At The Fault Lines

On

February 6, 2023 at 04:17AM


a devastating M7.8 earthquake struck southeastern Turkey.

FEBRUARY 6, 2023 • 04:17AM

M7.8 EARTHQUAKE
STRIKES SOUTHEAST TURKEY

60,000

LIVES LOST

The combined death toll from these earthquakes is estimated at nearly 60,000 people, with more than 15 million people affected across 11 provinces.

230,000

BUILDINGS COLLAPSED

More than 230,000 buildings, housing 520,000 apartments, collapsed. Even newly built, luxury residences marketed to the public as “earthquake-proof” were reduced to rubble.

58%

OF THE HISTORICAL CITY OF ANTAKYA
WAS COMPLETELY FLATTENED

The cities of Hatay and Kahramanmaraş were among the most severely damaged by the February 6th earthquakes. Historic towns and city-centers, such as the millennia-old city of Antakya, were completely leveled, rendering the city unrecognizable to its residents.

30,000

SMALLER QUAKES AND AFTERSHOCKS FOLLOWED
ALL WITHIN 24 HOURS

Following the initial 7.8 magnitude earthquake, the region experienced relentless aftershocks—more than 30,000 in total. Some aftershocks reached magnitudes of 6.4, plunging survivors into a state of constant fear and uncertainty.

M7.8 EARTHQUAKE
STRIKES SOUTHEAST TURKEY

FEBRUARY 6, 2023 • 04:17AM

FEBRUARY 6, 2023 • 01:24PM

SECOND M7.7 QUAKE
OCCURS JUST HOURS LATER

9 Hours

TIME BETWEEN THE FIRST
AND THE SECOND EARTHQUAKE

The second major earthquake, considered a powerful aftershock, struck about nine hours later at 1:24 PM local time (10:24 UTC). This was a magnitude 7.7 earthquake centered northwest of Kahramanmaraş.

110,000 km2

TOTAL LAND AREA OF
THE EARTHQUAKE REGION

The disaster area covered a staggering mass of land, about 110,000 square kilometers (42,000 square miles). If the same earthquake struck the United States, the path of destruction would stretch from Washington D.C. to Boston.

$163.3 Billion

ESTIMATED COST OF DAMAGE

The cost of the disaster was estimated at over $163.6 billion, with the greatest damage across southern and central Turkey, as well as northern and western Syria.

CHAPTER 01

February 6

On February 6, 2023, two consecutive earthquakes with magnitudes of 7.8 and 7.7 struck southern Turkey and northern Syria, resulting in the deaths of more than 60,000 people. Now named as the Kahramanmaraş Earthquakes, largest since the 1939 Erzincan Earthquake of the same magnitude, is the second most powerful ever recorded in the region. Affecting 14 million people and leaving 1.5 million homeless across 11 provinces, it is now recognized as the deadliest natural disaster in the history of Turkey.

A Drone Shot Through the Streets of Antakya by NYTimes
A drone shot through the historic center of Antakya after the earthquake, showing the extent of destruction.
Visual investigation by the New York Times.

The earthquake region, encompassing the 11 southeastern provinces of Kahramanmaraş, Hatay, Osmaniye, Adıyaman, Gaziantep, Şanlıurfa, Diyarbakır, Malatya, Kilis, Adana, and Elazığ, totals up to over 110,000 square kilometers (42,000 square miles) - nearly the same size as the neighboring Bulgaria. This is equivalent to stretching from Washington, D.C. to Boston, MA, encompassing the entire northeastern metropolitan region of the United States.

US Northeast compared to Turkey earthquake region of the Southeast
The earthquake region is roughly as big as the Northeast United States
From Washington, D.C. to Boston, MA.

In Hatay province, the historic center of Antakya is now nearly uninhabitable. The city, which has stood for millennia, has been flattened, leaving it almost unrecognizable. The extent of the destruction has driven nearly all residents away, transforming familiar streets and landmarks into a stark, altered landscape. In Kahramanmaraş, the historic district has been decimated, leaving once-vibrant neighborhoods in ruins. Similarly, Gaziantep, known for its rich history and architectural heritage, has seen extensive damage to its old city center, including the Gaziantep Castle—originally built by the Hittite Empire and later expanded by the Romans—and its traditional markets. Despite more than a year of reconstruction efforts, most of these cities remain unrecognizable to more than 500,000 people who once called them home.

