Leveraging AI for megacity planning

3:48Hello.

3:51>> Hello. Marian,

3:53>> can you now hear me?

3:55>> Yes, I hear you.

3:58>> So, I just want to make sure that like

4:00others also on the on that platform can

4:03hear me, but I'm not sure if they can.

4:06That's the uh issue right now because I

4:09don't have any other audio from you know

4:11the event here.

4:13>> Perfect. All right. They say like we can

4:16hear you.

4:17>> All right. That's great.

4:19>> I don't see any any camera from the

4:22other speakers.

4:24>> Is it everybody with the with the camera

4:26turned off?

4:28>> All right.

4:29>> Yes. Now

4:31>> perfect.

4:33All right. I think like the technical

4:35issue is over now. Hello everyone. Um

4:40good morning and a very late good

4:41afternoon from San Francisco here. It's

4:44around midnight and yeah it's like 12

4:48like some 5 minutes past 12 here. So I

4:52am joining you from a very quiet Bay

4:55Area right now. I am Mariam Husini. I am

4:58an assistant professor at the University

5:00of uh California, Berkeley and my

5:04research is at the intersection of

5:06computer vision,

5:08urban planning and accessibility for

5:10specifically for people with mobility

5:12and vision impairment

5:15and um I am also like an open-source

5:19developer for the past five six years

5:22and even more and I create tools and

5:25techniques to tackle the data scarcity

5:28and also the problems that we have with

5:30the uh design of the urban public spaces

5:34specifically for them to be more

5:36inclusive and to have more inclusive

5:38cities.

5:40So I am very pleased to moderate today's

5:43session on the emerging uh applications

5:46of AI in metropolitan planning and

5:49public service. We have an excellent

5:52group of speakers today whose work has

5:54spent research institute and public

5:57agencies and also global networks

6:00focused on the challenging

6:02uh issues of the mega cities today. So

6:05let me very briefly introduce our

6:08speakers. We have Dr. Dr. Jonang Choi

6:12who is the director of the center for

6:15external collaborative at the soul

6:17institute and also leads the mega city

6:20think tank aliens secretariate he was

6:24work focuses on urban analytics

6:26metropolitan governance and AI supported

6:30planning systems

6:33uh we have Mr. Rammon Prunetta Flip who

6:37is the chief technology officer at the

6:40AMB

6:42Informasio is surveys SA the public

6:46company of the Barcelona metropolitan

6:48area that develops the uh and develops

6:52and operate the mobility applications

6:55and also information services for the

6:57millions of the daily users and

7:00apologies if I'm not pronouncing the

7:02names correctly.

7:04You did very well. Thank you.

7:05>> Thank you. Uh we have also Dr. Kimon

7:09Jang who is a posttock fellow at the MIT

7:12sensible city lab and he will also be

7:16soon to be an assistant professor at the

7:19HKUS

7:22and his research uh focuses on

7:25generative AI and explanatory

7:27computation methods for complex

7:30metropolitan systems.

7:32And last but not least, we have Dr. Alec

7:35Hoyer, the director of the international

7:38affairs at the Institute of Paris region

7:41and he serves as a chair of the MTPGN.

7:46He has an extensive experience in

7:48international metropolitan corporation

7:51and long-term regional planning.

7:54So our agenda today reflects the breadth

7:58of methods work and also the shared

8:01interest in understanding how AI can

8:04support more adaptive and also equitable

8:07and sustainable metropolitan systems.

8:10We will begin right now with an

8:13introduction of the meta uh followed by

8:16three forecast presentations.

8:19So um after our second talk we will take

8:24questions from the audience and then

8:27we'll go back to our third talk

8:30and also we will close with a very brief

8:32wrap-up. So um with that I will invite

8:36Dr. Choy to begin with an overview of

8:38the meta. The floor is yours.

8:42>> Thank you.

8:53Can you can you see my screen?

8:57>> It's good.

8:59>> Yes. But we are seeing

9:02>> Yeah.

9:03>> Yeah. Yeah. I will briefly introduce

9:06Meta and then uh right after the

9:08introduction I will uh follow my

9:11presentation.

9:12So

9:14>> I think we have the we have the

9:16presenter's note here.

9:19>> You want to share the other screen?

9:21Perfect.

9:23>> Yes.

9:24>> Yeah.

9:26Yeah. My name is Ch. I'm currently

9:29working at the S institute as a director

9:33of center for collaborative research and

9:36also the director of the meta

9:38secretariat. The meta is meta is the

9:41mega city think tank alliance.

9:47Uh meta is a international alliance of

9:50think tanks. We focus specifically on

9:52addressing the unique challenges and

9:54opportunities inherent in mega cities

9:57worldwide. Our primary goal is to bring

10:00together leading expert and institutions

10:02to share palace development

10:05strategic initiatives and research. Uh

10:08Meta was founded on July 11th, 2014. Our

10:15initial founding members included the

10:17soul institute, the Beijing Municipal

10:20Institute of City Planning and Design,

10:22BICP, the Singapore Center for Liable

10:25City CLC, the Shaian Planning and Design

10:29Research Institute,

10:31uh, and Hano and Himin City Institute

10:35for Development Studies. Furthermore,

10:37three international organizations cityet

10:40metropolis and elay supported

10:43uh uh and actively participated in our

10:47uh founding. The purpose of metal is to

10:50create a collaborative international

10:51network that facilitate the exchange of

10:54information on urban challenges and

10:56solutions among member think tanks by

10:59coordinating problem solving efforts and

11:02encouraging partnerships. Meta aims to

11:04support in fostering bilateral and multi

11:07multilateral cooperation to address

11:10urban issues effectively.

11:13Meta operate with a structured approach

11:15to ensure active engagement among

11:16members. English is the official

11:19language and soul institute serves as

11:21the as the secretariat. There are

11:24several types of meetings. Meta forum,

11:27general assembly, academic forum and

11:30working level meetings.

11:35Meta is currently include 14 member

11:37institutions from Asia, Europe and

11:41Middle East comprising urban and

11:44government affiliated think tanks as

11:46well as two international organizations

11:49and ELA.

11:51uh together these institutions

11:54collaborate on urban policy planning and

11:57innovations to address shared mega city

12:00challenges and promote sustainable

12:03development.

12:07Since 2014, META has held nine forums

12:11across Asia discussing key urban issues

12:14like public space, mobility and

12:16sustainability.

12:18The most recent, the ninth forum in 2025

12:23was held in soul and gathered

12:25representatives from 19 institutions

12:28across nine cities including observers

12:31highlighting meta's growing global

12:33collaboration.

12:36META's activities are guided by its

12:39articles of association covering areas

12:42such as membership, governance and

12:45responsibility.

12:49For meta's operation in next year 2026,

12:54three key joint joint research themes

12:56were proposed. First, AI mega city

12:59planning proposed by soul was formally

13:02adopted at the 9th general assembly

13:052025.

13:06Second, Paris suggested a datadriven

13:09policy approach especially for

13:11addressing challenges such as housing

13:14crisis during bilateral discussions.

13:17Third, Singapore proposed the theme of

13:21health city which was also shared during

13:24the general assembly. In terms of next

13:26forum, we we are we are pleased to

13:28announce that the Beijing Institute of

13:31City Planning and Design BICP has been

13:33accepted as the hosting institutions for

13:36the 10th META general assembly and forum

13:39in 2026.

13:42Yeah, that is a brief introduction of

13:45beta.

13:49And uh now I want to uh present the my

13:54presentation. It is a AI supported

13:57system for daily life zone planning

14:01focusing on how accessibility and

14:04mobility based approaches can define

14:07more realistic urban living zones.

14:10Although I'm from the S institute, the

14:14algorithm was tested in Busan, southern

14:17part of southern port city of South

14:19Korea to demonstrate its applicability

14:23at the national level.

14:26The so institute established in 1992 and

14:29fully funded by the so metropolitan

14:31government operates with six research

14:34divisions, one AI lab and one center. It

14:37is home to 300 researchers and step with

14:40a 2024 budget of 37 billion US.

14:46Our research this research is driven by

14:49three key three key factors. First,

14:53there is a growing interest in the daily

14:55life zone DLG open associated with the

14:58VP ministry plan. Following COVID 19,

15:02the importance of DG which focuses on

15:05short distance travel has been signific

15:08significantly highlighted. Second, we

15:11reflect on the limitation of

15:12administrative convenience. Currently,

15:15the setting of DGS often centers on

15:18administrative district boundaries which

15:21inherently limit the scope of actual

15:24residents daily activities. Third, there

15:27is a clear need for a datadriven

15:29approach to reflect real world

15:31activities. Our primary purpose is to

15:34prepare

15:36a datadriven foundation for setting up

15:39DG that genuinely considers the living

15:42activities of actual residents.

15:44Ultimately, this study aims to derive

15:47datadriven living areas that accurately

15:50reflect resident actual activities.

