Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

High-resolution Population Maps Derived from Sentinel-1 and Sentinel-2

Nando Metzger, Rodrigo Caye Daudt, Devis Tuia, Konrad Schindler

TL;DR

POPCORN introduces Bag-of-Popcorn, a data-efficient framework that derives high-resolution population maps from free Sentinel-1 and Sentinel-2 imagery using only coarse census counts for weak supervision. The model separably learns a built-up score and an occupancy rate, combining them to produce per-pixel populations that are dasymetrically calibrated to regional totals; ensembling across seeds and seasons enhances robustness. Across Switzerland, Rwanda (Kigali), and Puerto Rico, Bag-of-Popcorn outperforms several baselines, including those relying on high-resolution building footprints, and demonstrates strong data efficiency, transferability, and interpretability through explicit built-up and occupancy maps. This approach enables scalable, cost-effective population mapping in data-scarce regions, with potential for continuous monitoring and application to policy and humanitarian planning.

Abstract

Detailed population maps play an important role in diverse fields ranging from humanitarian action to urban planning. Generating such maps in a timely and scalable manner presents a challenge, especially in data-scarce regions. To address it we have developed POPCORN, a population mapping method whose only inputs are free, globally available satellite images from Sentinel-1 and Sentinel-2; and a small number of aggregate population counts over coarse census districts for calibration. Despite the minimal data requirements our approach surpasses the mapping accuracy of existing schemes, including several that rely on building footprints derived from high-resolution imagery. E.g., we were able to produce population maps for Rwanda with 100m GSD based on less than 400 regional census counts. In Kigali, those maps reach an R^2 score of 66% w.r.t. a ground truth reference map, with an average error of only about 10 inhabitants/ha. Conveniently, POPCORN retrieves explicit maps of built-up areas and of local building occupancy rates, making the mapping process interpretable and offering additional insights, for instance about the distribution of built-up, but unpopulated areas, e.g., industrial warehouses. Moreover, we find that, once trained, the model can be applied repeatedly to track population changes; and that it can be transferred to geographically similar regions, e.g., from Uganda to Rwanda). With our work we aim to democratize access to up-to-date and high-resolution population maps, recognizing that some regions faced with particularly strong population dynamics may lack the resources for costly micro-census campaigns.

High-resolution Population Maps Derived from Sentinel-1 and Sentinel-2

TL;DR

POPCORN introduces Bag-of-Popcorn, a data-efficient framework that derives high-resolution population maps from free Sentinel-1 and Sentinel-2 imagery using only coarse census counts for weak supervision. The model separably learns a built-up score and an occupancy rate, combining them to produce per-pixel populations that are dasymetrically calibrated to regional totals; ensembling across seeds and seasons enhances robustness. Across Switzerland, Rwanda (Kigali), and Puerto Rico, Bag-of-Popcorn outperforms several baselines, including those relying on high-resolution building footprints, and demonstrates strong data efficiency, transferability, and interpretability through explicit built-up and occupancy maps. This approach enables scalable, cost-effective population mapping in data-scarce regions, with potential for continuous monitoring and application to policy and humanitarian planning.

Abstract

Detailed population maps play an important role in diverse fields ranging from humanitarian action to urban planning. Generating such maps in a timely and scalable manner presents a challenge, especially in data-scarce regions. To address it we have developed POPCORN, a population mapping method whose only inputs are free, globally available satellite images from Sentinel-1 and Sentinel-2; and a small number of aggregate population counts over coarse census districts for calibration. Despite the minimal data requirements our approach surpasses the mapping accuracy of existing schemes, including several that rely on building footprints derived from high-resolution imagery. E.g., we were able to produce population maps for Rwanda with 100m GSD based on less than 400 regional census counts. In Kigali, those maps reach an R^2 score of 66% w.r.t. a ground truth reference map, with an average error of only about 10 inhabitants/ha. Conveniently, POPCORN retrieves explicit maps of built-up areas and of local building occupancy rates, making the mapping process interpretable and offering additional insights, for instance about the distribution of built-up, but unpopulated areas, e.g., industrial warehouses. Moreover, we find that, once trained, the model can be applied repeatedly to track population changes; and that it can be transferred to geographically similar regions, e.g., from Uganda to Rwanda). With our work we aim to democratize access to up-to-date and high-resolution population maps, recognizing that some regions faced with particularly strong population dynamics may lack the resources for costly micro-census campaigns.
Paper Structure (36 sections, 6 equations, 8 figures, 14 tables)

This paper contains 36 sections, 6 equations, 8 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Census counts and geographic boundaries
  4. Input Data
  5. Methodology
  6. Building Extractor
  7. Occupancy and Population Estimation
  8. Weakly Supervised Training
  9. Bagging
  10. Dasymetric Mapping
  11. Implementation Details
  12. Experimental Setup
  13. Evaluation Metrics
  14. Coefficient of Determination ($R^2$)
  15. Mean Absolute Error (MAE)
  16. ...and 21 more sections

Figures (8)

  • Figure 1: Schematic overview of our approach to population mapping from Sentinel-1 and Sentinel-2 imagery. A pre-trained dual-stream (DS) building detector estimates a per-pixel built-up score. Concurrently, a second, trainable dual-stream block estimates occupancy rates. The population map is derived as the per-pixel product of built-up score and occupancy. To supervise the training of the occupancy branch, the predicted population counts are aggregated within administrative regions and compared to the corresponding census data.
  • Figure 2: Dual-stream (DS) architecture proposed by hafner2022unsupervised.
  • Figure 3: Scatter plots for Switzerland, Rwanda, and Puerto Rico. Note the logarithmic scale of the axes. Values close to zero (below 0.5) have been grouped into a single bin.
  • Figure 4: Comparison of population maps for the Mahama refugee camp (Rwanda), Kingi (Rwanda), and Zurich, Switzerland. The first column shows very high-resolution images googlemaps2023, while the second column shows the Sentinel-2 RGB composites sentinel2_composite for reference. Bag-of-Popcorn maps were resampled to the same grid as WorldPop to ease visual comparison. All maps claim validity for 2020. Best viewed on screen.
  • Figure 5: Comparison of population density estimates in the Goma-Gisenyi border region: The vertical strip is the boundary, the left part lies in the DRC, and the right in Rwanda. Sources: VHR googlemaps2023, Sentinel-2 sentinel2_composite.
  • ...and 3 more figures