  • ANTAKYA, HATAY
  • ANTAKYA, HATAY
  • 12 SUBAT STADIUM, KAHRAMANMARAS
  • ISKENDERUN
  • OSMANIYE
  • ANTAKYA, HATAY

Yet, Turkey is no stranger to such catastrophic earthquakes, as the country geographically sits on top two major fault zones - the North Anatolian fault zone being one of the most seismically active in the World. 20 years ago when a similarly destructive earthquake struck near Istanbul, the country’s economic and population center, it was a wake up call for the country: The government changed shortly after being deemed “incompetent,” much stricter building codes and regulations were put in place, nation-wide first response organizations were founded, and a new “earthquake tax” was introduced to help fund urban renewal and recovery efforts. The recent earthquakes exposed that many of these efforts have been immensely inadequate, mostly due to large-scale corrupt political, regulatory, and construction practices throughout the 20-year AKP rule. Moreover, the country’s most prominent scientists have been incessantly alarming against a long-awaited “Istanbul Earthquake” to strike in the next decade, with consequences exponentially graver.

“The devastation we witnessed in Antakya could easily happen in Istanbul. We are sitting on a ticking time bomb.”

-Prof. Naci Görür, Geologist, Istanbul Technical University

This study aims to dissect these corrupt practices, identify their traces in the architectural level across an urban scale, and develop an alternative risk assessment mapping method that may offer critical insight for local authorities, municipalities, as well as individuals. Recent developments in interctive mapping techniques, artificial intelligence and machine learning have allowed large urban-scale risk assessment models to be implemented in order to detect possible structural defects before any failure or collapse. In this article, we will explain the structural malpractices that led to such destruction in the area, and will propose an automated, faster, and more scalable alternative “risk assessment method” that could be used to create an earthquake risk map of the bustling metropolis of Istanbul that is home to 16 million people.

In order to better understand this, we first need to zoom out and understand the country’s geological, social, and political relationship with the earthquakes — starting with the East Anatolian Fault, which is where the Feb 6 earthquakes happened.

The East
Anatolian Fault

Powerful earthquakes such as the ones that occurred on Feb 6 along the EAF may induce tectonic events along the neighboring faults.
The East Anatolian Fault is neighboring to East the North Anatolian Fault, which is considered one of the most active in the World.
The NAF has a track record of being triggered by neighboring tectonic events, which is why the recent quakes are concerning for many.

A Long History of
Devastating Earthquakes

Despite its relatively modest land area, Turkey has experienced numerous M7.0+ earthquakes since antiquity. Positioned at the intersection of the Eurasian, African, and Arabian tectonic plates and crisscrossed by several major fault lines, it is one of the most seismically active regions in the world.

CHAPTER 02

The North Anatolian Fault

A soft story structure diagram

The North Anatolian Fault (NAF) is a major strike-slip fault zone stretching approximately 1,500 kilometers (900 miles) across northern Turkey, from the junction with the East Anatolian Fault in eastern Turkey to the northern Aegean Sea.

Since the mid-20th century, seismologists have understood that earthquakes along the NAF move in a westward progression, with each quake triggering the next. This phenomenon became evident following the 1939 Erzincan earthquake, which resulted in over 32,000 fatalities and more than 100,000 injuries.

Research indicates that nine out of ten major earthquakes along the NAF between 1939 and 1992 were brought closer to failure by the preceding shocks due to the transfer of stress along the fault line. This progressive failure mechanism suggests that each large earthquake can set up the next one by altering the stress distribution along the fault.

A comparison of before 2002, after 2002, and the current state of disaster preparedness efforts in Turkey.
Feb 6 quakes along the EAF have caused the ground shift laterally by up to 3 to 7 meters.
Some of these ruptures could be seen via satellite imagery. Photo: Maxar technologies.

