💻 computer science

LOCATE: Context-Aware GeoAI Decision Support for Population-Oriented Air-Quality Sensor Placement in Data-Constrained Cities

The paper introduces LOCATE, a data-light GeoAI framework that optimizes pre-deployment air-quality sensor placement in data-constrained cities by prioritizing population proximity and urban context, demonstrating that a context-augmented population-weighted p-median heuristic outperforms other methods across multiple African cities.

Original authors: Fidel Raja Wabinyai, Deogratius Okedi, Priscah Adrine, George William Settala, Usman Abdul-Ganiy, Jennifer Kutesakwe, Alex Ndyabakira, Sharifah Buzeki, Sarah Zalwango, Engineer Bainomugisha, Deo Okure

Published 2026-11-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fidel Raja Wabinyai, Deogratius Okedi, Priscah Adrine, George William Settala, Usman Abdul-Ganiy, Jennifer Kutesakwe, Alex Ndyabakira, Sharifah Buzeki, Sarah Zalwango, Engineer Bainomugisha, Deo Okure

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Breathing clean air is a fundamental human need, yet in many of the world's fastest-growing cities, the air is often thick with invisible pollutants from traffic, industry, and burning waste. To understand and fix this problem, scientists and city planners rely on networks of sensors that measure air quality. However, placing these sensors is a difficult puzzle. If they are placed too far apart, they miss dangerous pockets of pollution; if they are clustered too closely, they waste resources and fail to tell the whole story. In cities with limited data and funding, finding the perfect spots is even harder because planners often lack detailed maps of where pollution actually is. They must decide where to put the sensors before the sensors are even installed, relying on what they know about where people live and how the city is built.

A team of researchers from Makerere University and partners in Uganda, Kenya, and Nigeria has developed a new way to solve this puzzle. They created a digital planning tool called LOCATE, which acts like a smart guide for city officials. Instead of needing a perfect map of pollution—which often doesn't exist yet—the tool uses open maps of roads and buildings, combined with data on where people live, to figure out the best places to put air-quality sensors. The researchers tested this tool in four very different African cities: Kampala and Jinja in Uganda, Nairobi in Kenya, and Lagos in Nigeria. They wanted to see if they could design a sensor network that would get close to the most people, while also capturing the unique character of each city's neighborhoods, without needing expensive or complex pollution data to start.

The researchers approached the problem by treating the cities as a collection of potential spots. They generated thousands of possible locations for sensors within the city boundaries. For each spot, the tool looked at two main things: how many people lived nearby and what the immediate surroundings looked like, such as how many roads or buildings were within a short walking distance. They then ran a series of computer simulations to see which method of picking the best spots worked most effectively. They compared ten different strategies, ranging from simple random selection to complex mathematical algorithms designed to group similar areas together or to maximize the distance between sensors. They tested these strategies with different numbers of sensors, from as few as five up to eighty, to see how the coverage changed as the budget grew.

The results showed that one specific strategy stood out as the most reliable for the goal of protecting public health. This strategy, known as a population-weighted p-median heuristic, worked by prioritizing locations that were close to large groups of people while also ensuring the sensors were placed in areas that represented the city's diverse environments, like busy commercial streets or dense residential zones. In every city tested, this method produced the best overall layout. It managed to place sensors so that nearly all the modeled population in the smaller cities of Kampala and Jinja would be within two kilometers of a sensor when eighty sensors were used. In the much larger and more complex cities of Nairobi and Lagos, the same number of sensors covered slightly less of the population, reaching about 98 percent and 90 percent respectively. This difference highlights that a single number of sensors does not work for every city; a network that is perfect for a compact city might leave huge gaps in a sprawling metropolis.

The study also found that while some computer methods were very good at grouping similar data points together, that skill did not always translate to the best real-world sensor placement. For instance, a method called KMeans was excellent at creating neat mathematical groups, but it did not perform as well as the population-focused strategy when it came to actually covering the people who needed monitoring. This distinction is crucial: a mathematically perfect group of data points does not necessarily mean a sensor network that protects the most people. The researchers also noted that their tool is designed for the planning stage, before any hardware is bought or installed. It produces a shortlist of recommended sites, but it cannot guarantee that a specific spot is safe, has electricity, or is allowed by local authorities. Those final checks must still be done by people on the ground.

Ultimately, the work demonstrates that cities can plan effective air-quality monitoring networks even without expensive pollution maps. By using freely available data about population and city infrastructure, planners can make informed decisions about where to place sensors to get the most value for their money. The tool does not claim to solve the pollution problem itself, nor does it measure the air directly. Instead, it provides a clear, evidence-based starting point for governments to begin building networks that can eventually protect the health of millions of residents. The study confirms that there is no universal rule for how many sensors a city needs; the right number depends entirely on the size of the city, how people are spread out, and the specific goals of the monitoring program. With this approach, cities can move forward with confidence, knowing their sensor networks are designed to reach the people who need them most.

