ShelterMap Toronto

Inspiration

Toronto is currently experiencing one of the most severe homelessness crises in Canada. The tools available to municipal planners typically show what is happening, for example, shelter occupancy or general homelessness counts but they do not answer where new shelters should be placed or which areas are most underserved.

Furthermore, encampment data, shelter availability, and geospatial patterns are often dispersed across multiple incomplete datasets. We wanted to build something that unifies these datasets and provides actionable, data-driven, and map-based insights that can meaningfully support:

Urban planners
City officials
Community organizations
Emergency responders
Social service agencies

ShelterMap Toronto was created as a project dedicated to using data science for social good, aligned with

UN SDG 11 : Sustainable Cities & Communities
UN SDG 1 : No Poverty

Our guiding question:

How can we use real data, geospatial analysis, and clustering algorithms to identify where new shelters would have the greatest impact?

What It Does

ShelterMap Toronto is a comprehensive geospatial decision-support tool that integrates multiple data sources into a unified platform.

Key Features

Shelter Layer
Shows the locations of existing shelters across Toronto
Displays capacity, occupancy, utilization rate, and type
Allows filtering and comparative analysis of service coverage
Encampment Layer
Includes a custom dataset of encampment hotspots across Toronto
Uses geocoded coordinates to map locations such as parks, ravines, underpasses, and high-density intersections
Severity Modeling

How We Built It

Data

We integrated several datasets, including:

City of Toronto Shelter System Occupancy Dataset (2024)
Required extensive cleaning due to missing bed-level data
Custom Encampment Dataset (2025)
Created using journalistic and public-source encampment reporting
Fully geocoded and cleaned
Street Needs Assessment (SNA) 2024
Used for validating outdoor homelessness counts by district
Geocoding via leaflet
Used to convert encampment names into coordinates
Includes caching to avoid repeated API calls

Backend (Python)

Implemented components include:

Data ingestion and cleaning scripts
Feature extraction
Weighted clustering
Geodesic distance calculations
Export of processed data to BigQuery

We used: -pandas for cleaning

numpy for vectorized calculations
geopy for distances
sklearn for clustering
google-cloud-bigquery for database operations

Frontend (Next.js + React)

Frontend includes:

Map rendered using React Leaflet
Layer toggles for shelters, encampments, and recommendations
API routes that fetch processed data from BigQuery
Summaries optionally generated using Gemini

Database

We used Google BigQuery as the central datastore for:

Raw datasets
Cleaned encampment and shelter data
Cluster results
Scoring outputs

This allowed us to integrate Python ML and Next.js frontend smoothly.

Challenges We Ran Into

Missing Shelter Data

The primary City of Toronto dataset was missing most bed-level occupancy fields. Solution: -Replaced it with complete room-level occupancy data.

Geocoding Failures on macOS

Nominatim HTTPS requests failed due to certificate issues. Solution:

Implemented SSL bypass logic and offline caching.

BigQuery Permission Issues

We initially lacked the required IAM roles to create tables. Solution:

Updated service account roles to include BigQuery Admin, Data Editor, and Data Viewer.

ML + Next.js Integration

Integrating Python-based clustering with Next.js in a short timeframe was challenging. Solution: -Wrote all ML outputs into BigQuery for easy consumption by serverless API routes.