By studying the sequence of past earthquakes, seismologists can identify patterns and potential stress points, enabling them to predict future earthquakes. The most recent major earthquake along the NAF, the 1999 Izmit earthquake, which killed over 17,000 people, has led the scientific community to anticipate the next major earthquake hitting Istanbul.

Due to Istanbul’s proximity to the NAF, the warning time for a major earthquake could be as short as 2-6 seconds, making a rapid response challenging. Combined with its dense population of 20 million and inconsistent infrastructure quality, the city could face an unprecedented level of destruction.

A Century of Earthquakes

Over the past century, eight +M7 earthquakes occurred along the North Anatolian Fault Line. Each happening further west of the one preceding it.
The most recent of which is the 1999 Izmit-Gölcük Earthquake, which struck 50 miles outside of Istanbul, igniting widespread political, social, and economic change across Turkey.
Experts are concerned that the February 6 earthquakes could potentially trigger a future earthquake along the North Anatolian Fault (NAF) near or beneath Istanbul.

1999 Gölcük-Izmit Earthquake:
A Disaster Powerful Enough to Change Governments

The most recent major tectonic event along the North Anatolian Fault, the 1999 Izmit-Gölcük Earthquake has served as a catalyst for the country, causing widespread changes in the government, and brough new approaches to disaster management, urban planning, construction regulations, and raised the public's awareness about seismic risks.

CHAPTER 03

1999 Golcuk Earthquake

The most recent major tectonic event along the North Anatolian Fault, the 1999 Izmit-Gölcük Earthquake, struck on August 17 at 3:01 AM local time. With a magnitude of 7.6 and an epicenter just 90 kilometers east of Istanbul, this 37-second tremor claimed over 17,000 lives and displaced 500,000 people. As Turkey’s most destructive earthquake until February 6, it exposed critical vulnerabilities in the country’s disaster preparedness, many of which were due to rampant corruption in the construction industry.

The incumbent government’s organizational failure and inability to respond and provide help to the area, combined with a long-standing financial crisis, has left the public with a bitter sentiment that “the state has left them on their own.” As a result, the government received less than 10% of the vote in the 2002 elections.

A comparison of before 2002, after 2002, and the current state of disaster preparedness efforts in Turkey.
The states of disaster preparedness in Turkey,
informed and taxonomized by major tectonic events.

The Justice and Development Party (AKP), which still is the ruling party in the country, rose to power in the elections with Mr. Recep Tayyip Erdogan, back then the very popular and highly-acclaimed Mayor of Istanbul, promising drastic changes. The newly appointed government codified what was initially planned to be a short-term recovery act, the infamous “Earthquake Tax,” making it permanent. It also launched the country’s first nationwide disaster response organization, AFAD, fore-fronted much stricter construction regulations and building codes, and introduced massive-scale urban renewal programs.

Since 2002, the AKP has been able to consolidate its base largely due to initiatives that many believed had transformed Turkey from a run-down, old country to modern standards, with new highways, residences, and other infrastructure improvements. However, the February 6 earthquakes have led to widespread disillusionment. The devastation revealed that many of the government’s commitments were not fulfilled, exposing significant inadequacies and corruption in regulatory enforcement, infrastructural reconstruction and disaster preparedness efforts.

A comparison of before 2002, after 2002, and the current state of disaster preparedness efforts in Turkey.
Decades of failed urbanism and development policies have led to cramped streets, shoddy buildings, and very few open spaces.
Photo: Kostas Tsironis, Bloomberg.

This, along with the westward pattern of the seismic activity along the North Anatolian Fault suggesting an increased possibility of a tectonic event taking place around Istanbul, directly to the west of the most-recent quake zone Izmit, has left a city of 20 million worried about the safety of their homes and workplaces.

The 1999
Izmit-Gölcük Earthquake

The M7.6 earthquake lasted 37 seconds and took lives of more than 17,000 people, and displaced 500,000 others. It was the biggest that the country has ever seen until the Feb 6 earthquakes.
The devastating event has caused significant changes, such as the Ecevit government to lose in the following elections to AKP, which still holds the office today.
The AKP campaign promised significant changes to the country's natural disaster response strategies, including a new Earthquake Tax, and stricter construction codes.