Technical Summary: LOCATE – Context-Aware GeoAI Decision Support for Population-Oriented Air-Quality Sensor Placement

Problem Statement
Air-quality monitoring is critical for environmental health surveillance, yet many low- and middle-income cities lack the reliable, citywide pollution surfaces, detailed emissions inventories, or validated dispersion models required to guide sensor placement. In these "data-constrained" environments, pre-deployment planning cannot rely on historical pollutant measurements. Furthermore, existing network design approaches often fail to account for the specific urban contexts (e.g., road density, building footprints) and population distributions of diverse African cities. There is a need for a reproducible, data-light framework that can generate defensible sensor layouts prioritizing proximity to residential populations and urban context without requiring prior pollution data.

Methodology
The study introduces LOCATE, a GeoAI decision-support framework designed for the pre-deployment stage. The methodology operates in five stages: input definition, spatial/contextual feature preparation, model fitting, feasibility repair, and ranking.

Data Inputs: The framework utilizes WorldPop gridded population data (as demand weights) and OpenStreetMap (OSM) data (road networks, intersections, building footprints) to characterize urban context. It does not use operational pollutant measurements.
Study Scope: The framework was evaluated across four cities with distinct urban forms: Kampala and Jinja (Uganda), Nairobi (Kenya), and Lagos City (Nigeria).
Experimental Design: Nine sensor budgets ( $N = 5$ to $80$) were tested. A minimum pairwise separation of 0.5 km was enforced.
Placement Algorithms: Ten distinct approaches were compared:
- Direct Allocation: Context-augmented population-weighted $p$ -median heuristic and feature-space maximin.
- Fixed-Count Clustering: KMeans, Approximate K-medoids, Gaussian Mixture Models (GMM), and Ward-linkage agglomerative clustering.
- Density-Based Clustering: DBSCAN and HDBSCAN.
- Uninformed Baselines: Random selection and a hexagonal-grid approximation.
Evaluation Metrics:
- Internal Diagnostics: Silhouette coefficient, Davies–Bouldin index, and Calinski–Harabasz score (calculated on pre-repair labels for clustering methods).
- Final-Layout Performance: Population-weighted nearest-site distance ( $D_{pop}$ ), 95th-percentile geographic distance ( $D_{95}$ ), and population coverage within 2 km ( $P_{pop}(2)$ ).
- Composite Ranking: Methods were ranked using a weighted composite score emphasizing population proximity (0.35), 2-km coverage (0.20), and contextual representation (road/building density).

Key Results

Top-Performing Method: The context-augmented population-weighted $p$ -median heuristic ranked first in all four cities across all nine sensor budgets under the declared population-and-context composite objective. It effectively reduced population-weighted proximity gaps while prioritizing candidates in high-context areas.
Clustering Performance: Among fixed-count clustering methods, KMeans produced the strongest mean internal diagnostics (highest Silhouette and Calinski–Harabasz scores; lowest Davies–Bouldin scores) across the seven-budget diagnostic subset ( $N=5$ to $60$).
Coverage by City: Coverage improved with budget but varied significantly by city scale and form. At $N=80$ , the $p$ -median layouts placed approximately 100% of the modelled population within 2 km in Kampala and Jinja, 98% in Nairobi, and 90% in Lagos City. The lower coverage in Lagos and Nairobi was partly attributed to large water bodies and protected non-residential lands (e.g., Nairobi National Park) within the administrative boundaries.
Baseline Performance: The hexagonal-grid baseline consistently ranked second across all cities, suggesting that simple spatial balance provides a competitive planning benchmark.
Internal vs. External Validity: The study highlights a critical distinction: strong internal clustering diagnostics (e.g., KMeans) do not necessarily translate to superior final-layout performance for population-oriented objectives. The $p$ -median's success was attributed to its direct alignment with the population-proximity and contextual endpoints.

Significance and Claims
The paper claims that LOCATE provides a transparent, reproducible starting point for sensor network planning in cities lacking prior pollution data. Its primary significance lies in:

Data-Light Planning: Enabling population-oriented and context-aware planning using only open geospatial data (WorldPop and OSM) before any sensors are installed.
Methodological Separation: Demonstrating that internal clustering quality and deployment coverage are distinct metrics that must be evaluated separately; optimizing for one does not guarantee the other.
Contextual Awareness: Incorporating urban context (roads, buildings) to ensure selected sites represent diverse urban settings, not just population density.
Decision Support, Not Determination: The authors explicitly state that LOCATE outputs are planning shortlists for institutional and field review, not evidence of pollution reconstruction, exposure accuracy, or field feasibility. The framework narrows the field-review problem but does not replace the need for regulatory review, site verification, host engagement, and post-installation validation.

The study concludes that no universal sensor count is appropriate for all cities; budget requirements must be evaluated based on city-specific evidence regarding scale, boundary geometry, and population distribution.

More like this