15:54We focus on two main approaches for

15:56defining daily life zones. First, we

15:59analyzed daily life zone DG based on the

16:03accessibility of neighborhood living

16:05facilities. This includes key facilities

16:09like kindergartens, elementary schools,

16:11libraries and parks. Second, we analyze

16:14DG based mobile data reflecting actual

16:18population flows and mobility. Our study

16:21area, the Gangdong region in Busousan

16:24city offers an interesting context

16:27characterized by being surrounded by the

16:30Nakdong River and mountous terrain and

16:33featuring a mix of industrial and

16:35residential areas. This diverse

16:37geography is crucial for observing how

16:41DG is formed.

16:44This model supports community planning

16:46through three core functions. DG

16:49planning diagnosis

16:51plan zoning and facility location

16:54support. It evaluates accessibility to

16:57neighbor services and public transport

16:59and helps determine optimal locations

17:02for single or complex service

17:04deployment.

17:06This study focuses on daily life zone

17:10resoning using both living facility

17:12accessibility and and mobile data as key

17:16inputs.

17:18This study focuses on the Gangdong

17:20region of Busan city covering about 15%

17:24of the city population. The area

17:26includes mixed industrial and

17:28residential zones and is part of Busan's

17:312040 abandon plan. We applied the

17:34datadriven approach to define daily life

17:37zones using two criteria accessibility

17:40to neighborhood living facilities and

17:43mobile data patterns. We then compared

17:45the spatial accuracy of each method

17:48using the IOU index.

17:52This study uses three key data sets.

17:55First, the population data in a 100x 100

17:59meter grid helps identify residential

18:03patterns. Second neighborhood living

18:05facilities like schools, parks and

18:08welfare centers were mapped across 12

18:11facility types and 545 locations.

18:15Finally, mobile visual location data in

18:1850 by 50 m grids was used to analyze

18:22actual movement patterns. This mobile

18:24data was reaggregated to match the

18:27population grid for consistency.

18:32uh our study applied two key method for

18:36data analysis. First is the community

18:39community detection method using the

18:41lubang algorithm. This algorithm

18:43automatically identifies groups of

18:46strongly connected nodes by maximizing

18:49the modality modularity of the network

18:52which helps us detect the spatial

18:54communities based on structural

18:57connection patterns. It is especially

19:01efficient for analyzing large scale

19:03urban data. Second, we uh use the

19:07intersection of union or IOU to evaluate

19:10the spatial similarity between different

19:13daily life zones. The IOU value ranges

19:17from 0 to one with one indicating

19:19perfect spatial overlap. This metric

19:22helps us compare the accuracy and

19:25consistency of community boundaries

19:27generated from different data sets.

19:32Our analysis follows a threestep

19:35framework. In step one, we extract

19:38analysis specific grid sales based on

19:40three data types, population,

19:42neighborhood, living facilities, and

19:44mobile data using a 100x 100 meter

19:48spatial grid. In step two, we perform OD

19:50metric analysis using OD pairs located

19:54within a 5 kilo radius and calculate the

19:5820 minute walking distances with OSRM,

20:02the Python library to evaluate the

20:04spatial reachability. Finally, in step

20:07three, we calculate the spatial weights

20:08using boost posility accessibility and

20:11mobile population data. These weights

20:14are often applied to identify daily life

20:17zones through community detection using

20:20the lubing algorithm. We compare two DG

20:23outputs, one based on accessibility and

20:26the other based on mobile movement

20:28pattern.

20:30This step outlines the data

20:32prep-processing

20:34process with within our analysis

20:36framework. We use a 100 by 100 m spatial

20:40grid as the as the analysis unit

20:43overlapping three core data sets

20:45population data neighborhood living

20:47facilities and mobile user movement data

20:50facility coordinates matched to grid

20:54sales and mobile data originally in 50 m

20:58unit is aggregated to align with the 100

21:01m scale. We then select the grid based

21:03on the presence of population facilities

21:07or the mobile population data. This

21:11spatial structure structuring ensure a

21:14consistent and comparable data set for

21:16the further analysis in daily lipo zone

21:20modeling.

21:22In step two, we conduct a matrix

21:25analysis to understand the spatial

21:27connectivity. We define odiparas as grid

21:30cell centroidid that are within a 5 kilo

21:34meter radius and reachable within 20

21:36minute on foot. To do this we

21:38pre-process the grid cells set origins

21:41and destination and use OSRM opensource

21:45routing machine to compute walking

21:48walking travel distance travel times

21:50between cells. Finally, we extract OD

21:53pairs that meets the time and distance

21:55criteria which become the basis for

21:58constructing the OD metrics used in

22:01community detection and accessibility

22:03analysis.

22:05In step 3-1, we calculate the waiting

22:08values to perform community detection

22:10using the lubing algorithm. Two types of

22:13waiting which are applied. Post is

22:15postulate accessibility measured using

22:17the two SACA method. Twostep floating

22:21catchment area. This assesses the

22:24special balance between supply and

22:26demand by evaluating

22:28how much population each facility can

22:31serve and how accessible it is to

22:34different grid sales. The more

22:35facilities and the fewer computing users

22:38a greedy sales has, the higher its

22:41accessibility score. Second is the

22:44mobile mobile based weight which

22:46reflects the volume of population flow

22:48between grid cells using audi data from

22:52mobile telecom users. We extract the

22:55origin destination parasol recorded

22:57between 8:00 a.m. and 8 p.m. and

23:01calculate travel time within a 20 minute

23:04walking distance.

23:06This weight are normalized and then used

23:10to generate two versions of daily life

23:12zone.

23:15One based on accessibility and the other

23:17on mobile movement pattern.

23:21This slide explain the analysis

23:23framework used to derive daily life

23:25drone or DLG. The process consists of

23:29three main steps. First we extract

23:31relevant spatial grid sales based on

23:33population facility or mobile data.

23:38Then we analyze a pairs within a 5 km

23:40raders or 20-minut walking distance.

23:43Step three, we calculate weights using

23:45facility accessibility and mobile

23:47population counts and apply the Lubang

23:50community detection algorithm to

23:51identify DGS. Two types of DGS were

23:54created, one based on accessibility and

23:57one based on telecom mobile data. We

24:00adjusted the resolution to 1.1 to ensure

24:04both DG types produced a similar number

24:08of communities. After running the

24:10algorithm 100 times

24:13one 100 times to account for randomness,

24:16we selected the final result with the

24:19high highest modularity normalized

24:22mutual information NMI and consensus

24:26score. Lastly, small communities were

24:29merged to improve clarity and

24:32interpretability.

24:35In in step 3-3, we explore three main

24:39approaches for solving special

24:42optimization problem. First, linear

24:44program offers mathematically rigorous

24:47solutions and is

24:51effective when the problem has clear

24:53linear structures. Tools like Groby,

24:56Clex and Pulp are commonly used for this

25:00method. Second, curistic and

25:02mathematical approaches such as genetic

25:05algorithms, simulated analing and tab

25:08search don't guarantee optimality but

25:11are practical under time or resource

25:14constraints and are highly flexible for

25:17complex real world problems. Third, deep

25:21learning and reinforcement learning

25:23approaches including reinforce and tools

25:26like spanet and leonet use neural

25:29networks to learn decision paralysis.

25:32These are particularly useful for larger

25:35scale problems with irregular spatial

25:38pattern.

25:40This slide compares the life zones DGS

25:43derived from three sources. the

25:46government to government existing plan

25:49accessibility based analysis and mobile

25:52data. We observe that DGs based on real

25:55data often extend beyond beyond the uh

26:00administrative boundaries better

26:02reflecting the actual activity ranges of

26:05residents. In fact, multiple DGs can

26:08appear within a single or dimminitive

26:11tibu dong which highlights low special

26:15alignment with the government defined

26:17zones. Geographic element such as roads,

26:21rivers and terrain along the travel time

26:23also significantly influence DG

26:27boundaries. This indicate the importance

26:29of using real world mobility and service

26:32access data when redefining community

26:36zones for open planning.

26:39This analysis shows that daily life zone

26:42or DG often do not match administrative

26:44boundaries. Jones defined by

26:46accessibility or mobile data extend

26:48beyond the official bound official

26:50borders reflecting when people actually

26:53move and live. We also see that one

26:56administrative dome can include several

26:59DGs highlighting the need for more

27:01flexible planning. Geographic features

27:04and working time like rivers,

27:08roads and terrain play a major role in

27:10shaping these jones. This slide compares

27:14government defined DGS with those

27:17generated from data using IOU values to

27:21measure spatial similarity in

27:23residential focused area. We observe

27:25high alignment with between datadriven

27:28DGS and official plan which supported

27:30their use in managing infrastructure.

27:33However, however, in industrial and

27:35mixed use zones, the accessibility based

27:38DG shows low spatial alignment.

27:41Interestingly, the mobile based DLG

27:44display higher similarity especially in

27:47areas like the the southern part where

27:51the IOU which is 0

27:53uh 8. This suggested the mobile

27:56population patterns can better represent

27:58real activity zones in complex urban

28:01settings.