The Impending Istanbul Earthquake

Decades of corrupt practices that led to the February 6 aftermath and warnings about an impending tectonic event, informed by historical patterns, now cause widespread concern about the safety of current buildings in a city of 20 million people.

Population Density

Home to a quarter of the country, Istanbul is operating way above its capacity as a result of failed migration policies.

State of Structural Testing

Only less than 15% of the buildings are tested. Many don't opt in due to potential legal implications.

Low Damage

The February 6 earthquakes has once again reminded the residents of 20 million city that the impending earthquake may happen anytime.

Medium Damage

The February 6 earthquakes has once again reminded the residents of 20 million city that the impending earthquake may happen anytime.

Heavy Damage

The February 6 earthquakes has once again reminded the residents of 20 million city that the impending earthquake may happen anytime.

Extremely Heavy Damage

The February 6 earthquakes has once again reminded the residents of 20 million city that the impending earthquake may happen anytime.

Estimated Loss of Life

Proejcted figures show that the densest neighborhoods are at the highest risk of loss of life.

CHAPTER 04

The Big Istanbul Earthquake

The devastating earthquake in February has heightened anxiety among Istanbul’s residents about the inevitable arrival of “the big one.”

A city of 16 million (20 million, including undocumented residents according to the experts) living in highly dense neighborhoods with apartment buildings of five stories or taller, many of which were built before the 1999 quake thus not up to the latest building codes. These older structures, more than 625,000 of them, are particularly vulnerable, says Mehmet Özhaseki, Minister of Environment and Urbanism citing studies conducted by both the city and Ankara government.

POPULATION

20
Million

1/4 of the Country Lives in Istanbul

BUILDINGS

1.2
Million

Only a fraction of them tested

EXPECTED TO COLLAPSE

+52%
Under Risk

625.000 Buildings Built Before 1999

Since then, the Municipality has conducted various studies that tries to simulate the extent of damage (both buildings, people, and infrastructure) and shared it on its Open Data portal as a spreadsheet. The below map is an interactive visualization of this data, showing the number of damaged buildings, injuries, death toll, and shelter availability per neighborhood. This study allows us to have an overall idea by ZIP code, yet it lack the granularity of a high-resolution map at the building level - which is critical, as the extent of damage varies per structure dramatically.

IMM Risk Simulation, Mapped


Hover over a neighborhood for more

Since then, the Municipality has conducted various studies that tries to simulate the extent of damage (both buildings, people, and infrastructure) and shared it on its Open Data portal in the form a CSV document. The above map is an interactive visualization of this data, showing the number of damaged buildings, injuries, death toll, and shelter availability per neighborhood. This study allows us to have an overall idea by ZIP code, yet it lack the granularity of a high-resolution map at the building level - which is critical, as the extent of damage varies per structure dramatically.

It may take up to 21 years to get
a structural test for your building.

The situation becomes even more dire as the wait time for structural testing can stretch up to 21 years at the current rate of 120 buildings per day. With the North Anatolian Fault indicating an imminent earthquake, local architects, engineers, and students have initiated alternative methods for identifying structural issues. Using makeshift lists of easily identifiable deficiencies, these grassroots movements on social media have become highly popular among the public with posts sparking conversations and sharing of information at the scale of millions in helping people make crucial decisions involving their safety. This and many other initiatives has also shown that the solution to Istanbul's testing problem had to be more agile yet granular, and easily scalable.

The Upcoming
Istanbul Earthquake

Researchers predict that the likelihood of a M7.0 quake to hit Istanbul in the next 5 years is 70%.
That is nearly 20 Million people, living in more than 1.2 million buildings.
625.000 of them are to be heavily damaged - that's more than 52%.

Although multiple studies have identified the most at-risk structures and communities across the city, only a few offer meaningful resolutions that can be effectively communicated to the public. The Municipality’s risk simulation study is available to the public as a spreadsheet through the city’s Open Data portal, but it lacks the interactivity and user-friendliness of a widely accessible tool.

Risk Simulation data, retrieved from Istanbul Open Data portal.
Risk simulation studies and datasets are open to public, although mostly in unfriendly formats.

