Assessing the accuracy of census-independent small area modelled population datasets
This study validates the accuracy of census-independent modelled population estimates by comparing them with Zambia's 2022 census data, finding strong correlations at national and sub-national levels while identifying that refining residential settlement stratification is crucial to reducing systematic overestimation.
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
Counting people is one of the most fundamental tasks a society undertakes, yet it is also one of the most difficult. Governments and aid organizations rely on accurate population numbers to plan schools, build hospitals, and distribute food during crises. Traditionally, this counting happens through a census, a massive, expensive effort where officials visit every home, usually once every ten years. But in many parts of the world, these censuses are delayed, disrupted by conflict, or simply never happen. When the official count is missing or outdated, the numbers used for planning become guesses, and the further away we get from the last real count, the less reliable those guesses become. To fill these gaps, researchers have developed a different approach: using satellite images to count buildings and then estimating how many people live inside them. This method, often called a "bottom-up" model, builds a picture of the population from the ground up, using the physical evidence of where people live rather than waiting for a government to send out a team of counters.
In Zambia, this new approach met a rare opportunity for a real-world test. In 2020, researchers produced a detailed map of the country's population using satellite images of buildings and some limited survey data, creating an estimate of where people lived without relying on a national census. Two years later, in 2022, Zambia finally conducted its official national census, the first to record the exact location of every household using GPS devices. This timing allowed a team of scientists to compare the satellite-based estimates directly against the official government count. They wanted to see if the computer models, which had been built without any census data, could accurately reflect the reality of the population. The results offered a clear picture of both the power and the current limits of this technology.
The researchers found that the satellite-based estimates were remarkably good at capturing the general shape of the population. When they looked at the country's large provinces and districts, the numbers from the model and the census moved in lockstep. If a district had a lot of people according to the census, the model predicted a lot of people too. The correlation was so strong that the two sets of numbers were almost indistinguishable in their patterns. However, a closer look revealed a consistent problem: the model tended to guess that there were more people than the census actually found. On average, the model overestimated the population in most areas. This was not a case of the model missing entire towns; rather, it was counting people in places where they might not actually live.
The key to understanding this overestimation lay in what the model was looking at. The computer program counted every building footprint it could see in the satellite images and assumed that each one was a home. But buildings are not all the same; some are houses, while others are factories, warehouses, or offices. Because the satellite data did not include labels to tell the difference between a home and a non-residential building, the model counted every structure as if it were occupied by a family. When the researchers filtered the data to include only the grid squares where the census had confirmed the presence of actual households, the overestimation dropped significantly. By removing the areas that looked like buildings but were likely not homes, the model's numbers came much closer to the official count.
Even with this adjustment, the comparison was not perfect. The researchers had to account for another issue: the timing of the data. The satellite images used for the model were taken between 2017 and 2019, while the census happened in 2022. In the years between, some buildings may have been demolished, or new ones constructed, and people may have moved. The model also missed some people entirely because their homes did not have a visible building footprint in the satellite images, perhaps because the structures were too small or the images were not clear enough. When the team adjusted the census numbers to exclude these missing households, the gap between the model and the census narrowed even further. The final comparison showed that the model was highly accurate for the areas it could see, but it struggled to distinguish between a house and a warehouse, leading it to guess too high in areas with many non-residential buildings.
The study highlighted a specific challenge in the capital city, Lusaka. Here, the model predicted a population that was substantially higher than the census count, even after the researchers tried to filter out non-residential buildings. The city is dense and complex, and the difference suggests that the model may have counted many commercial buildings as homes, or that the population distribution in the city changed significantly between the time the satellite images were taken and the census was conducted. Despite this, the overall performance of the model was impressive. It successfully identified where people lived across a vast and diverse country, proving that satellite-based methods can serve as a reliable alternative when a traditional census is not possible.
This work does not suggest that satellite models should replace censuses entirely. A census collects far more than just numbers; it gathers information on age, education, and health that a satellite camera cannot see. However, for countries that cannot afford or organize a full census, these models offer a vital lifeline. They provide a way to know, with a high degree of confidence, how many people live in a specific village or district, allowing for better planning and resource allocation. The study confirms that while the technology is not yet perfect, it is a powerful tool that is getting better. As satellite imagery improves and researchers learn to distinguish more clearly between different types of buildings, these digital estimates will become even more precise, offering a clear view of the human landscape even in the most remote or unstable corners of the world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.