28:03This slice compare compares the

28:06government defined living zones with the

28:09datadriven DGS while accessibility and

28:12mobility based DGs generally align where

28:15they often uh diverge from

28:19administrative zones. For example, in

28:21areas in the the middle uh middle part,

28:25the government designated them as a

28:28separated zone, but data shows they

28:30function as a single continuous living

28:32area. This suggests that planning based

28:35solely on administrative boundaries may

28:38overlook how resident residents

28:41actually move and use urban facilities.

28:44Additionally, we observe a positive

28:46correlation between the residential land

28:48use ratio and the special overlap IOU

28:51indicating that databased DG better

28:55reflect reflect uh live the urban

28:58reality.

29:00In conclusion, daily life zones or DG

29:03are dynamic spatial unit shaped by real

29:06world mobility patterns that fixed

29:09administrative boundaries. Our analysis

29:12shows that datadriven DGS often extend

29:15beyond the dome boundaries or form

29:19multiple zones within a single dome

29:22indicating the need for planning based

29:25on actual human activity data in

29:27residential area. Higher residential

29:30land area land usage ratio correlated

29:34with uh stronger special congruence

29:38between different D types. This

29:40highlights the role of physical

29:42accessibility and facility placement in

29:44stabilizing DG structures. Datadriven

29:48methods can support more adaptive DG

29:51planning and help close the gap between

29:54auditative planning and real behavioral

29:57patterns ultimately enhancing spatial

29:59equity and urban connectivity.

30:03For for the for future work, we propose

30:05expanding the spatial and temporal scope

30:07and including external population such

30:10as inbound commas and visitors to

30:13validate the DG model more

30:16comprehensively.

30:19Thank you.

30:22Thank you very much for the presentation

30:26and um I just now want to ask and invite

30:30Rammon to present his research as our

30:34second talk of the session.

30:38>> Thank you Marian. Let me to share my

30:40screen

30:43with you.

30:46Will you confirm that? Are you seeing my

30:48presentation?

30:49>> Yeah. Uh

30:50>> yes.

30:51>> Is it the AI to improve public services?

30:54>> Correct.

30:55>> Perfect.

30:55>> Yes, this is the cover. Perfect.

30:57>> Yes.

30:58>> So thanks uh Mariam for your kind

31:00introduction introduction and for your

31:02perfect pronunciation

31:04>> and thanks to ITU and the rest of the

31:07organizers to invite at this interesting

31:10webinar. My name is Marian Satis Ram.

31:13I'm CTO at the Metropolitan Public

31:15Company AMB inform.

31:18Um from works for the Barcelona

31:22Metropolitan Area Administration,

31:23division administration that manage

31:25services for the 36 municipalities

31:28including Barona city. For instance, the

31:31mobility and public transport services.

31:33Here you see the main competencies like

31:35the entry public transport network

31:37management, taxi licenses, bike and

31:40scooter sharing services, uh park and

31:43ride or the management of low emission

31:44zones.

31:46So uh let me show this table. AMB

31:49Informasio our company develops uh

31:51technical solutions and digital tools

31:53for these services. Um and we are in

31:56charge of managing multiple sets of data

31:58and information related to this. As you

32:01can see here in the table with some data

32:04volumes managed by organization

32:06uh like the administration of concession

32:09card for the senior citizens uh to

32:11access with benefits with benefits at

32:13the Barcelona public transport network

32:16uh where we are managing more than

32:18400,000 users or the public transport

32:22services covering more than 6,000

32:256,000 uh bus trips per year or handling

32:295,000 service disruption every

32:32So last uh the registration for access

32:34at the control uh access and control to

32:37the low emission zones where we are

32:39registering more than 2 million vehicle

32:41plates every week in our data lake. So

32:44we are um or we have the framework to

32:47manage and and take out values from from

32:49this data of course.

32:52Now let me introduce a specific project.

32:54It's called IDA project. uh is a

32:56solution that we implemented for the

32:58Barcelona Barcelona IIO shuttle bus IDA

33:02is a project built around machine

33:04learning uh what we often call

33:06traditional AI and it's designed to

33:09train an algorithm capable of predicting

33:11passenger demand. Our goal uh was to

33:14provide bus operators with accurate uh

33:17datadriven insights uh so they can adapt

33:20services levels dynamically and improve

33:23both efficiency and passenger

33:25experience.

33:27We decide to start the project uh on

33:29this particular bad lane because it's

33:31not conventional urban route. Instead,

33:34it functions as a shuttle service

33:36connecting Bastrona airport with the

33:38city center, which means that the demand

33:41patterns are highly variable.

33:44They depend on, for instance, flight

33:46schedules, uh seasonal peaks, uh tourist

33:49flows and even exceptional events. All

33:53these factors make the line an ideal

33:55candidate for a predictive AI model. So

33:59as I said h the main objective of the

34:01project is to develop and train a

34:03machine learning model capable of

34:05forecasting demand by uh by

34:08understanding and anticipating passenger

34:10volumes uh we are able to optimize fit

34:13allocation reduce waiting times and uh

34:17ultimately make the entry transportation

34:19system more resilient and efficient of

34:21course but beyond that uh we also use

34:24this project as an opportunity to

34:25analyze the impact of external factors

34:28on demand. This include this includes uh

34:31variables such as weather conditions,

34:33flight schedules, seasonal tourist

34:36patterns, etc. By incorporated

34:39incorporating this external data source

34:41into the model, we can gain a much

34:43deeper understanding of uh the

34:46underlying dynamics that shape demand.

34:50So uh in this model in our model we

34:52integrated historical demand data uh

34:55using several Kmetrics like the trip

34:58validation ticket sales the number of

35:01bus trips operated and any service

35:03depion that occurate. In total we built

35:07a data set containing uh two full years

35:11of this historical information. But as I

35:14said we didn't uh stop there. uh we also

35:17incorporate these external variables by

35:19training the model um on this

35:22combination of internal and external

35:24data sets. We were able to obtain a much

35:27accurate prediction. As I said, we test

35:30different machine learning algorithm

35:32evaluating their performance using

35:34metrics like the RS squared and

35:36eventually selected gradient boosting

35:38progressor as the most uh reliable model

35:41for this scenario.

35:44Now I don't want to spend too much time

35:46on on the technical details or or this

35:49for example this heat maps graphics that

35:51you that you can see here but the key

35:54message that I want to share is that we

35:56now have a tool capable of predicting

35:58demand of this service with about 85%

36:03accury one week in advance. So we have

36:06this possibility with this predictable

36:07tool to obtain this um forecast the mind

36:12with this 85% accuracy.

36:15This mean that our public mobility

36:17technicians or also the bus operator now

36:20have a complimentary datadriven insights

36:22to help them plan service more

36:24efficiently improving the operational

36:27quality.

36:29So at this point I would like to share

36:31the three main conclusions of this

36:33machine learning project project. The

36:36first one is uh that there are

36:37significant difference between

36:39generative AI and traditional AI or

36:42machine learning.

36:44uh of course the generative AI helps us

36:46to automate task and improve

36:48productivity but uh machine learning

36:50solutions create direct value for public

36:52and private organization

36:54because they extract they they extract

36:57extract uh insights from their own

36:59internal data.

37:02The second conclusion uh there is a

37:04clear need to truly believe in and and

37:06commit to that a data-driven strategy

37:09uh a datadriven strategic transform an

37:11organization with uh with the data and

37:14the core decision making is not easy. It

37:16requires it requires addressing several

37:19key aspects like security, data quality

37:23and availability also very important.

37:25Not only the the quality of the data if

37:27not also the availability, the access

37:29control at this data anonymization uh

37:32standardization, deploy technical

37:35platform and infrastructure. Um maintain

37:38analyze the economic viability of of

37:40this implementation the legal and

37:43ethical frameworks. acknowledgement

37:45training of your teams. So in short uh

37:48solid data governance is essential to

37:50become a data driven uh datadriven

37:52organization.

37:54And the last conclusion the third one is

37:56that implementing these machine learning

37:58solutions uh requires a significant

38:00investment

38:02both in time and in a specialized pent

38:05and resources. So these three points

38:08reflect the main learnings uh from our

38:10project and highlight what organizations

38:12must consider when adopting machine

38:14learning technologies.

38:16I always like to to show this sentence

38:18because it's very clear that that about

38:20the idea that I want to transmit. It's a

38:23sentence from Mr. Edward the mind is a

38:26American consultant that works with

38:28Japanese government at the end of of the

38:31second world war and that he said that

38:33in God we trust all others must bring

38:36data and let me add uh all others must

38:39also be data driven as as I said.

38:43Okay. Now uh let me show you a high

38:47level overview of our data platform from

38:49a technical technological perspective.

38:52The goal of this light is to give you a

38:54sense of the complexity and the

38:57importance also of managing both data

38:59and metadata across different domains.

39:02What we use is a data mesh architecture.