However, some high-risk areas are well-known: The southern districts of the European side and the Princes’ Islands are at highest risk due to their proximity to the North Marmara fault lines. Similarly, low-income neighborhoods developed as “gecekondu”s (makeshift buildings constructed overnight by factory workers in the 1960s) are poorly documented, with many built before modern construction regulations. The contrast between well-established, high-income neighborhoods and these low-income areas is stark, with the lack of planning and open spaces visible even in satellite imagery.

Satellite imagery, average rent, and green space analysis chart

$1,587

$442

Levent Mahallesi

Çeliktepe Mahallesi

Satellite images retrieved by using Google Maps API.
Real estate data is retrieved from Zingat.com

Many planned high-income neighborhoods and the mentioned “gecekondu” areas are located side by side, often only separated by highways, hills, or high-rise buildings constructed as part of urban renewal programs. Most urban renewal projects have taken place in middle to high-income neighborhoods with higher profit margins, often overlooking low-income areas where people are more vulnerable and at higher risk.

Before
Atakoy Sirinevler

This is why we chose “Hurriyet/Sirinevler” as the pilot site for our project. Separated only by a highway strip from one of the city’s most well-planned housing developments, Atakoy, it frequently tops risk assessments and simulation studies. Many buildings here were constructed before updated regulations, sharing a common set of structural flaws and malpractices known to cause failures during earthquakes. We will explain these issues in detail in the next chapter.

//Anatomy of A Collapse

CHAPTER 05

Anatomy of A Collapse

The 1999 Earthquake was different in the way that it was widely studied and well documented. A detailed field investigation conducted by the Earthquake Engineering Field Investigation Team (EEFIT) offers particularly useful insights here, highlighting that many structures collapse in a similar way, commonly referred as “pancake collapse”, that occurs in buildings with poor structural configurations, specifically those with soft stories. Nearly 24 years after regulatory amendments were introduced to eliminate these outdated construction practices, numerous photos and videos from the February 6 Earthquakes reveal that even newly constructed buildings collapsed due to the same malpractices these regulations were meant to address.

Below is a frame-by-frame excerpt from a social media post depicting a building in Şanlıurfa collapsing due to soft-floor failure to better illustrate this concept.

MOMENTS OF A PANCAKE COLLAPSE CAPTURED DURING FEBRUARY EARTHQUAKES IN SANLIURFA, TURKEY.
VIDEO BY VOX MEDIA. SOURCE

This building type, ubiquitous across the nation, is a mid-rise residential structure with street level retail and auxiliary spaces. The upper floors jut out beyond the building footprint to increases residential square footage—a tax loophole that lets real estate developers maximize profits.

In the first few seconds of the quake, the building appears stable, with only subtle signs of stress, such as the increasingly intense shaking of the glass shopfronts. Soon, however, the ground floor begins to buckle under the weight of the upper floors. The collapse initiates at this level, causing the ground floor to crumple and lose its support, marking the moment when the soft-floor failure becomes evident.

Once the ground floor gives way completely, the now-unsupported upper floors begin to collapse in a domino effect, with each floor pancaking onto the one below. In a matter of seconds, the building is reduced to a pile of debris and dust.

//Common reasons behind structural failures

Common Reasons Behind
Structural Failures in the Region


Many videos and photos like the one above from the moment of collapse all hint at a common type of structural failure: First, the ground floor suddenly collapses, followed by all the floors above shortly after, pancaking on top of each other layer by layer in just a matter of seconds.


These “pancake collapses” are recognized as the primary cause of devastating consequences resulting from earthquakes in the region, such as the 1999 Golcuk earthquake, as reported by the infrastructure assessment team sent to the area. The factors leading to this phenomenon have been researched extensively and can be narrowed down to four primary construction/design shortcomings.



1. Soft Story Buildings

A soft story structure diagram

Soft story buildings are multiple story buildings that has a ground floor that has large windows, wide openings, and large open commercial shopfronts where a shear wall or vertical support element would typically be needed to ensure that the structural stability. Especially in earthquake-prone areas such as Turkey, at least 30% of the ground floor area must be dedicated to structural elements to ensure that the structural system could withstand the stress of the above floors. Soft story buildings are highly common in the region, with many multi-story apartment buildings accommodate commercial floor space on the ground floor, which is allowed by mixed-used zoning practices. The 1999 quake has led to an updated building code that prohibited such structures, many developers did not feel pressed to follow the regulations due to the lack of a strictly enforced control mechanism.


2. Heavy Overhangs

A building with heavy overhangs/protruding parts

Another common practice is excessive cantilevering above the ground floor, as this allows developers to go beyond the site land usage limitations to maximize profits over extra square footage. Although cantilevering structures are a common practice in architectural design, combined with other malpractices such as soft story building designs, or using lower-quality building materials significantly increases the structural stress on the ground-level structural system. This may lead to a structural failure in the moment of an earthquake, even when the building may be able to withstand this load under normal conditions.


3. Pounding Effect

A building with heavy overhangs/protruding parts

Pounding effect is a common cause of collisions between adjacent structures during seismic events due to insufficient gaps between buildings that share one or more walls. In fact, pounding effect accounted for over 40 percent of all structural failures, with at least 15 percent collapsing mainly due to this phenomenon. Similarly, 200 out of 500 surveyed structures after the 1989 Loma Prieta earthquake were damaged due to the pounding forces of adjacent buildings.

Although Turkish construction codes have clear guidelines for proper seismic gaps, this remains a widespread issue in many metropolitan areas worldwide due to financial and architectural constraints. In metropolitan cities, the gap is often narrow or nonexistent. For instance, statistics from Eskisehir, Turkey, show that only 36 percent of adjacent buildings are adequately separated.

In addition to seismic gaps, studies show that the distance between floor levels, floor slab misalignments, and mass differences between adjacent buildings contribute to the extent of damage.

4. Short Columns, Misaligned Sturctural Elements and Other Post-Construction Alterations

A building with heavy overhangs/protruding parts

Structural alterations such as removing vertical supporting elements on the ground floor to open up space for commercial activities is another common practice many building owners employ to maximize profitability. Although there are multiple laws and regulations strictly prohibiting this, many buildings that has such alterations go unnoticed as it’s mostly done in the inside of the buildings.


5. Usage of Prohibited Materials and Techniques

Sea sand mixed cement

The 1999 Earthquake has shown that the usage of certain construction materials, such as “sea-sand mixed cement” or rebars with no ribs, can be a significant risk factor during a earthquake. Although most of these materials cannot be identified without a structural test, some easy-to-spot visual hints can offer useful information. Many locals are advised to look for sea shells on the surface of the concrete to identify whether the cement is mixed with sea-sand.

"I come from the kitchen" in this business, as we say, meaning all my family, under my father, were also a part of this sector, the construction sector. I grew up working with him, witnessing all of the development at that time.

...

70 percent of all the settlements in Istanbul, I would say, are vulnerable to a major earthquake. Without the proper diagnosis treating a patient is not possible. The construction materials used for various settlements in different parts of Istanbul used to be of poor quality - I personally know this myself, as I was one of those who extracted, sold, and used this type of cement.



-Ali Agaoglu, Turkish Real Estate Mogul

//AI + Computer Vision

CHAPTER 06

AI + Computer Vision

The original idea for FaultLines came from a deceptively simple hypothesis: if architects and civil engineers could recognize common structural defects just from looking at a building, could we teach a computer to do the same based off images of buildings?

Before
An early study of the FaultLines algorithm. May 2023.

Stress testing using a physical sample has always been the gold standard in assessing the structural integrity of a building. Yet, many citizens don’t have the motivation to pursue this option due to years-long wait lines, potential legal implications, and municipal/bureaucratic backlogs.

Traditional risk assessment studies, nevertheless, mainly depend on human-generated datasets (that are highly accurate, but time and resource consuming to create) that mostly took place in rather static GIS environments. Besides, rarely do these findings are shared in an easy-to-understand, accessible way to inform the broader public.

CADASTRAL

GEOLOGICAL

LEGAL

IMAGERY

HISTORICAL

Patterns that allow an architect or an engineer to make an “educated guess,” could very well be detected by multi-modal image processing algorithms. We now have an ever-increasing amount of data about our cities, be it live bus schedules, maps with live-traffic feed, or panoramic images of virtually every street in any given city. This data —that we all are harnessing today, through tools like Google StreetView— combined with the recent advancements in AI and computer vision algorithms, allow us to make sense of the built environment in various new ways and combinations, at exponentially larger scales, much faster, and by using very few resources.

STREET LEVEL IMAGERY

StreetView images that our workflow intelligently collects and pre-processes for semantic image segmentation.

To test our hypothesis, we first had to find points at which two buildings were close or explicitly touching each other at one or more surfaces. In order to automate this, we developed a custom python script in QGIS that enabled us to query a Google StreetView image that showed only the adjacent buildings sharing a wall using the StreetView Static API. Once this was done, we rectified these images to mitigate any image-segmentation or calculation issues that might have occurred due to varying perspective levels in the StreetView imagery.

Windows from both sides of the adjacency line are (mostly) segmented for processing.
Windows of each building of both sides of the adjacency line are segmented for processing.

These rectified street images then had to be intelligently segmented by left and right buildings, as well as windows within each facade. To solve this, we developed a custom Geospatial AI workflow that builds on the Grounding DINO and Segment Anything models. (Both available publicly on HuggingFace and Meta AI) Combining these two models enables us to semantically process any given object (e.g. buildings, sidewalk, cars, or windows, doors, and walls) across Istanbul’s street imagery. For this study, we have further tuned this workflow to detect and calculate misalignments in adjacent buildings — which is one of five aforementioned visual cues that may hint at a critical structural deficiencies.

Windows from both sides of the adjacency line are (mostly) segmented for processing.
Facade segmentation enable us to make calculations - such as measuring the distance between windows,
which allowed us then to measure difference between floor levels.

Semantic image segmentation has allowed us to generate grouped masks for each window on either facade, which we used red and blue color maps for easier identification. Then, we created subgroups for vertical and horizontal arrays of windows, which allowed us to create an index of facade elements (such as making calculations using only a select group of windows by floors or bays) — which then further enabled us to measure distances from either side of the adjacency line, and compare those with the other side. This way we were able to generate an adjacency score — an integer value of the length difference between corresponding floor slabs of the left and right buildings.

Indexin and naming convention.
INDEXING AND NAMING CONVENTION THAT ALLOWED FOR
TRANSFERRING, TRACKING AND MATCHING ATTRIBUTES BETWEEN GIS AND ML ENVIRONMENTS.

We then simply saved the processed image with a naming convention (latitude, longitude, index numbers for buildings on both side of the adjacency line, as well as the adjacency score) that allowed for easy transfer of information between GIS and image segmentation environments. This way, we were able to match these numbers with their geospatial counterparts on QGIS, where we then able to visualize and see the results, as seen below.

Merging adjacency calculations back in QGIS for visualization.
MERGING ADJACENCY CALCULATIONS BACK WITH CORRESPONDING BUILDING IN QGIS FOR VISUALIZATION

In short, our workflow encompassed four main steps:

1. Finding street-facing points at which two buildings share one or more surfaces.
Query a street-level image of this adjacency line for further processing.
2. Semantically segmenting the image to find out floor slabs and distances based on street facing elevations.
3. Calculating the distance between floor slabs of buildings from both left and right side of the adjacency line.
4. Assigning this calculated score back onto the GIS object.

Evaluation


This experiment serves as a proof-of-concept for using zero-shot text-to-mask segmentation in conjunction with raster and geospatial data. For the scope of this study, we utilized a zero-shot image segmentation model. Consequently, we neither generated nor trained our model on any specific set of images or datasets. However, Meta AI Research indicates that the SAM image dataset SA-1B was trained using over 11 million images, with Turkey being prominently represented in the geographic distribution, contributing over 300,000 images to the dataset.

