Sensing population distribution from satellite imagery via deep learning: model selection, neighboring effect, and systematic biases

Xiao Huang, Di Zhu, Fan Zhang, Tao Liu, Xiao Li, Lei Zou

Abstract

The rapid development of remote sensing techniques provides rich, large-coverage, and high-temporal information of the ground, which can be coupled with the emerging deep learning approaches that enable latent features and hidden geographical patterns to be extracted. This study marks the first attempt to cross-compare performances of popular state-of-the-art deep learning models in estimating population distribution from remote sensing images, investigate the contribution of neighboring effect, and explore the potential systematic population estimation biases. We conduct an end-to-end training of four popular deep learning architectures, i.e., VGG, ResNet, Xception, and DenseNet, by establishing a mapping between Sentinel-2 image patches and their corresponding population count from the LandScan population grid. The results reveal that DenseNet outperforms the other three models, while VGG has the worst performances in all evaluating metrics under all selected neighboring scenarios. As for the neighboring effect, contradicting existing studies, our results suggest that the increase of neighboring sizes leads to reduced population estimation performance, which is found universal for all four selected models in all evaluating metrics. In addition, there exists a notable, universal bias that all selected deep learning models tend to overestimate sparsely populated image patches and underestimate densely populated image patches, regardless of neighboring sizes.

Introduction

Accurate and timely population distribution data is essential for urban planning, disaster response, and public health initiatives. Traditional census data, while reliable, is often updated infrequently and is constrained by administrative boundaries, leading to the Modifiable Areal Unit Problem (MAUP). Remote sensing (RS) coupled with deep learning (DL) offers a promising alternative for generating high-resolution, near-real-time population grids. However, there is currently no consensus on which DL architectures are most effective for this task, how spatial context (neighboring effects) influences accuracy, or what inherent biases these models might possess.

Methods and Datasets

This study utilizes an end-to-end deep learning framework to map satellite image patches to population counts.

Data Sources

The primary inputs are Sentinel-2 Level-2A multispectral images (10m resolution) including Red, Green, Blue, and Near-Infrared (NIR) bands. Ground truth population data is derived from the LandScan USA 2019 grid, aggregated to a 30 arc-second (~1km) resolution to align with the satellite imagery.

Figure 1: Study areas in Metro Dallas (TX) and Metro Atlanta (GA) used for model training and cross-validation.

Model Architectures and Training

Four state-of-the-art architectures were evaluated: VGG-16, ResNet-50, Xception, and DenseNet-121. To accommodate the 4-channel input (RGB+NIR) into models pre-trained on 3-channel ImageNet data, a $1 \times 1$ convolutional layer was used for feature fusion. The target variable was the logarithm of the population ($p_i^{lg}$) to mitigate the impact of the heavily-tailed distribution. Models were trained using the log-cosh loss function, chosen for its robustness against outliers.

Results and Discussion

Model Performance and the Neighboring Paradox

DenseNet-121 emerged as the superior model, achieving an $R^2$ of 0.915 in the Atlanta test set. Its performance is attributed to "dense blocks" that facilitate feature reuse and capture complex urban patterns more effectively than the simpler stacking in VGG or residual connections in ResNet.

A significant finding of this study is the "neighboring paradox." Contrary to previous assumptions that more context improves prediction, increasing the neighborhood size from a $1 \times 1$ patch to an $11 \times 11$ area consistently degraded performance across all models. For DenseNet, $R^2$ dropped from 0.915 to 0.811 as the neighborhood expanded. This suggests that the high heterogeneity of urban environments and the dilution of central patch features in fixed-size inputs negatively impact model precision.

Systematic Biases

A universal bias was observed across all models: the overestimation of sparsely populated areas and the underestimation of densely populated urban cores.

Figure 2: Distribution of log-transformed population counts ($p_i^{lg}$) and representative Sentinel-2 image patches for different density levels.

As shown in Figure 2, the distribution of population is highly skewed. The underestimation in dense areas is likely due to the "vertical blind spot" of 2D satellite imagery, which cannot distinguish between building heights. A mid-rise and a skyscraper may appear similar from a top-down view but house vastly different numbers of people.

Conclusion

This research provides a benchmark for deep learning-based population estimation. While DenseNet shows high accuracy, the negative impact of neighboring context and the systematic bias in high-density areas highlight areas for future work. Incorporating 3D structural data (e.g., LiDAR) and exploring multi-scale analysis will be crucial for developing more robust global population models.

Summary updated

Abstract

The rapid development of remote sensing techniques provides rich, large-coverage, and high-temporal information of the ground, which can be coupled with the emerging deep learning approaches that enable latent features and hidden geographical patterns to be extracted. This study marks the first attempt to cross-compare performances of popular state-of-the-art deep learning models in estimating population distribution from remote sensing images, investigate the contribution of neighboring effect, and explore the potential systematic population estimation biases. We conduct an end-to-end training of four popular deep learning architectures, i.e., VGG, ResNet, Xception, and DenseNet, by establishing a mapping between Sentinel-2 image patches and their corresponding population count from the LandScan population grid. The results reveal that DenseNet outperforms the other three models, while VGG has the worst performances in all evaluating metrics under all selected neighboring scenarios. As for the neighboring effect, contradicting existing studies, our results suggest that the increase of neighboring sizes leads to reduced population estimation performance, which is found universal for all four selected models in all evaluating metrics. In addition, there exists a notable, universal bias that all selected deep learning models tend to overestimate sparsely populated image patches and underestimate densely populated image patches, regardless of neighboring sizes. The methodological, experimental, and contextual knowledge this study provides is expected to benefit a wide range of future studies that estimate population distribution via remote sensing imagery.
BibTeX
@misc{huang2021sensing,
  author = {Xiao Huang and Di Zhu and Fan Zhang and Tao Liu and Xiao Li and Lei Zou},
  title = {{Sensing population distribution from satellite imagery via deep learning: model selection, neighboring effect, and systematic biases}},
  year = {2021},
  eprint = {2103.02155},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.02155},
}
Figure