 meters of staircase.

So it's perceived as 5 times as long or almost 10 times as long as walking along a regular link. And this influences how attractive some links are. And this influences in turn our route search. And this influences which routes we use and then where we can expect flows. And this is also something that we did a while back in Singapore. So now our question is when we use different link costs, can we create better pedestrian forecast models?

So just here an example. This is a sketch of Raffles Place mrt.

You might not see it, but in Raffles Place MRT you have this beautiful straight hallway going down to La Basut. And then you have here to. What is it called again this is a while back to oceans, Ocean Tower, something called Ocean Tower. And then you have the high tower one Raffles.

And now nobody. Or if you're there, the question is whether people would use this route which is slightly shorter or whether people use a route with less angular change. There's only one turn and this is beautiful and this is wide.

And now this is one of the questions. Do we have which route do people prefer? Do people prefer to take less turns? Do people prefer to take straight line or many turns? And now.

If people prefer this, then we would need to make this for instance wider or more comfortable or higher that we can cope with flows. Whereas if people prefer this way we would need to make this wider. Etc.

Now think about your own trip through Raffles Place MRT and you will note that during commute hour many people pass through a 1.5 meter corridor here.

Instead of taking the 6 meter wide corridor.

You can go there tomorrow morning and have a look. So in this case, and these are commuters, they prefer simply to take the shortest path which is only like 10 meters shorter than the.

Than the beautiful.

So now as an analyst and as a researcher, I have different choices to make with distance metric. Do I use in my analysis, do I use angular distance metric distance, do I use something called perceived distance? Which measures should I incorporate between this or accessibility? And then there's something called radius. How far should I look when I calculate these measures?

Good.

Now keeping this in mind and if you now know this outline, you think well, okay, I'm going to start and there's all these different softwares that you can use and this doesn't make it much easier. And I'm not going to into depth, I'm happy to email. We use many of These tools also are teaching and for practice. And you have differences in the platform that you can run these tools on.

You have differences in the type of measures that they generate. You have differences in speed, you have differences whether weights are possible, you have whether you can incorporate it into a spatial data flow, like attaching data to links. Is that possible directly? Is it free? That might be important.

Is it suitable for planning? And what I mean there is, you have tools that work with OpenStreetMap.

Which are very good, but you cannot run your own analysis, you cannot add a link. You have differences in how well something is documented.

There are different tools that are very well documented, like the Play Syntax tool or the SDNA is quite well documented. Whereas certain R libraries like SD Planer, you might need to search in the code, which is possible because it's open source, to look at what the code is doing. And now I'm working with a PhD student here in Switzerland and he just prefers to write his own. So that's. If you have a. If you're a PhD student again, you can write your own code to do this type of analysis at C. Tra.

Good. I have two case studies and dependent on time, I will present them both or I just will present one of them. The first one is Singapore, as promised in the title.

Now, this is a picture of 1914 of Singapore. And you will note that the street network didn't change at all. This is just to give you an idea. It's very hard to change planning. And people from London also know that London's network didn't change much between different. So Raffles Place is right here. And what has changed, of course is the density. And what has changed is.

That there is a whole new island here, Marina Bay.

That there is a whole new land here, and that there is an underground network, et cetera. But in general.

The Ed Grade network didn't change at all over these years.

So in our project, which was funded by the URA, the LTA and HDB in an L2 NIC project is we had several objectives and I'm going to talk about only one, which is about quantifying pedestrian flows. We did this in four areas. One is Raffles Place that I just shown. Another one is Drone east, which is located in the east as the name indicates, but which is a planned environment. We have Orchard Road, very commercial, and we have Tampines, which is more residential.

So when we set out, the first thing that we needed was pedestrian networks. And what we did is we obtained networks from different sources. We drew networks we did field visits and.

We digitized these walkways in a spatial database.

This was not a 2D network or a 2.5D network, but this was a true 3D network.

In this case here for Jurong east, the red lines indicate indoor network.

With bridges between different buildings. We have escalators, we have the outdoor network. And what we have is something which is up for discussion is sidewalks.

We explicitly chose to include these as these roads are.

Three lanes in each direction with a fence in the middle that prevents crossing. So we consider this as truly separate walkways despite them moving along the same line of sight.

Now, how we worked, and this was a painstaking was a painful process. How to get here is that we wanted to have this 3D. We want to have data attached to the network. And what we finally chose is a proprietary solution with S3RGIS server, which we where we use an open source database on the back end. But S3 allowed us to draw some in Autodesk AutoCAD 3D. And it allowed us to attach data to the network at the same time. By using Postgres, we could use QGIS for some visualizations, we could use R for modeling and we could use other open source tools just to interact with the data. We could write queries to update. And we weren't too dependent on on the S3 environment.

Now this is a graph and this is something that I will get back to later is that why do you draw the indoor network? This is commercial. You shouldn't care about this indoor network.

This is just where people go shopping. But it is truly part of the.

Network, of the publicly accessible network. And.

Simply most people walk there. It's integrated with transport. And this just shows how much length of the network per site is indoor and outdoor. And for instance, in.

Jerome east, what we see is that, and this is kilometers, these figures that you see here is that 36 kilometers of the network within 400 meters of the main transit station there is indoor, is private, and only 17 km is outdoor.

Now, even if we move further away from the train station, we see that 8km is indoor still and 36 is outdoor. Our radius is getting larger. This is something that happens.

And we see this is directly integrated with the transit hub. So this is where we need to plan for. This is where we need to look at as well.

What we did and this is we wanted to have pedestrian count. And.

Unfortunately we couldn't tap into towns for the shopping malls. Have they have their own counting systems. But we needed to install our own devices. And just thinking of the process. Again, this was a painstaking process because we wanted to employ counters in private developments and we want to employ.

I need to share my screen again. Right.

No technical issue.

Cool. But we're back. So counting and all we did is we did manual counting with manual gate count for 10 minutes and we did this for three days. So we had at the end we had approximately 80 counts per site, which is a for modeling. It's not a high number as compared to vehicular transport, but it is quite high for pedestrian studies. And this was all pre Covid, so we had sufficient.

Cool. So what we did is. What we did is we generated different network measures. We use different rates, we use metric distance, we use perceived distance, we used angular distance, et cetera. And we used.

Accessibility metrics. We generated tool phase between these metrics. This is something that we'll show in the next slide. And we use different radii. So what we did is we use different radii to be able to capture distance decay. So that we can say between this between 0 and 200 meters is more important than between this between 200 and 400 meters. And I will show you briefly a comparison how these different methods compare.

First, this two phase between this metric. This is new, that's why I chose to show this is that we have for instance in origin an MRT station. And we know that 100 people arrive there in a certain amount of time. And we know that 80 people go to this destination the hospital and 20 people need might go to another destination. Now what we would.

Do in a normal network analysis is that we would just simply say I have origin links and destination links and these two magic numbers indicate to me that these both links are equally important or almost equal, equally important without using weights at all. But now I do have a priori knowledge on origin and destination, so why shouldn't we use this?

And then I can say, for instance, I know that along this link there will be 100 people walking, along this link there will be 80 people walking. And along this link there will be 20 people walking. This is an example with perfect knowledge right on route choice network. It's a very simple example.

Key point is that we use land use or we use land use and bus stops and transit to inform our betweenness metric and we can use different origin destination.

Now in our case we chose not to use.

This perfect knowledge, but rather we represented.

MRT station as one.

Because we didn't know or we chose not to use the amount of passengers arriving and what we did use is a certain G of A. And now we get a magic number. But this magic number now is between this figure 0.8 and 0.2 does represent that this link is likely to have more pedestrians walking because there's a higher weight at this end.

Now, fortunately for us, we had.

Rather good knowledge on the land usage and we use different land use types. So we calculated these metrics for retail, office, residential units.

And now we have series of links or a series of metrics for each origin destination. So I know I use this trip generation something from this transport demand modeling process. And I know something about drip attraction. And this is something that I put into a model that I will estimate later first into correlations and later into a model.

Just to indicate.

How such a metric looks like that. You better understand this. This is the betweenness from an MRT station which is located right there smack in the middle with to retail to these different shopping malls. And we have here we have the JQ with the ice skating ring, if it's still there.

Here we have Jam and here we have imm. And here we have the location formerly known as Big Box. I'm not sure what's there now.