We evaluated the accuracy of zero-shot prompt-based image segmentation for estimating floor-misalignment-related pounding effect in adjacent buildings, across a series of metrics encompassing various processes applied during score generation, on a randomly-selected test group of buildings located in our pilot ZIP code, 34188. In order to achieve this, we compared machine-generated image segmentation results and floor misalignment scores with human-evaluated results across the same images in three distinct ways:

1. Annotators rated floor misalignment levels from A (perfectly aligned) to F (completely misaligned).
2. Annotators compared AI-generated results with corresponding human evaluations.
3. If an issue was found in the post-processed image or score, annotators specified the error based on a comprehensive list of points of failure, ranging from object obstruction to facade misindexing.

POST-PROCESSING
SUCCESS RATE

65.38%

SEGMENTATION AND POST-PROCESSING (E.G. FACADE INDEXING)

AI ESTIMATION
SUCCESS RATE

84.61%

EXCLUDING MISINDEXING & POST-SEGMENTATION PROCESSING FAILURES

POTENTIAL
SUCCESS RATE

97.11%

IN THE CASE OF SEGMENTATION SUCCESS RATE BEING 100%

Across these results, we have achieved a %84.61 success rate in semantic segmentation of facade elements, with the remaining samples could not be successfully segmented due to extreme facade irregularities, vehicle and object obstruction, extreme or miscalculated perspective, angle and/or heading prior to obtaining street imagery and low raster quality after image processing.

Despite challenges like facade irregularities and obstructions, the results highlight the potential of zero-shot models for practical applications in raster + geospatial use-cases, such as our experiment. Future improvements could leverage the recently released SAM 2.0 model by Meta AI, which offers much higher accuracy rates, and incorporate higher-accuracy geospatial datasets along with more up-to-date raster imagery to enhance segmentation precision.

Merging adjacency calculations back in QGIS for visualization.
Merging adjacency calculations back in QGIS for visualization.

INTERACTIVE MAP

By intelligently detecting adjacent buildings, and successfully calculating critical measurements across neighboring structural frames within a GIS environment, we were then able to share these findings on an interactive map interface, where the information was not represented with static geometries, but also were enriched with calculation results, explanations, and machine-generated process images. This way, we not only ensured that our workflow accurately represented real-life test subjects, but also these tests were held up to the judgement of the viewer for transparency.

CHAPTER 07

Interactive Map

After delving into the history of earthquakes in Turkey, examining corrupt construction practices, and exploring how AI can extract information from street view images, we present the culmination of this research: an interactive map.

This interactive map, the product of a year-long effort at the Center for Spatial Research, allows residents to access critical structural information about their buildings through a simple search interface. The map provides risk insights in a transparent, immersive, and easy-to-digest manner, embodying our commitment to accessibility and user-friendliness.

Initially available for Hurriyet Mahallesi, a densely packed pilot neighborhood with buildings constructed mostly before the 1999 Earthquake, the map offers risk insights on “misaligned floor calculations.” This area, with minimal open spaces or greenery, is notably highlighted in the IMM Earthquake Risk Simulation dataset for its high projected damage and death tolls.

Currently, the map covers selected ZIP codes, but it will gradually expand to encompass all boroughs and ZIP codes across Istanbul. This expansion will increase the number of tests performed and the data used to calculate the overall risk score for each building, ensuring more comprehensive coverage over time.

To use the map, residents can search for building names, door numbers, addresses, and more, limited to the available ZIP codes for now. By hovering over each building, users can access an in-depth risk report and a visual representation of the performed tests. This interface allows anyone to look up their address and access “machine-generated” structural insights.

Transparency is vital when dealing with AI-generated data, especially since much of it results from “blackbox” processes beyond our control. In developing this project, we made sure all computations made on the initial images were visible, explainable, and shareable. We clarified our aims with each computation and used language that encouraged users to critically assess every AI-generated result. This approach ensured that these “risk insights” were understood as fallible and provided citizens with our “training methods,” helping them recognize structural cues in their environments.

We hope this tool will serve as an insightful resource, helping residents understand the structural integrity of their buildings and prompting necessary actions to enhance safety. By making critical information accessible and transparent, we aim to empower residents to engage with their built environment and take informed steps toward improving their safety and resilience.

FaultLines PRE

Index Number: