Blog article: Abandoned Places And Where to Find Them (with Toronto’s Open Data)

Published on October 2, 2026

Guest post by Amely Su

Disclaimer

Urban exploration of abandoned places can be exhilarating and thought provoking but also dangerous; floors may be unstable, walls may contain hazardous materials, and places that may look abandoned could belong to someone. If you plan on pursuing urban exploration, make sure you bring the correct safety equipment and thoroughly research the places you plan to visit. Ensure you know the risks of each location. Never cause intentional damage; take only photos, leave only footprints.

Excellent online resources that can teach you to explore the right way include the Urban Exploration Resource (UER.ca) or this blog post on accessing abandoned, beautiful places to photograph.

Introduction

In this blog post, I’m sharing my latest project: a data-driven love letter to all the hidden beauty Toronto has to offer. I’ll be walking through how I set out to uncover some of the city’s beautiful abandoned spaces, with a little help from Toronto’s Open Data.

You are in the dark with nothing but a headlamp to light your way. Something crunches beneath your booted foot: plastic maybe, or glass. Your light catches the tangle of pipes overhead, stretching out like fingers, and you carefully duck through a doorway which hasn’t been new in fifty years.

An explosion of colour hits you. Graffiti art is splashed over the walls from floor to ceiling, some of it crude, some of it painfully intricate. You can see flowers, faces, and words simultaneously at war with each other and coexisting in harmony.

You approach a ladder and carefully ascend, watching the floor getting further and further away from you, and light, real daylight, spills across your face. You haul yourself onto the roof and remove your respirator, taking your first breath of fresh air in the last hour. You step out, unsteady from your odyssey below, and face the sun.

On the horizon is the shop where your mom used to work, and just beside it is the park where you and your sister used to play when you were children. Your whole world is before you. You watch it revolving around you and wonder how you could have gone so long in your life without seeing it this way.

That was a journal entry I wrote about my very first urban exploration trip in Toronto. Since then, I’ve gone on many more, and each trip has brought something new and worthwhile into my life.

I know that not everybody finds the thought of venturing into a potentially filthy, smelly, dangerous place appealing, but I find that something magical happens when a building becomes abandoned. Objects left behind become immensely interesting; you can’t help but wonder who used them. I once visited an abandoned office where everything remained mostly untouched for years. I could still see writing on the whiteboards where people made supply chain plans and pricing estimates. It was like venturing through some fantastic realm made all the more extraordinary by how ordinary it was.

Urban exploration changed the way I thought about handling my possessions. When you see an abandoned chair, you wonder why people left it behind. Same goes for when you find hundreds of CDs ruined by water, or hundreds of old computers filled with dust and mould.

Spaces evolve over time, too, and people (and animals!) repurpose them to fit their changing needs.

Urban Exploration with Open City Data

As a computer science student at York University, it is only natural that I bring my love of urbex and photography to my love of data. When I discovered the City of Toronto had a dataset of Active Building Permits, I wondered if I might be able to use it to find more beautiful places to photograph. It felt like it should, as the permits in the dataset tell us something about the buildings that are coming down in Toronto and those that will replace them. So, I chose to ask the following research question of the data:

Can the active building permits associated with my favorite derelict buildings help me find others that are equally beautiful?

The Dataset

A building permit is formal permission from the city to build, demolish or renovate a property. Abandoned or derelict places, the kind in my photographs, are often associated with a building permit of some kind, but inhabited buildings can have them too.

The Active Building Permits dataset contains features associated with all of Toronto’s open building permits, including the date the permit was issued, the work type that was proposed (e.g. demolition or plumbing work), its estimated cost and purpose, and the person or company responsible. The city typically issues about 9000 such permits a year, and there are more than 200,000 permits in the current database. Some of the most common features of the currently active permits in the database (as of September 2026), as well as descriptive statistics related to them, can be found in Table 1.

	All Buildings
Total permits	241,163
Date range	2005–2025
Most common year	2024
Avg permits / building	3.3
Avg units created / building	3.81
Mean est. cost	$1.42M
Cost range	$1 – $1B
Permit Types	Total (percent)
Plumbing(PS)	53,688 (22.3%)
Small Residential Projects	50,220 (20.8%)
Mechanical(MS)	44,317 (18.4%)
Building Additions/Alterations	34,752 (14.4%)
Drain and Site Service	18,632 (7.7%)
New Houses	16,444 (6.8%)
Fire/Security Upgrade	6,272 (2.6%)
Zoning Review	6,158 (2.6%)
Other	4,386 (1.8%)
Demolition Folder (DM)	3,524 (1.5%)
New Building	2,770 (1.1%)
Work Types
Building Permit Related(PS)	45,300 (18.8%)
Building Permit Related(MS)	41,923 (17.4%)
Other	40,676 (16.9%)
Interior Alterations	30,414 (12.6%)
Multiple Projects	22,241 (9.2%)
New Building	21,201 (8.8%)
Building Permit Related (DR)	13,562 (5.6%)
Addition(s)	5,903 (2.4%)
Backflow Prevention Devices (Water only)	4,315 (1.8%)
Other(BA)	3,721 (1.5%)
Demolition	3,524 (1.5%)
Other(SR)	2,860 (1.2%)
Garage	2,802 (1.2%)
Back Water Valve (Sewer only)	2,481 (1.0%)

Table 1. Descriptive statistics: permits in Toronto’s Active Permit Database. PS permits relate to plumbing work while MS permits relate to maintenance work. SFD stands for “single family dwelling”.