39:05So you can see here in this in this uh

39:08scheme that our data mesh architecture

39:10which you can think of as a hybrid

39:12between a traditional data warehouse and

39:14data links. This approach enables us to

39:18hand data ingestion, ETL processes and

39:22exploitation of data through business

39:24intelligent tools and AI developments.

39:27To generate real value, it's essential

39:29to build architectures capable of

39:31integrating data from multiple domains

39:33in a unified and consistent way. This

39:37cross domain integration is what allows

39:39us for example to understand what

39:41happens uh for instance when a park and

39:44ride services uh what happens in in a

39:47park right services when we implement a

39:49new low emission zone uh or for instance

39:52how bus ridership increase when a bike

39:55sharing station is temporarily out of

39:57service. So at the end is understand um

40:00in integrating view these uh different

40:03domains

40:05this capacity to connect and analyze

40:07data across domains is key to supporting

40:09better decisions and a smarter mobility

40:12uh services.

40:14So now uh allow me to show you some

40:17examples of our business intelligent

40:19tool and this uh tools is available

40:23because we implement this data

40:24government strategy and this technical

40:27platform that I show you here. For

40:29example, you can see the key K

40:33indicators of the metropolitan bus

40:35services operated through different

40:37concessions with private operators.

40:40As you already know um with this visit

40:43intelligent tools we can integrate real

40:45time data. We can select the tempor

40:48temporal granularity that we want if we

40:50want to see data from one day or one

40:52week or we can inspect details by

40:55municipality or by bus line or and so on

41:00at the end. So keep in mind that our

41:02public managers for instance administer

41:05dynamic contracts uh dynamic public

41:07contracts sold by public tender with the

41:10operators with private operators with an

41:13economic penalties and bonuses depending

41:15of the quality of the service that they

41:17provide. So our public technicians are

41:20managing these dynamic controls

41:23monitoring these services and make

41:24decisions through this vision

41:26intelligent tool that uh you show here

41:29in these slides. Here you have more

41:31examples and for instance uh the request

41:34and operations for uh delivery freight

41:37for loading and uploading of goods in

41:39the city. So we can see by municipality

41:43we can see in a data in a maps etc.

41:48And now to finish uh let me very briefly

41:51share with you some of the main projects

41:53that we have been in developing with a

41:55generative AI because I show a use case

41:59on the example uh done by traditional AI

42:01as I said or by machine learning

42:03solution but we're implementing also

42:06projects uh through genetic AI. Um this

42:09is uh for example a chatbot built using

42:12open AI APIs. This is the typical rack

42:16uh uh chatbot based in Iraq architecture

42:20based system that provides a specialized

42:22information about the metropolitana

42:24service services.

42:26Um one of the biggest efforts in this

42:28project was preparing the vector product

42:30knowledge base. So the model now could

42:34answer citizens accurately within the

42:36specific domain that we needed and that

42:38we specialize the the chatbot.

42:41Today the service handles more than

42:443,000 questions every month and we are

42:47seeing a satisfaction rate uh of over

42:5092%

42:52based on the responses that it provides.

42:54No.

42:55And of course following the

42:58requirements uh of the European Union AI

43:02act we assess the risk of the solution.

43:05We clearly informed users about the

43:07technology behind it and about the

43:10possible limitation in its source or the

43:13possible mistakes and we have also

43:17published an algorithm transparency

43:19sheet uh for this tool in our website.

43:23Now this is another project. This

43:26project is about automating an

43:28administrative process using generative

43:30AI. So for forage vehicles entering to

43:35Barcelona city must register to obtain

43:38permits to access at the low emission

43:40zones.

43:41And in this case for Spanish vehicles we

43:44retrieve the data via the license plate

43:48from the general Spanish vehicle

43:50registration. So with the number of the

43:53plate in the Spanish vehicles we can

43:55check uh or we can look for this

43:57information in the general Spanish

43:59vehicle registration but this

44:01information is not shared between

44:03European Union countries. So with a

44:06number plate of uh France or for

44:08instance from Italy that comes to

44:10Barcelona we cannot know what is the

44:12mechanical or power or or power train

44:15characteristics of this vehicle to

44:16process the permit to access at the low

44:18emission zones.

44:20So now the process the French or Italian

44:23or Belgium driver who want to register

44:25in the low image zone have to upload the

44:28vehicle technical sheets and afterwards

44:32administrative staff of our company had

44:35to manually check the technical sheet to

44:37determinate whether the vehicle was

44:38electric hybrid or Euro 6 diesel diesel

44:42for instance. Uh but this is the old

44:44process. Now we have implemented a model

44:48that um doing a little bit of joke uh we

44:52call this model as the AI party because

44:55we put in the same tool a Google

44:57Microsoft and OpenAI to work together in

44:59this same process where um in an

45:03automated workflow we use Google Vision

45:05to extract data in JSON or in text um uh

45:10format from the photos or the PDF uh of

45:13the technical sheets that the user sent

45:16We run the data store anonymization

45:18process with precedio Microsoft solution

45:21and then we structure the response for

45:23the back office using uh open AI APIs.

45:28As a result, our public administrators

45:31um

45:33not longer validate the technical sheets

45:35one by one. Instead, they supervise the

45:38result of this automated process,

45:40drastically improving process

45:42efficiency.

45:45And last to finish and as I'm a member

45:47of of AI working group from the UITP is

45:51the international organization of public

45:53transport. Uh I want to share with you

45:56this uh this publication. We uh we

45:59publish uh this is a a book of different

46:03use case based on AI in the public

46:05transport from different countries and

46:08you have here the link to to look for

46:10this publication. It's very interesting.

46:12Here you have some examples. I will show

46:14you very quickly because it's not over

46:16exempted. It's done by by other entities

46:18like uh for instance in Singapore you

46:20can see this uh digital sign uh language

46:24avatar for the death. So with genetic AI

46:28they are giving information with sign

46:31language avatars for the deaf users in

46:34the public transport network or for the

46:37visually impaired users um using also AI

46:42generative solution to learn about the

46:44bus stops what is the information of the

46:47arrival time or the bus lanes etc. they

46:50developed these projects or for instance

46:52a chatbot for for the staff at the side

46:55chamber in done in Netherlands. [snorts]

46:58So that's all uh I think that I'm time

47:01thanks for your attention and I will add

47:04your questions after the webinar.

47:07>> Thank you very much for the amazing

47:09presentation

47:11very interesting and I think maybe our

47:13audience also will have some questions

47:15to ask. So we want to uh have some

47:20minutes just uh dedicated to take the

47:23questions. If anyone has any type of

47:26questions they can use um there is this

47:30um button on the platform

47:33uh I think it's called video wall that

47:36you can use like there's a tab on the

47:38platform that you can use to post your

47:40questions there. Also we have one

47:43question from one of the audience Dr.

47:45Natalie Aris

47:47um who asked Dr. Choy if um he's if do

47:53you see any opportunities to collaborate

47:55with other initiatives and city networks

47:59for instance um Milan urban uh food

48:02policy pact which is a network of 330

48:07cities working on food policy and

48:10developing sustainable food systems that

48:13are inclusive, resilient, safe and

48:16diverse.

48:18So um Dr. Troy is it something that you

48:21would or you want to consider?

48:24Yeah, the the originally this project is

48:29funded by the national R&D program and

48:33we are the priority is on the the

48:37applying this approach to the local

48:41government of Korean city local

48:43government of Korea but the S institute

48:46as the secretary of ETA mega city tank

48:50alliance and also the global

48:53collaboration In terms of global

48:55relation, we are open to the the other

48:58cities to collaborative on on this

49:02algorithms. Yeah,

49:05>> perfect. Thank you. So,

49:09do we have any other questions from the

49:11audience that we want to go to or we

49:14should move um to the next question like

49:17to the next um speaker?

49:23All right. So if we don't have any

49:25questions, I think we can move to our

49:28next speaker

49:30and

49:32um

49:36let me see if there is any other

49:37questions here. No. So Dr. Kimon Jang, I

49:43think the floor is yours and we're

49:45waiting to hear about your research and

49:48talk.

49:50>> Thank you. Uh let me share my screen.

50:00>> Perfect. We can see your screen now.