And what we did is we counted at all these red lines. We use these different methods to count. We counted at different distances and we counted at different levels. Okay. Just to know where there would be how much this between this metric influences.

The footfall.

Now just getting to the key point here is that when we use three different model types. One is a model simply this classical space syntax model where we have length weighted angular change. Between this another model where we use accessibility which indicates the proximity of amenities to each link. And we use this two phase between this metric and now what I show here is the row square a goodness of fit. We use different goodness of fit measures. I chose here to use goodness of fit method called rho square. This is something we'll show the model estimates later. But including.

Land use in Jurong east dramatically improved our model fit. So if we use a betweenness metric without land use, we get a rho square of 0.17. If we use land use in our model in our betweenness metrics, we get a row square of almost 0.7.

And this is a trend that we saw almost in all sides.

When we include land use in our betweenness metric, we get a very high rho square rather good models than we estimate at Orchard Road model that also works quite well without including land use. And very exciting to see when a raffles plays in this Old district, we get a very low row square almost like you would throw away this model using angular change as a between this metric. But you get a better rho square still far from the other sides. We can discuss that later why that is when we use these two phase between these metrics.

And this is something that we fitted on average weekday peak flows.

Well, how does such a model look like? And I've chosen to show you two models.

This is Tampines.

This is a top view. Not ideal, but it does the. It does the trick. We have the RT station here and a dark color indicates how many people we expect there. And now here we use an accessibility model. And what an accessibility model does and you see that beautifully. Let me just think right here in this area.

Is it uses the characteristics of adjacent links on each link. So it's not very sensitive to betweenness. And there we have a beautiful landscape. Where we see here we have lots of activity, lots of pedestrians, but it's equal on all links judging by the color. And then sudden it, it drops off very gradually.

Now if we use between this metrics, what between this does it shows it searches the shortest path. And.

I'll show that on the next slide right here. Now this is for two sites. And this is Drone East. And here we have a pedestrian flow.

From the.

Train station right here to the different developments. And it goes right there. And here we have a good flow or here we have the pedestrian flow continuing to the final wall in the area.

Now it's important to understand that between this only results in one shortest path.

So what I mean by that is that there might be a second path going outdoor. But this path will not be considered in the route search. Between the train station and this mall there will be only one path leading over the.

Jaywalk over these bridges.

And the same for Orchard Road is. And then I need to twist my head a bit here is that we have.

See this is Iron Orchard right here, the MRT station. And then we have the flows going to Mekong City.

Different locations. One thing that we see is that these, these pedestrian flows, they're going through buildings. And this is something that I tried to say earlier. We need to plan for these because these are the most important routes that people take not only to a single building, but also to subsequent buildings. So we need to plan for these link.

Now just that you see which variables we put in our to our model. And this is a selection. I think we have a bit more.

Is that shows us that dependent on the site different variables contribute differently to pedestrian flow. For instance, we have in Jurong east the highest contribution is provided by retail, whereas in Raffles Place the highest contribution is, not surprisingly by office land usage.

Another thing that you would see if you spend a bit longer looking at this is that there is something called distance decay is that 200 meters contributes most, then we have 38, which is a bit less, then we have 25 MRT to retail, which means links that are retail, links that are further away from the train station get less foot flow, footfall from the train station. And then further away we have almost nothing, but at the same time there is nothing there. So this is driven by spatial structure. In Orchard Road you don't see the same trend that that sharply, especially not near to the train station, which has something to do with.

That. We have difficulty with employing counters very close to the train station. So this effect might be there. And this is the difficulty. If I use these two phase between these metrics, I need to have counts at all these different land use combinations. And something that we can discuss later if you want.

Now I spoke about this distance and, and this is one side where I show the goodness of it. And we didn't see.

We didn't get clear indications which distance metric works best, unfortunately. So we see that metric distance actually performs quite well, judging by this rho square. We see the angular distance doesn't.

Does well.

But we also see that perceived link distance works well and it is very sensitive to how we attach our land use to the links. And it's very sensitive to network topology.

So maybe drawing a very detailed network.

Drawing a very detailed network makes it very hard for these between these metrics to find the correct route at the end. And introducing these different distance metrics makes it also more difficult. But we see that angular distance does it, as I mentioned, rather well in this case, in this side that I show.

Now, what time do we have?

Now you can also do this for other areas in planning. This is an area called Dreischbitz in Basel. There's a train station right here. And.

Check the orientation north.

Yes, there's a train station right here. And this has this very strange layout. And this strange layout is because of the railway. There's a railway coming in and all these streets have tracks to load and unload. Now this is not happening anymore. And now the question is, or this area is being redeveloped.

And in this case we simply didn't have counts, but what we did have was number of transit passengers arriving, different stops. We had approximately the number of planned people in each building, number of jobs.

We had, where the car parks were. And now the question is, can you support us using this method? What are the most important links in the network? And can we attract people to this railway track which currently looks like beautiful as in this image and can we make that the main pedestrian through there?

So what we did is, what we do is and this is something that people in urban planning are common with. We have two networks and we ran the current situation and we looked at the current situation right here and we see well, people don't want to take this railway track. They're not using it for whatever reason. But while people are not using it, our analysis indicates that people will not use it.

So we might need to rethink our network. And when you do that, and this is something that you do iteratively with planners is you say, well, we might add a new link here or better still we could add a new link through an existing plot and this is again where private meets public. How can we make this link more attractive that it appears on people's path?

Or can we even make an upgrade to a plot where we can route people through a building at the same time? What such an image shows you is where should I prioritize infrastructure at the moment? And it clearly shows if I need to set priorities at the moment, better not to do it on this side because I don't expect much pedestrians here. It's too far from the stop. We have many people arriving by car here but may probably we can better use our funds to create on this.

Just as a case study to show well, I don't need these elaborate networks. I can also do it for an existing site without pedestrian counts simply using assumptions on transit passengers, a real number n number of jobs, a real number and then my between this metric will yield me a number that I can directly use.

Just as a summary.

Is that what we see is that pedestrian flows they occur indoor and outdoor. Especially what we saw in Singapore is indoor Understanding of the flow through and between private developments is necessary. And these developments start at the MRT station and go to retail office and retail to retail.

Now we as planners should aim to account for these flows and plan for these flows. Despite that these are off our usual outside of our comfort zone and it might not be everywhere the case, but still it's not in the public realm, which makes it a different mindset.

Now when setting out guidelines for developments and districts around MRT stations, it's clear that people prefer these direct paths. And this makes it incredibly Difficult. Where would you plan a district like Hoarsha Road at one go? You don't do that. Right. It's a process of years and this is a challenge for phase developments.

You want to set standards for widths. How wide should a walkway be? You might want to set a standard for height, but you also need to set out a standard for connection points. And these connection points to make it more difficult are on different levels.

And that is a tremendous challenge. And I'm not involved in these challenges as I'm more into research than really into this hands on planning. But what you need to do is you need to design this network. And developers should build around a network eventually. And that makes it very difficult, especially because it's vertical.

So it's the same as with the Manhattan grid or with the other grids of planned networks is that we need to consider how does my network look like? And inside each plot that I want to be redeveloped by a pipe developer.

Now.

I don't see why you shouldn't put land use into your network measure unless it's not available. Incorporating land use in your metric makes complete sense. And there's multiple tools that can do this. Now we use sdna, but.

Urban Network Analysis Toolbox can also do this. Now.

At the same time, and this is a real difficult idea that we had is that not all land uses are equal. And this is something that we will do in the next study now here in Switzerland is how can I predict the number of people entering and leaving the store? For instance, I have a Uniqlo. This is the Uniqlo in Somerset, day of opening or something, huge queue outside. But you know, Uniqlo tend to be very, very busy, like the H and M et cetera.

At the same time we have a Louis V Torsion. There is a queue. I know, but this queue is not the same queue. There's a queue in Singapore for Louisiana, but it's not the same queue and it doesn't attract the same footfall as a food Republic or as a Uniqlo. And now you need to think about if I want to model these flows on these level and this level of granularity, you really need to take into account this detailed land use. And this is a function of the type of retail. It's a function of the visibility in the network, it's a function of the integration in the network or the accessibility. And to make it more complicated, it's a function of the betweenness. There is a self reinforcing loop, probably if a shop is located along A certain link people would simply just walk in because it's there. So there's this reinforcement loop. And to make it even more complicated, developers have their own science on where to put different shops to improve.

Footfall and different operations.

Another thing that I would like to highlight is this, between this metric, which is very sensitive to how you draw the network.

Now, for instance, Crispin Cooper has put in work into random link costs, where you say, well, I don't know exactly what my link costs are, but I sample them from a distribution. And here I said, well, these are just three examples. I have here three costs along different links. Now I sample this and I have multiple analyses. And it might be that this path here is 80% of the time on the shortest path and this path here will be 20% of the time on the shortest path. And by doing so I get better between these metrics. And by doing so I account for uncertainty in travel times, which can traffic lights, for instance. And I can account for differences in taste preferences between different people.

But of course this takes much longer to run. One run in our case took maybe four minutes for our networks. Now, if I want to run 100 between this analysis 100 times.

Even with parallelization, I might still need 10 times as long for.

Or for 100 between this analysis. But I do see that this is really a way forward in using between this and getting better.

Good. I'm happy to discuss. I didn't hope that I took up too much of your time. If you have any questions, please ask them now in the chat or discussion. Or just drop me an email if you have any questions. Great. Thank you very much for your attention.

Thank you, Michael. This was a fascinating talk.

Please join me in giving Michael a virtual round of applause for sharing with us both the theoretical underlying.

Principles behind spatial network analysis and also how Michael has been applying these methods in Singapore and Switzerland.

Are there any questions for Michael? I think he presented a very rich presentation.

You can either voice them. I can already see one question from Tanvi on the chat.

Yeah, that was just to. I just want to clarify, Michael, you showed three models, accessibility, space, syntax and then a combined model, Right? So when you're talking about accessibility, that is weighted by the gfa, correct?

Yes, exactly.

So that's a significant improvement, even after waiting by gfa when you include the space, syntax, measures.

Yes, in most cases it is. It is a very.

Significant improvement.

It has, yeah, I can go into the details, but it has to do with. With accessibility, all links will get. All links in close proximity to each Other will get the same weight. And this is something.

Where between this or these. Is between. This metric only would attach weight to a single link. Yeah. Mm.

So I do think that accessibility is a very good metric to forecast other things like how many people might go somewhere or how many people would linger somewhere. But I think it can serve very well. But to forecast pedestrian flows. In our case, accessibility didn't work. That work.

We have one question from Iwana Stavurulaki from Chalmers University who is asking, how do you exactly calculate perceived distance?

Yeah, thank you for that question. And I realize now that I didn't put a slide in the appendix on that.

What we do is if I have my network. Yeah.

This is a good example.

What I do is that the most simple case is that I say that a level change in my network costs.

Something more. So in this case, this is above grade. I might say that this is perceived as one. So this here I multiply the distance by one. And now I know that an escalator which is located somewhere in this building is perceived as a distance multiplied by five, so it gets longer. And now if I use a route search algorithm, I have different weights on my edges and a link will be perceived. Will not be included in the shortest path. So that's the most simple case. And I think that verticality is something that.

Has.

Is not up to discussion as much.

And you can say that this represents different things. This represents that you don't know the path because it's on a different level, because you don't see the path, or it can simply because you dislike climbing a staircase. Now, what we did in Singapore is we did a stated preference study and we showed people lots of images on different streets. And we asked whether people would be willing to detour or take a certain path to avoid.

For instance, a route along a busy road. And we saw that people would say, well, I prefer walking along the corridor, or I prefer walking along the street with shops, above a street without shops. And you can also put that into your network. And you say, well, a street with shops is perceived as 16% less. So I reduce my weight of my link by 16% because there are shops along. If it's along a major road with lots of noise, I increase the distance by 16%.

Now, these numbers are not universal and there's no universal route as the climatic conditions are different, preferences are different.

But you do see that in London, I think a staircase, as I said, was perceived as 50 meters. And in Singapore they found approximately the same. So you can try to include that as well. Is that clear? Iota?

No worries.

We have another from Don. I'm not sure if you saw this.

Yeah, I see it now. What are the variable used for the weightages and how are they derived? So what we use is we use in this case, in this image, what you see here is, you see, for instance, we use whether there's an MRT station and what we do is we say there's an MRT station at this link.

That's weight one. It doesn't matter too much which weight I use because I use a regression model after. Now for GFA is what we have for this development, we know what is the actual not GFA but the net floor usage by land use. So we know how many shops are there, we know how many restaurants are there.

By square meters. And now we make an assumption, and this is simply a data assumption. We distribute this.

Proportionally according to link length in the building. So each level in a building has the same amount of GFA and that's what we use as a weight.

Now that means that a longer link within a building.

Has more weight because there's more shops or there's more GFA according to our method. But between the buildings, if this building has more gfa, so if there's more surface within the building.

This will have per meter. It might have more GFA than this building. Now what you can do is you can.

Instead of that you can use trip generation rates and you can say I know that so many that 10 square meters of office generates one person and I can use that one person as a metric.

What we did is, and this is something in Orchard Road is we really had difficulties with the.

We had difficulties with exactly what I show here. So you have for the people familiar are not sure which dan this is, but we have people.

Here underground and we have people above ground, but both shops have the same square footage. So in our model it assumes that both shops generate the same amount of people. And our predictions were really a bit off when we used a homogeneous distribution. So in this case we simply said there's a luxury mall which is above ground, which is the case. And then we have a auto mall which is below ground. And we.

Attach a bit more GFA to that to represent that there is more football generated by the model below. So what did we include at the end of the model is we included these factors that you see here. You see MRT to office, MRT to residences, MRT to retail, you see retail to retail. Then on the other page, there would be bus stops and there would be.

Simply a link to retail, for instance, and link to mrt. And these are the variables that we included in the.

In the model.

One more question from Tanvi.

I think, yeah, I'm just curious, like, if you have any broad general understanding of methods of doing this pedestrian floor analysis based on the land use and district type. Because I see it's very different from all different. The four different district types, right. The office district, mixed use, more residential and more commercial. So how is it different when you use different methods in these.

So we can. And this is what we did is I estimated like a combined model for Tampines and Jerome. And that worked. The model fit is less good, but it's still acceptable. That means that we put these two together, these two columns. We put these together.

So they're not that different.

In that sense that they generate the same amount of footfall.

And you can combine it. You just get. Because you have a larger heterogeneity in your variables, you can.

Generate.

You can get a model that have it less predictive power now.

But is there is a reason. I mean, I. The previous slide with the. Where Raffles model fit was 0.01. That was really drastic difference.

Yeah, this is a different. This is Raffles. I feel that we have difficulties because we don't use random link costs. What we.

Did is we counted it, for instance, at intersections in Raffles, and we counted at both sides of the road. And I feel that.

This. Between. This metric, as I said, is very sensitive to what is my shortest path. There's only one shortest path from origin to destination. But in real life, your shortest path will depend on whether a traffic light is green or red. You might say, I take this route because I don't have to wait two minutes for the traffic light. So I take a different route. But in our model, we don't have this because we have these sidewalks.

That. That didn't work too well, that idea of only one shortest path. Whereas in Orchard Road and Jerome east, this network, there is almost no. You. You really model these lines of sides because there's only one corridor. And.

That really helps in getting this good model fit. And the Raffles Place network is much messier because of the sidewalks and also because of the train station, which is very difficult. But there's much more heterogeneity and tastes of people of the routes that they take. And I feel that that's the issue there.

But this is something that we want to look into.

Mike, if I may, I have A question as well about the.

Transferability of these models because in a way you show that, you know, we can get very high quality predictions by adding more information about GFA calibrating between us.

What was your experience with for example translating some of those thoughts in. In a new environment where you don't have as many pedestrian counts to calibrate them? Does it matter? Do you get.

How do planners utilize that information?

It's a. So the case that I showed for Basel is we don't have any counts. It's an assumption. And I think the first thing that you see in these analyses is how about we know that it's very.

At least for me because I do lots of network analysis and I never find the results surprising. Because you think ah, well, I would know what is happening here.

And then you start thinking closer. And then for instance this conclusion that we don't need to intervent immediately in a district. This is something that I wouldn't look at because I always look at the high flows and other low flows and then a planner would say ah, but then I can prioritize for instance by using this.

Another thing is that this really shows. And this is now clear on this image. This is a very. This intersection is really terrible. There's only one crossing. It's.

There's only one traffic light instead of four. It's. It's really a bad intersection. And now it really becomes obvious to the transport plan is okay now we need to do something about this intersection. If we expect all these flows and this is without count. What we do here is we count bus stop to jobs and bus stop to students and we're train stop etc. We have four different types of flows and we add these together.

So it. It does help in an understanding. But you do get questions in which routing algorithm do you. Which distance do you use?

But it. I would say that it's very interactive in that sense that where do I need to add a link? It's much more in a discussion fashion.

Yeah.

But the. The method with the land use it's. It's very sensitive as I said to having the right counts at the right locations. And this makes it very difficult with the two phase between these metrics. Especially if you use the small distance bins because then you need to have account at each distance bin to be able to estimate a model a parameter for that.

Two phase between this combination. But you might have a counter hand.

But your general takeaway is that when we can we should be using GFA because.

Or what Andres did recently Andrevsky did a paper where he used trip generation rates.

As a, as a method. But I.

I feel I don't understand why you shouldn't use it, especially not in environments that might not have been grown so organically, like.

Old city environments where a high density of links represents a high density of footfall that you expect.

So that's.

Why you should use these methods simply to get an understanding of. And what this method of course does is it shows you the sum of shortest paths, right? It's not one shortest path, but here it shows the sum of shortest path between all these buildings here from this origin. And this is something that might be harder to compute from the top of your head, you might be able to compute one shortest path combination.

From this district to this university, but not to all these small destinations. And this is what it shows you, the cumulative outcome of this.

And perhaps one last closing question. I don't see any other question from the audience. So how do you think this approach of computationally.

Estimating pedestrian flows and pedestrian activity fits with the broader.

Vision and mission to make cities more pedestrian friendly, walkable? You know, we hear about the 15 minute CT as an approach for urban planning. How does this complement our current methods? Since you have both research and practical experience.

First, I think it's still, and this is a challenge also when working with students and with other people, these methods.

They allow you to zoom out too far.

And what I mean by that is if you're a pedestrian, it's about the micro level environment.

It'S about these small things that matter. It's about sweating at the traffic light for two minutes, it's about the car passing by too fast. And it's maybe not about this 400 meter walk or what is this 800 meter walk. And these GIS methods, they make you zoom out very far. And it's still not the perspective that a pedestrian had. And that I think is a real challenge. And to plan for pedestrian actually means going out there and experiencing the world as pedestrian and not only as yourself, but using different glasses and walking with the pair of glasses of child and elderly person.

A couple, a group, and to experience the environment in that way. And I think still think that you should plan for pedestrian infrastructure that way and maybe not on this very high sundam level.

A challenge with these methods it's still the mode choice. Question is we assume now pedestrians being constant and there's not more pedestrians being added if we make a better walkable environment, the cake remains the same.

And we would like to get this pedestrian cake to get bigger. Or if there is a bicycle infrastructure, you might compete with the bicycle. And these direct demand models, they don't. It's very difficult. The cake remains the cake and I divide the cake over my area, but the cake doesn't get a wedding cake or whatever with multiple. It doesn't get bigger. And now this is something where you need to have feedback that you say, if I have a more friendly environment here and if I do all these interventions, I will get more pedestrians. And this is something which these models still not have, these direct demand models don't have very well, which is a very important question because that's as I said. Now I can do a cost benefit analysis maybe for this. Somehow to make this improvement, I get more people, but I don't get more people in general.

That's a very difficult question to answer. This is also a very important question to answer. But on the other hand, it's just a question for why not do it? If we build infrastructure for cars which cost millions or billions, we can do an improvement which costs a couple for pedestrians.

Thank you.

That's very insightful.

Are there any more questions from our audience?

You've done a great job answering all of our questions.

So please join me all with another round of virtual applause. And Michael, thank you so much for joining us today from Basel, Switzerland, all around the world where we are.

And present this work that has been very important both for us, for fcl, for the pedestrian projects, and also for the world of urban planning at large. It's kind of putting together the pieces.

Of the puzzle on pedestrian mobility. Thank you very much, Panaza. Thank you all for staying until this time of the evening for you guys.

And also one round of thanks to Helen and to Anjana for organizing this talk and coordinating with everyone. We'll have a series of talks at FCL every Wednesday, coordinated by Tanvi, Helen, Jana and our colleagues. So for those of you that join us from outside fcl, you can subscribe and join future research talks. Thank you.