To better understand the relationship between these permits and properties that make for a good photograph, I extracted all permits connected to buildings in my prior photos from the permit database. This led me to 76 permits associated with 18 known, beautiful derelict buildings that have been my subjects in the past. I also identified 555 permits that were associated with 6 buildings I had previously visited but found visually unsatisfying. A comparison between these two sets of permits, for both the beautiful and unsatisfying buildings I visited in the past, can be found in Table 2.

	Beautiful Buildings	Unsatisfying Buildings
Total permits	76	555
Date range	2005–2025	2005–2025
Most common year	2025	2025
Avg permits / building	4.2	92.5
Avg units created / building	51.72	1.83
Mean est. cost	$25.99M	$670.5K
Cost range	$3.3K – $240M	$1K – $20M
Permit Types	Total (percent)	Total (percent)
Plumbing(PS)	12 (15.8%)	96 (17.3%)
Small Residential Projects	0 (0%)	0 (0%)
Mechanical(MS)	13 (17.1%)	119 (21.4%)
Building Additions/Alterations	9 (11.8%)	253 (45.6%)
Drain and Site Service	10 (13.2%)	9 (1.6%)
New Houses	0 (0.0%)	11 (2.0%)
Fire/Security Upgrade	0 (0.0%)	54 (9.7%)
Zoning Review	5 (6.6%)	1 (0.2%)
Other	9 (11.8%)	10 (1.8%)
Demolition Folder (DM)	12 (15.8%)	2 (0.4%)
New Building	6 (7.9%)	0 (0.0%)
Work Types
Building Permit Related(PS)	10 (13.2%)	95 (17.1%)
Building Permit Related(MS)	12 (15.8%)	115 (20.7%)
Other	19 (25.0%)	90 (16.2%)
Interior Alterations	3 (3.9%)	214 (38.6%)
Multiple Projects	1 (1.3%)	8 (1.4%)
New Building	10 (13.2%)	9 (1.6%)
Building Permit Related (DR)	4 (5.3%)	4 (0.7%)
Other(BA)	3 (3.9%)	8 (1.4%)
Backflow Prevention Devices (Water only)	1 (1.3%)	0 (0.0%)
Addition(s)	1 (1.3%)	10 (1.8%)
Demolition	12 (15.8%)	2 (0.4%)
Other(SR)	0 (0%)	0 (0%)
Garage	0 (0%)	0 (0%)
Back Water Valve (Sewer only)	0 (0%)	0 (0%)

Table 2. Charts summarising some of what the permits associated with my existing beautiful photos and unsatisfying photos are like.

While many features are similar between the two seats of buildings, the permits associated with beautiful photos proved much more likely to be associated with demolitions while the others were more likely to be associated with mild construction (e.g. Building Additions/Alterations or Security Upgrades). In addition, many of the permits associated with my more beautiful photos detail costly work that will create many new units; these buildings may give way to condos. The other permits, by contrast, relate to less costly work that won’t create very many new units.

Left: Gorgeous! Beautiful!

Right: Apartment buildings I visited but that were not exactly what I’m looking for, unfortunately.

Analysis

After identifying some features of permits that seemed to distinguish the better looking buildings in my existing photographs from the unsatisfying ones, I chose to try to encode some of this information in a decision tree. While decision trees are typically trained on much larger sets of data, I thought a decision tree might help me in this instance to encode some general “rules” for detecting beautiful buildings based on permit data.

A decision tree is a flowchart-like model that makes predictions by splitting data into subsets based on feature values until a decision about the properties of each data subset can be reached. In my case, feature values were related to permits for buildings and the decision was whether I might find the buildings beautiful. Many decision trees use a statistic called entropy to split data into subsets in a way that reduces uncertainty about the properties of data in subsets, like the beauty of buildings. Entropy measures how mixed up, or uncertain, data points in a subset are; for example, it is certain that any given building in a subset of beautiful buildings will be attractive, while the beauty of any given building in a subset with equal parts beautiful and unsatisfying buildings is much more uncertain.

To build my decision tree, I first aggregated information about all the permits associated with my prior photographs by building address. Aggregated features for each building included the most recent year a permit was applied for or issued, a sum of the estimated construction costs across all permits, the number of dwelling units created or lost as a result of the work, the estimated square area for specific categories of work (Industrial, Mercantile, etc.), and the percentage of building permits in each permit category for each building. Each building address was also associated with my own binary assessment as to the building’s attractiveness (beautiful or unsatisfying), and my decision tree was trained to distinguish the buildings in the two sets from one another.