50:03>> Okay. Well, uh hello all. Uh my name is

50:07Kimun Jane. You can call me Ki. Uh thank

50:10you for having me for this webinar

50:12series uh from AI for good. Today I will

50:17be talking about uh generative AI uh not

50:21going too much deep into the technical

50:23details but it'll be more of an uh like

50:27an exploration of how generative AI

50:29technology uh pretty much an open source

50:32formats can be used for various urban

50:35tasks and what are the good what are the

50:37uh major implications when it is to be

50:39applied in uh metropolitan planning or

50:42city planning in general. Um so we all

50:46know about chachi it has been released

50:49uh pretty much about like three years

50:51from three years ago and uh it is

50:53dominating the world and it's not only

50:55about chip but since uh the release of

50:58GPT models there has been bunch of

51:00different kinds of genative AI or large

51:02language models in a more technical term

51:05that uh has been published in different

51:07kinds of uh their own uh also being

51:12published in different kinds

51:13of uh private services that you can also

51:16subscribe yourself for various different

51:18kinds of uh use cases. Um and then it

51:23has been tested since then uh for its

51:26application in different domains. Uh one

51:28of the most famous videos when the GPT4

51:31was released was when it was tested for

51:33tutoring services. Uh so it can now have

51:36a multimodality capabilities

51:38understanding the images like hand

51:40drawings uh interpret the information

51:42out of that and then um uh also being

51:45able to interact with natural language

51:47prompings. Uh there's also online quiz

51:50solver. So now people are coding it uh

51:52making a plug-in that can be embedded in

51:55your web- based platforms so that it can

51:57search through the web automatically and

51:59get you the the most relevant

52:01information of your interest. And it has

52:04also tested for various kinds of

52:05professional exams especially uh the

52:08well-known cases is that it has passed

52:11the bar exams uh to become an attorney

52:13or also the medical uh the medical

52:15school exams uh to become doctors but as

52:19urban planner or urban designers I was

52:21wondering what it would be like if it is

52:23applied to uh any kind of city planning

52:26tasks. So that is going to be the main

52:29question the underlying question for uh

52:31using generative AI for urban uh for

52:34urban uh pro uh practice.

52:37So I'll today briefly introduce about

52:39two uh previous projects that I've

52:41worked on. Um starting off with this one

52:44titled multimodal large language models

52:46as built environment auditing tools. So

52:49this will be a more of a practical uh

52:51exploration of how it can be used in a

52:53very specific task of built environment

52:55auditing.

52:57And uh the built environment includes

52:59the physical makeup of the the cities uh

53:02especially of the streets uh that which

53:04is one of the main focus when it comes

53:06to uh built environment auditing and

53:09auditing can be done for very specific

53:12use cases such as to assess the

53:14infrastructure elements uh or maybe very

53:18strictly focusing on whether the streets

53:21are walkable, whether the streets are

53:23bikable uh how well is the food

53:25environment uh prepared. prepare for the

53:27city residents and so on and so forth.

53:29Um so how it has been traditionally done

53:32is uh these field audits. So pretty much

53:35the city government people would go out

53:37in the streets with their checklist uh

53:40that looks like this. This is a built

53:42environment auditing report particularly

53:45for bike ability. As you can see from

53:47the list there's a bike rack

53:48accessibility, bike rack availability

53:50and so on. But the problem here was that

53:53this uh people going out on the streets

53:56checking all these information are very

53:58labor intensive sometimes timeconuming

54:00and very costly especially when you want

54:03to scale up at a very large scale. Uh

54:05and it especially for metropolitan

54:07cities it becomes much much more

54:09timeconuming and costly. Uh probably

54:13virtual audits were uh one alternative

54:15uh to solve these issues. So they try to

54:18have these GoPro or maybe any sort of uh

54:21visual information that they can collect

54:23uh of the streets and take it back to

54:25the office and then look at the images

54:27or videos and then use that for the

54:29auditing processes and then maybe even

54:32further than that would be how if you

54:34can either automate it uh with the

54:37pre-developed deep learning based

54:38computer vision methods. So this is one

54:41uh uh study that has tested the

54:43availability of uh image segmentation

54:46models in detecting specific objects or

54:48specific urban urban elements that may

54:51be or that is commonly in included in uh

54:54normal bu uh the street environment

54:56assessment tasks. So as you can see from

54:58here is assessing whether there is a

55:00tree uh whether there's a street walk uh

55:03and then there's a crossroad and so on

55:05and so forth.

55:06However, this also has a a remaining

55:09challenge in that it requires technical

55:11expertise and of course it requires a

55:14lot of uh access to computing resources.

55:16Um we're not going to make all the city

55:19governments be uh very well equipped

55:21with Python capabilities uh learning

55:24about YOLO technologies uh applying deep

55:27learning for these simple tasks.

55:30So and then now this was done last year

55:33when the the GPD 40 model was released

55:36and looking at its and given its

55:38multimodality capabilities we are

55:40interested in if it can interpret images

55:44as much as possible uh as conventional

55:46deep learning technologies and then if

55:48so pretty much we assume that it can be

55:52used uh in a very applicable manner to

55:55conduct built environment auditing tasks

55:56in a uh uh also very uh readily

56:00available. So if we find so we believe

56:03that this such a user-friendly tool can

56:05be very much applicable regardless of

56:07the technical competence of users and uh

56:10thus reducing the time and cost involved

56:12in such auditing tasks. Um so what we

56:15tried to do was we prompted uh these two

56:19different models uh GPT40 and Gemini Pro

56:22models uh with this natural language

56:24prompt saying I'll provide you with a

56:26photo of a streetscape from the United

56:28States and please detect the following

56:3010 objects in the image. So these are

56:32the 10 objects that we prompted in uh

56:34which pretty much could be included uh

56:36very likely to be included in auditing

56:38tasks such as vegetation like trees,

56:40vehicles, bicycles, roads, sidewalks,

56:43benches, trash bins, street lights and

56:45traffic signals. And then uh having a

56:48little bit of more awareness of how

56:50these tools are now operating. We try to

56:52also give a more structurized uh uh to

56:56output in a more structured format such

56:58as saying if the object is present

57:01respond as one. If it is not present

57:04respond as zero. If the image itself is

57:07invalid so that you cannot interpret

57:09anything out of it. uh uh output nine so

57:12that we can automatically filter it

57:14afterwards during the post pro uh

57:15post-processing uh steps. So these were

57:18uh uh uh this is a prompt and we try to

57:21uh have have Columbus, Ohio which has a

57:24very clear urban and rural uh uh

57:26gradient in terms of their geography as

57:29our case study. So we we did uh we did

57:33it for 2,000 Google Street images in

57:34Columbus, Ohio and compare it with uh

57:37deep learning based image segmentation

57:38methods and surprisingly these LLMs

57:43[clears throat]

57:44uh showed

57:46showed over 90% agreement with uh the

57:48the deep learning based uh image

57:50segmentation methods. uh even you mean

57:54the the average was 9.9 for tragic PT

57:56and 9.7 which is very much similar

57:58between the two uh but there were still

58:01some considerable variations in the

58:03agreement scores when you looked at

58:04individual built environment elements uh

58:07as you can see from here the tree

58:09bicycle and roads which are very much

58:11obviously uh being uh able to be

58:14detected nearly showed a perfect match

58:17uh with the deep the deep learning

58:19technologies

58:20uh almost uh it approaching 100% whereas

58:25in specific elements there was uh chacha

58:28being better in some cases Gemini being

58:30better and in some cases both uh not

58:33reaching 100% but still exceeding 70 or

58:3680% which is still considerably uh uh uh

58:39a good result

58:42but uh we also observed uh that there is

58:44still a potential geographic disparity

58:47in the usability of such um methods. So

58:51uh what we try to do was we try to do a

58:53global morons eye test. So this is a

58:55spatial clustering method to see if a

58:57high results cluster among themselves

59:00and low results cluster among themselves

59:02which were pretty much true only for the

59:05case of using Gemini. So in case of

59:08tragic t it it had a more had a

59:10relatively even distribution of the

59:12results which means that even in the

59:14downtown even in the rural uh the

59:16outskirts of the city the the results

59:18were pretty much dissimilar. However,

59:20when using Gemini, high results

59:22clustered mostly in the city center,

59:24whereas if you go out in the periphery,

59:26the results were uh going uh uh lower

59:30than expected. uh which means that uh it

59:33does underperform in specific regions of

59:35the city particularly uh it will be

59:37where it's less urban uh less populated

59:41and if so the people in there or or for

59:44for interpreting the uh doing the

59:47auditing task for images there pretty

59:49much would not be performing as well as

59:51um possible. So we should be also aware

59:54of such um uh uh limitations. But still

59:58overall [clears throat]

59:59uh we we feel that uh the these LLMs are

1:00:03actually more accessible to users and

1:00:05can actually de democratize access to uh

1:00:08advanced tools uh that [clears throat]

1:00:11once only the experts could do for built

1:00:13environment auditing such as using the

1:00:14deep learning technologies for that

1:00:16task. However, uh now this can

1:00:18democratize the the such task for urban

1:00:21planning uh without any coding skills

1:00:23and uh without any high high performance

1:00:25computing resources and and by doing so

1:00:28it can also benefit these local

1:00:29governments or the city planners as

1:00:32[clears throat]

1:00:33uh it can it can be done in a very easy

1:00:35way

1:00:36just by prompting uh with natural

1:00:38language text and then uh we try to we

1:00:42try to suggest a potential need to shift

1:00:44the focus toward assuring data quality.

1:00:46So it's no more about the models. It's

1:00:47no more about the spec specific methods

1:00:50to do the task. It's about how can we

1:00:52guarantee that there is good quality

1:00:54good amount of data uh uh evenly

1:00:57distributed across the city so that not

1:00:59only the city center it is benefiting

1:01:01from such technology but also uh it is

1:01:03just an equitable equitably benefiting

1:01:06the entire city as a whole. So probably

1:01:09this will be another task that um the

1:01:11data collection uh at a regular scale to

1:01:14to update the street images at a very

1:01:16high quality will be an important task

1:01:18for the city governments and that will

1:01:20be uh that we believe that that will be

1:01:21a very important assets uh in the

1:01:24future. Uh so this is more of an a

1:01:26practical exploration how it can be used

1:01:28in a specific task that the city

1:01:31governments would do in the future.

1:01:33However, uh we try to also ask even

1:01:37though it is powerful, even though it is

1:01:39promising, would it be actually used? So

1:01:41what would be the actual uh factors that

1:01:44people would consider if they were to

1:01:46accept these technologies?

1:01:48Uh and this is actually a known notion

1:01:52uh as promise promise gap. Uh which

1:01:55means that there's a gap between the

1:01:56perceived and realized benefits of

1:01:58technology. And here according to the uh

1:02:01uh global city leader survey last year

1:02:03by deote uh they surveyed these global

1:02:06city leaders with uh this question um

1:02:10how how are you feeling that these

1:02:13technologies are powerful enough or

1:02:15effective compared to how would you

1:02:18actually use it within the threeear span

1:02:21and uh on the on the rightmost column is

1:02:23the generative AI technology and 58% of

1:02:25student users actually expected that's

1:02:28um uh it is powerful. You know, there's

1:02:30potential to uh uh uh use it. However,

1:02:34uh only 1% currently view it as very

1:02:37effective. So, this disconnection uh uh

1:02:40implies that there's a there implies the

1:02:43historical patterns in in planning

1:02:45support systems when adopting a new

1:02:47technology. So, it would be powerful,

1:02:49but would you actually use it? So this

1:02:51has been a an ongoing debate when a new

1:02:53technology is coming in and if the city

1:02:56governments are actually trying to adopt

1:02:57it full uh for practice. So so having

1:03:00this prompt scap also in the case of

1:03:02generative AI we try to ask these

1:03:04questions which capabilities of visual

1:03:06genai are perceived as most useful for

1:03:09city design workflows and how would

1:03:11these perceptions shape attitudes toward

1:03:13adoption. So what are the perceptive

1:03:14qualities uh that is important for the

1:03:16adoption and then how would hands-on

1:03:18engagement with these tools shape

1:03:21planning community members uh uh shape

1:03:23their perceptions and attitudes toward

1:03:24the adoption in a participatory manner.

1:03:28So what we did was we we designed this

1:03:30uh uh workshop uh at MIT this which was

1:03:34a which was a three-week or three-time

1:03:36workshop uh one per one per week uh

1:03:39where we invited participants to uh

1:03:42freely join regardless of their

1:03:44familiarity or experience with engaging

1:03:47with uh the generative technologies uh

1:03:49especially Dali which is a visual

1:03:51generative AI of of of open AI using GPT

1:03:54models and we asked them to uh uh design

1:03:59a street based on like their uh uh their

1:04:03their interests. So try to design a

1:04:05street that is much safer, much livable,

1:04:09much beautiful and so on and so forth.

1:04:11Uh and then after all that we try to do

1:04:13uh we based on a technology acceptance

1:04:16model we try to survey the perceived

1:04:18usefulness, the perceived ease of use

1:04:21and their attitudes toward using such

1:04:23technology. uh and this was done by of

1:04:26with the members of the academic

1:04:28planning community. Uh the interesting

1:04:30thing was that uh the initial findings

1:04:33was that familiarity strongly increase

1:04:35post intervention which is very obvious

1:04:37because you you after the intervention

1:04:39you get more familiar and then also the

1:04:41perception metrics did increase for a

1:04:44strong uh uh if you have higher

1:04:46proficiency. So if you know generative

1:04:49AI or if you are aware of it uh your

1:04:51perception metrics do increase in a

1:04:53positive tone. Uh the more interesting

1:04:56part was when we had an interaction term

1:04:58between time and proficiency together.

1:05:01Uh it was found that lower proficiency

1:05:02users did show stronger interest and

1:05:05engagement in city design after the

1:05:06intervention. So we were able to

1:05:09increase their engagement with these new

1:05:11tools if they were actually having a

1:05:13lower proficiency from the beginning.

1:05:15But who were already high proficiency

1:05:17participants saw the tool as more

1:05:19indispensable which means that they knew

1:05:21already the tool that it has some effect

1:05:23it has some impact it can be powerful

1:05:25tool but after using it that kind of

1:05:28reinforced that mindset and said oh

1:05:30maybe this can be actually really

1:05:31necessary. So it's the difference

1:05:33between whether who had low proficiency

1:05:35and high proficiency and what they value

1:05:37more in engaging with these technologies

1:05:40and how they try to adopt it and overall

1:05:43and after all that we also try to look

1:05:44at the uh uh uh relationship between uh

1:05:48their perceived usefulness and their

1:05:50attitudes toward adoption. uh it was

1:05:53found that it was the cognitive and

1:05:55participatory dimensions that were more

1:05:57effective whereas the operational

1:05:59perceived usefulness were actually not

1:06:01significant in pre-intervention stage

1:06:04and even after the intervention it was

1:06:06very low. So the conclusion here was

1:06:09that uh the cognitive and participate

1:06:11patterns did matter the most while

1:06:13operational concerns are far less

1:06:15influential which uh for example were

1:06:18the accuracy or realism. So it's not

1:06:20only the performance that it's it's not

1:06:22actually the performance or how well the

1:06:25outputs look like. It's how the users

1:06:27would feel or uh feel or more engaged to

1:06:30while using these tools. Um so and then

1:06:34and then through having these engagement

1:06:36participatory

1:06:37workshops we believe that hands-on

1:06:39engagement can also make it more easier

1:06:42for people uh to to to feel more

1:06:45familiar and and feeling a positive

1:06:47attitude toward adoption and by and by

1:06:51finding these uh uh uh con uh results we

1:06:55feel that visual geni tools do have

1:06:57strong potential for educational

1:06:59purposes as well. So it is trying to

1:07:01lower the barrier for firsttime users

1:07:03cuz it is actually the low perfection

1:07:05users who have who feel even a stronger

1:07:08interest and engagement after all uh

1:07:10these these survey experiences.

1:07:13So these are two fun um experiments that

1:07:16we did while trying to use how it can be

1:07:18adopted in either built environment

1:07:20auditing or city design workflows. But

1:07:23is it only promising? uh we do find we

1:07:27do want to suggest some uh uh some some

1:07:29limitations or some uh important

1:07:31messages behind uh to use it in a in a

1:07:34good way. Uh one one reason is we find

1:07:37that there is a sign of stigmatization.

1:07:40Uh we we prompted midjourney to say um

1:07:44to to to to

1:07:46draw an image of a residential apartment

1:07:48building in Brazil and this was how it

1:07:50looked like. However, if we add favlla,

1:07:52which means informal sediments, it just

1:07:55suddenly degrades the quality of the

1:07:57built environments. Similarly, we try to

1:08:00uh add aski

1:08:02uh the street level scenes in Boston,

1:08:04the residential area of the white

1:08:06community, which look like the very nice

1:08:08brownstone uh housing areas. However, we

1:08:11just change white to black suddenly

1:08:13degrades the built environment uh with

1:08:15all the cracks uh unmanaged the the

1:08:18bushes, the the the street poles, the

1:08:21wires and so on and so forth. These uh

1:08:25the white and black is just a mere is

1:08:26merely a color. It doesn't necessarily

1:08:28imply any uh bias behind it. However, it

1:08:32is reinforcing the stigma behind what we

1:08:35might have as humans and is actually

1:08:37portraying

1:08:39and and we all know that the outputs are

1:08:41also can also be uh re reed into the

1:08:45training process. So it may have a power

1:08:48to reinforce such stigma that we as

1:08:51humans have in the real world. And

1:08:53there's also a geographic bias in the

1:08:55responses when we prompted with

1:08:57environmental justice questions of each

1:08:59uh of the contiguous US counties. Uh

1:09:02this was how it looked like. So the blue

1:09:04ones were where it did output successful

1:09:07results for environmental justice issues

1:09:09of the corresponding county and the red

1:09:11was when where it did not respond uh uh

1:09:14properly and it it has a very clear

1:09:17division in the geography. So the the

1:09:18the coastal areas uh uh uh were were

1:09:22performing much much more well much much

1:09:24well. Uh and it was found that counties

1:09:27that are more rural and poor have a

1:09:29higher chance of not receiving local

1:09:30specific responses uh when using chachi

1:09:33particularly when prompting for

1:09:35environmental justice issues. So for

1:09:36example, if the local government is uh

1:09:38uh people try to use tragic for their uh

1:09:41in practice in real life asking for what

1:09:44may be the the local issues here, they

1:09:47might not get the good results in

1:09:48specific areas. So there would be people

1:09:50who are benefiting with these tools,

1:09:52there would be areas who are not

1:09:53benefiting with these tools. So how

1:09:55should we address these issues? So these

1:09:56are some of the biases, some of the

1:09:58limitations of these tools that I would

1:10:01try to equally highlight and uh try to

1:10:03raise a question so that we can all

1:10:05collaborative think about together to

1:10:06see uh to to discuss how we can use this

1:10:09in a better way for city city planning

1:10:11tasks. So this is truly a black box. We

1:10:14don't know how it's operating fully.

1:10:16However, I was trying to show through

1:10:18this presentation what are the good

1:10:19sides, what are the promising sides. In

1:10:21the meantime, what are the limitations

1:10:23and biases that we should be very

1:10:24carefully uh looking into at the same

1:10:26time. So, I do not want to put an answer

1:10:29for this for now, but I would like to

1:10:30open up the discussion uh in the in the

1:10:33in the in the Q&A part and trying to uh

1:10:35have a more balanced talk about these

1:10:38tools when it comes to uh being adopted

1:10:40for city planning and design tasks. So,

1:10:43uh thank you.

1:10:48>> Thank you, Kimo.

1:10:51And all right, so before we go to the

1:10:56question and answer,

1:10:58um I'm going to talk very briefly about

1:11:01my research since we don't have that

1:11:03much time. I'm going to like um

1:11:08just dedicate a very small part to this.

1:11:13Um let me share my screen.

1:11:17All right.

1:11:22And

1:11:41I hope that you all can see my screen

1:11:43right now.

1:11:51Okay,

1:11:53if you can see my screen, it would be

1:11:55great to just give me some signs, but um

1:11:59I'm going to go through with this.

1:12:03Perfect. All right. So

1:12:07with the amount of information being

1:12:09continuously collected, it is very

1:12:11tempting to think that we now have the

1:12:13world at our fingertips and we can run

1:12:16any analysis we desire. But if we look

1:12:19closer, we see a large gap in the

1:12:21available data at the global scale.

1:12:24While every corner of every large US

1:12:26city is captured in Google street level

1:12:30images, the streets of capitals of many

1:12:33of the countries in the global clouds

1:12:36where their names are not even labeled

1:12:38on the maps.

1:12:41Even within the data affluent nations

1:12:44such as the United States, we see a

1:12:46clear inequity in what data we have and

1:12:49whose interest it represents. Data

1:12:52availability drives where decisions are

1:12:54focused. For instance, while a third of

1:12:57the US population are non-drivers,

1:13:00federal transportation funds have

1:13:01historically been skewed towards

1:13:03automobile and overlooked active

1:13:06transportation.

1:13:07This pattern of systematic

1:13:09disinvestment, we can see it in the

1:13:11normal lives of people. For instance,

1:13:14these are some of the tweets and talks

1:13:16and experiences of day by day of people

1:13:22dealing with these type of

1:13:23infrastructure that we have.

1:13:26And the problem is not only limited to

1:13:29the surface or like inaccessibility.

1:13:32There is a lack of reliable data on

1:13:34where sidewalks, food paths, and

1:13:36crosswalks are. Most navigation apps

1:13:40still rely on the road networks that

1:13:42assumes that there are sidewalks on both

1:13:44sides of the road. And as we all know,

1:13:47it's not a correct assumption

1:13:48specifically in the United States.

1:13:53Having access to all services or cities

1:13:56have to offer is a human right.

1:13:59Is driving a choice or is it a surrender

1:14:02to the current circumstances?

1:14:05Is it possible for an adult non-driver

1:14:08to rely on the current condition of the

1:14:12aeron infrastructure to address their

1:14:14daily needs? To be able to answer these

1:14:18type of questions, we need to be able to

1:14:20do a comprehensive analysis of our

1:14:22pedestrian infrastructure.

1:14:24And any comprehensive analysis of

1:14:26pedestrian infrastructure needs a

1:14:29three-fold understanding of where

1:14:31sidewalks are, how they are connected

1:14:33and what their condition is. While we

1:14:36have sensors all around us collecting

1:14:39data and even like the types and um

1:14:42species of the trees and the numbers of

1:14:44them, the data that we have from

1:14:46sidewalks is hardly enough to draw any

1:14:49informative picture.

1:14:52This lack of data leads to us not being

1:14:56able to assess the condition and then we

1:14:59cannot address something that we know it

1:15:01is it extent and we cannot know what the

1:15:05problem is and to know that we need

1:15:08data. So my in my research my goal and

1:15:12aim is to address this lack of data and

1:15:16to tackle it through providing

1:15:18open-source and easy to use tools for

1:15:21people to use and on CTS scale I'm using

1:15:25aerial imagery and on human scale I'm

1:15:27using street level imagery. I'm going to

1:15:30only talk about the CTS scale where I

1:15:32use aerial imagery and I created this

1:15:35open source tool called tile to net

1:15:38which is an endto-end open-source tool

1:15:40for creating pedestrian networks from

1:15:43autorectified aerial imagery. I use

1:15:46computer vision model to detect

1:15:49sidewalks, footpaths and crosswalks from

1:15:52high resolution aerial imagery. And

1:15:54despite the complexity of the source

1:15:57code, we created a very userfriendly

1:16:00interface where the user can run the

1:16:03whole system for the areas that we

1:16:05support only using just one line of code

1:16:07from command line.

1:16:10And when tileet was published, it really

1:16:14received like lots of uh attention and

1:16:16the attention was a testimony to how

1:16:19much this tool was needed, not that much

1:16:21how good the tool was. And um the tool

1:16:25has been recently added to the ArcGIS

1:16:28um living atlas and has been downloaded

1:16:31around like 18k times by now.

1:16:36And briefly this is the pipeline of how

1:16:38tile 2 works. It takes it has like two

1:16:41different fe like main parts feature

1:16:44detection where you like where we have

1:16:47like this we train this uh computer

1:16:50vision semantic segmentation model that

1:16:53can detect these features from the

1:16:55aerial imagery and these features are in

1:16:58raster format. Then we take these raster

1:17:00predictions and we georreerence them

1:17:03creating maps that has like the uh

1:17:06geometry and also like the information

1:17:09of where they land on the surface of the

1:17:11earth. You have a polygon and then from

1:17:13polygon we create the line

1:17:15representation.

1:17:17And here is the result. And as you can

1:17:19see because this model was trained on

1:17:22the um planetric data we can see that it

1:17:27could actually like overcome lots of the

1:17:29obstruction that we see from the shadow

1:17:31or vegetation.

1:17:33And also this is the result of applying

1:17:35the model on the tiles from the Boston

1:17:38common. And you see the aerial imagery.

1:17:41In the middle there is this polygon

1:17:43data. And then on the um right panel we

1:17:46have like the line representation

1:17:49and this is the city scale application

1:17:51of tile to net over Boston, Manhattan

1:17:54and Washington DC. And the strength of

1:17:58the model here is that it can actually

1:18:00like um tackle this on a larger scale

1:18:04with a lower cost.

1:18:11Perfect. All right.

1:18:15So

1:18:17now I want to invite everyone to first

1:18:21like our audience to send their question

1:18:24and also for us to get together for a

1:18:27and get ready for a like brief

1:18:30discussion.

1:18:37All right. So we will like the session

1:18:40will end in around

1:18:42like 11 minutes. So we have 11 minutes

1:18:46to do the discussion together.

1:18:53If we have any question from the

1:18:55audience I'm happy to uh share it right

1:19:00now.

1:19:05I have a question for you.

1:19:08So, we have a question and I'm going to

1:19:11read it out right now. So, the question

1:19:14read,

1:19:16I'm currently developing a method for

1:19:18generating building in project areas.

1:19:22However, not all generative methods have

1:19:25metrics that can be used to evaluate the

1:19:27generated buildings themselves, but only

1:19:30for example the quality of the generated

1:19:33image,

1:19:35not the building themselves. Do you have

1:19:37any case metric selecting for the

1:19:40evaluation performance of generative

1:19:43methods in relation to the generation of

1:19:45physical object such as buildings? The

1:19:48most obvious metric for me right now are

1:19:50F, building density, etc.

1:19:56All right. So, I wanna um give the floor

1:20:00to Eric right now and um I think he can

1:20:04lead the discussion from this session

1:20:06on. Eric,

1:20:08>> thank you very much for this uh all the

1:20:11presentation very um very uh technical

1:20:16and with different topics that is uh

1:20:20that is useful for I can say all cities

1:20:23at all scales. Um so it's u um related

1:20:30to the um to the data analysis and to uh

1:20:34and how to to use AI in a prediction

1:20:38manner to uh to to to deliver better

1:20:41services for cities. So this is very

1:20:44interesting. [snorts] Um I would like to

1:20:46um to ask the uh the panelist on the on

1:20:50the scale because we're we are here in

1:20:53the webinar of meta so looking for mega

1:20:57cities and how we can use this analysis

1:21:02at the scale not at the scale of the

1:21:05neighborhood of local scale but how we

1:21:07can amplify the uh this the tools you

1:21:10are de developing regarding the at the

1:21:13scale of the mega cities. what are the

1:21:14the main concern or the main tools that

1:21:18we can imagine using LM and the and AI

1:21:23to modelize to to prepare policies or

1:21:28territorial strategies at the scale of

1:21:30of mega cities. I think this um because

1:21:33when we speak about mega city were

1:21:37something which is more complex than a a

1:21:40small city and so probably there is some

1:21:43other methods to to develop really

1:21:46specific for this very complex bodies.

1:21:49So

1:21:51perhaps the panelists can elaborate on

1:21:52that please.

1:22:01Yes. Uh I think I think I can uh start

1:22:04answering to that question. Uh thank you

1:22:06for uh raising a very important question

1:22:09especially when it comes to the mega

1:22:10city context how it can scale up uh in

1:22:14the mega city scale. Uh related to the

1:22:18presentation that I did uh for my

1:22:20projects uh I mentioned about the

1:22:22importance of now the data quality. I

1:22:24think the mega city governments can

1:22:26focus on uh how to how to ensure that

1:22:29there is a very good quality data set um

1:22:32at a large scale and that's that's where

1:22:35I think the the big city governments can

1:22:37uh can can contribute to in securing a

1:22:41very good asset for urban analysis at a

1:22:44massive scale. Now the models especially

1:22:47because I presented about the LLMs

1:22:49models can be openly accessible or uh

1:22:52through subscriptions you could be able

1:22:54to fine-tune it yourself for specific

1:22:56purposes but still uh if the open models

1:22:59are still very powerful in terms of the

1:23:01performances it's now how like which

1:23:03city or which government is having a

1:23:06very good quality data set at a large

1:23:07scale. I think that's where big city

1:23:09governments can contribute to more. Uh I

1:23:11I know that Amsterdam uh our lab has a

1:23:14very good connection with the Amsterdam

1:23:16Institute AMS. Um, and we learned a lot

1:23:19from the Amsterdam city that the city

1:23:21government is putting a lot of efforts

1:23:23to have their own like street view

1:23:26imagery data set at a citywide scale uh

1:23:29a LAR point cloud data set that they the

1:23:31city government people or the the

1:23:33researchers that are inhouse in the city

1:23:35governments themselves collect and

1:23:37gather and and and pre-process it for

1:23:40that that in in a quality that is

1:23:41readily accessible for research. Um why

1:23:44this is important is because um also in

1:23:47the street field imagery realm people

1:23:49are trying to diverge from relying on uh

1:23:52like private sector such as Google maps

1:23:54because then if so we always have to ask

1:23:57for like API requests and sometimes they

1:24:00are even deprecating the APIs that uh in

1:24:03one day we might be we might be having a

1:24:04limited access to the data and if so the

1:24:07research capabilities would decrease at

1:24:09some point. So they're trying to uh uh

1:24:12gather their own assets that they have

1:24:14their own rights of usage. So I think

1:24:16that's where the big the governments can

1:24:18also contribute to a lot. Another is

1:24:20that very briefly the another is that I

1:24:22mentioned about the educational purposes

1:24:24especially for the low proficiency

1:24:25users. Um I I believe that if the

1:24:28government can also offer some

1:24:30educational programs uh so that

1:24:33firsthand users can have their their

1:24:35very first experiences of these new

1:24:37tools or or of these AI technologies. It

1:24:40will be very helpful for them to get

1:24:42more engaged and be more willing to

1:24:43adopt the technologies in the future.

1:24:46>> Thank you.

1:24:48other panelists.

1:24:50>> Yeah, maybe from my side Eric I'm not a

1:24:53technical specialist maybe that like Mr.

1:24:56Jang but in our case um working with

1:25:00genative AI maybe we try to segmentate

1:25:04make make a segment a segmentation of

1:25:06the data sets no because as you probably

1:25:08know the large language model can works

1:25:12around a one million tokens of context

1:25:14window no so so maybe it's preferred to

1:25:17to make a segmentation of of these uh

1:25:20data sets but for machine learning u

1:25:23model developments

1:25:25I think that in some projects we use

1:25:28clustering techniques in machine

1:25:30learning projects it's more easy to uh

1:25:32make a more uh big context but but uh

1:25:36you will need more time of of process of

1:25:40your of your CPUs of your machines but

1:25:42you can use uh clustering uh techniques

1:25:45for for this kind of of AI traditional

1:25:48AI solutions.

1:25:52>> Yeah.

1:25:53Interesting.

1:26:00Yeah. I I would like to um talk about

1:26:04the the usefulness of this this kind of

1:26:09approach to the mega city um the the the

1:26:14deal with the mega city problems that is

1:26:17not just the the extend the approach

1:26:20distend approach using the more larger

1:26:24scale data but we we need to find some

1:26:28the

1:26:30fill the gap between the administrative

1:26:33planning and the the this kind of

1:26:37mobility based planning. That means we

1:26:40can uh uh we we can uh find the some the

1:26:47discrepancy or the special

1:26:50uh the unequality place and by detecting

1:26:56those kind of unequal place that the

1:27:00service is not evenly provided. It will

1:27:03very helpful by this kind of datadriven

1:27:06mobility based approach we can adapt to

1:27:10then you can be helpful to many the mega

1:27:13cities.

1:27:15>> Thank you Miriam.

1:27:19[snorts]

1:27:20>> So yeah thank you.

1:27:22Um

1:27:26I think we only have 2 minutes right now

1:27:29and yeah so like my very brief answer is

1:27:34I think

1:27:36like a big chunk of it depends on the

1:27:38quality of the data that we collect

1:27:41because right now I think one of the

1:27:43main problems of many of these let's say

1:27:46LLMs and the large language models that

1:27:49we have is garbage in garbage out and we

1:27:52don't have any data quality control over

1:27:54what is going in these models and what

1:27:57type of biases is going to be you know

1:27:59instill in these type of models and then

1:28:02what are the type of biases that they're

1:28:04propagate again after being trained on

1:28:08all of those data. So maybe having some

1:28:10filtration and some quality control will

1:28:13help us get way better result than

1:28:16focusing mostly on the you know like

1:28:20making them like 1% better than what

1:28:22they are right now. So this is I think

1:28:25where the specifically for the fields

1:28:27where we are working with humans and

1:28:30with the very sensitive data it is where

1:28:32our focus should be. Thank you.

1:28:36You can go to the question from the from

1:28:38the uh the audience.

1:28:41Yeah, actually I wanted to ask the

1:28:43audience that I read it their question

1:28:45to actually email me because this is a

1:28:48question that we can discuss and this is

1:28:50something that I work on so I can help

1:28:52them maybe like with the question of how

1:28:54to evaluate because this is a field that

1:28:57I'm actively working on developing new

1:29:00evaluation metrics for these models

1:29:02because we are not right now we don't

1:29:04have those evaluation metrics available

1:29:06for many of these type of analysis. So

1:29:10um yeah you can email me and maybe I can

1:29:12help you with that. Um and I think yeah

1:29:16we we are at the hour. So just as we

1:29:19close I wanted to underline that uh you

1:29:21know aside from like always having the

1:29:24machine learning and AI which is

1:29:26fantastic and it is helping us a lot to

1:29:28scale up our analysis. One thing that we

1:29:31should not forget is we need the human

1:29:34experience and expertise in the loop of

1:29:36training and in the loop of development

1:29:38and also having this connection between

1:29:41the research institute and different

1:29:44interdisciplinary researchers. And I

1:29:46think this is precisely the role that

1:29:48the mega city think tank aliens want to

1:29:51play to bring us like and all of these

1:29:55types of like a research centers

1:29:56together. And I hope that we see more of

1:29:59these happening across different you

1:30:01know interdisciplinary research because

1:30:03urban analytics in nature is a complex

1:30:07system that needs interdisciplinary

1:30:09approach in order to be you know like

1:30:12reaching a conclusion and a good um

1:30:15outcome with the all of these type of

1:30:19decisions that we want to make um

1:30:21whether using AI or whether using any

1:30:24other technique.

1:30:28Good.

1:30:31>> Is there any other questions? So you you

1:30:34want to we continue or because there is

1:30:37>> no I think I think there is no other

1:30:39question. I just checked. Okay.

1:30:41>> And with that I wanted to just close the

1:30:45meeting today. Thank you everyone for

1:30:47joining us for to our great panelist and

1:30:51to Dr. Choy and Eric and everyone. and

1:30:54I'm very happy to be here and hope to

1:30:57again collaborate soon.

1:30:59>> Very good.

1:31:00>> Have a great morning and every

1:31:01>> Thank you very much.

1:31:02>> Thank you.

1:31:04>> Thank you. Bye.