My trained decision tree is illustrated in Figure 4. In the figure, orange leaves represent subsets of beautiful buildings and blue leaves represent subsets that are unsatisfying. Given the information about a single building’s active permits, you can in theory use this tree to make a decision as to whether the building is likely to be beautiful by walking from the root of the tree to an orange or blue leaf, applying the rules in the tree nodes along the way. As you can see in the figure, my tree indicates that beautiful buildings have fewer fire/security upgrade permits, fewer new house permits, and are less likely to be single family detached buildings. This is consistent with my observation that the beautiful buildings in my prior photos are less likely to be associated with permits for minor upgrades (e.g., fire or security related) and more likely to be related to permits that create large buildings containing many new units.

Left: a mall associated with 11 permits, designated as a mix of building types (Other, Retail Store, College/Trade/Tech School) due to its nature as a mall. Seven permits were for Building Additions/Alterations, two were for Mechanical (MS), one was for Plumbing (PS) and one was a Non-Residential Building Permit.

Right: a house from a block associated with two permits, designated as SFD – Detached buildings. Both were Demolition Folder (DM) permits.

Once my decision tree was built, I used it to identify other buildings in the Active Permits dataset that might be beautiful as well. To do this, I first filtered the Active Permits dataset to remove buildings I had already photographed and those not associated with a Demolition Folder (DM) permit. Ryan Persaud, Manager of Building Planning, Performance, and Intelligence at the City of Toronto, confirmed to me that buildings in the DM subset were all primed for a full demolition; based on the descriptive statistics in Table 2, it seemed clear to me that these would be more likely to be “beautiful” for the purpose of photography. I then aggregated permits by building address to create more than 3000 individual data points, and asked my decision tree to identify those that were considered, by the tree, to represent “beautiful” abandoned buildings. Of those labelled as potentially “beautiful” by the machine, I chose 12 at random to research more closely for a potential visit.

The DecisionTreeClassifier model. Percentages represent the fraction of the permits which belong to that specific permit category. For example, the tree root splits where 2.6% of permits for a building were of the Fire/Security Upgrade type.

Results

To research the buildings I selected for a potential visit, I put their locations into Google Maps and reviewed both the satellite and street view imagery. Based on what I saw, I manually selected the one that seemed to be the most promising. One TTC trip later, I was there.

Hooray!

From the outside, this site checked all of my boxes: it had boards on the windows and derelict exteriors implying that none of the buildings (there were four of them!) had been used for quite some time.

I was excited to start exploring further, but unfortunately it seemed that I had arrived at a bad time. Construction workers were already busy working on the interior of the buildings when I got there.

Hard at work…

I was still able to poke my head around and find some interesting things, regardless. I spotted some graffiti on the roof and saw through some windows into the rooms. I also found a smashed door and a suitcase that had been thrown into a basement window grate. It made me wonder who might have lived here before, and whether other explorers had already seen the place before I had.

Oddities.

Goodbye beautiful!

Afterword

Have I described a perfect method for finding abandoned sites? Absolutely not. There are many limitations to my decision tree method and some are as follows:

Beauty is a subjective concept that depends on context. I am the eye behind the lens of my camera; a decision tree will never perfectly capture my subjectivity and sensibilities or the myriad of ways an abandoned place can capture my interests. My current sample size is small and not representative of all of the ways a building can be beautiful. Ultimately, I would like my decision tree to learn better rules, but I don’t think there are a fixed set of rules that define what I see as beautiful.
My decision tree, as it stands, has been “over-fitted”. This means that when I trained the tree on my dataset, it came up with criteria that perfectly categorises the buildings I photographed in the past into specific groups (beautiful or unsatisfying). The tree therefore conformed perfectly to the data I had in hand, but the data in hand may not represent buildings I have yet to see. My tree may therefore perform poorly when trying to categorize new, unseen buildings.
The dataset is imperfect and incomplete. Several abandoned places I have photographed did not show up in the Active Buildings dataset. Additionally, the details of the permits in the Active Buildings dataset generally reflect the intentions for the future of a site, while much of what I’m looking for in my photos reflects a building’s past. Information that I feel is important, like how well-preserved an abandoned building is, is missing from the dataset.

At the same time, this exercise saw me visit buildings in regions of the city that I might not have visited otherwise. Most of my exploring up until this point occurred in and around my local neighbourhood, and it was exciting to see new parts of the city which I’d never seen before. I still believe the best way to find abandoned sites is to randomly poke your nose around—urban exploration is all about exploration, after all. However, this project has taught me that using machine learning can provide a great launching point to start.

This project is also a tribute to Toronto. Since beginning it, I graduated from university and moved to a different country for work. I have been missing home quite terribly, because there’s really no other place quite like it. Through my photos, I hope that you have been able to see that.

Here is an abandoned washroom I stumbled upon organically while I was out for a walk in the woods in Toronto: