[Paper Review] Spatial sensitivity analysis for urban land use prediction with physics-constrained conditional generative adversarial networks

Adrian Albert, Jasleen Kaur+2 more|arXiv (Cornell University)|Jul 22, 2019

Urban Design and Spatial Analysis12 references7 citations

TL;DR

This paper proposes a physics-constrained conditional GAN framework to predict urban land use changes using only global remote-sensing data, enabling large-scale spatial sensitivity analysis without relying on local socioeconomic data. The model generates realistic urban form distributions and computes spatial gradients via backpropagation, revealing that changes in population or economic proxies propagate over 15 km in urban regions.

ABSTRACT

Accurately forecasting urban development and its environmental and climate impacts critically depends on realistic models of the spatial structure of the built environment, and of its dependence on key factors such as population and economic development. Scenario simulation and sensitivity analysis, i.e., predicting how changes in underlying factors at a given location affect urbanization outcomes at other locations, is currently not achievable at a large scale with traditional urban growth models, which are either too simplistic, or depend on detailed locally-collected socioeconomic data that is not available in most places. Here we develop a framework to estimate, purely from globally-available remote-sensing data and without parametric assumptions, the spatial sensitivity of the ( extit{static}) rate of change of urban sprawl to key macroeconomic development indicators. We formulate this spatial regression problem as an image-to-image translation task using conditional generative adversarial networks (GANs), where the gradients necessary for comparative static analysis are provided by the backpropagation algorithm used to train the model. This framework allows to naturally incorporate physical constraints, e.g., the inability to build over water bodies. To validate the spatial structure of model-generated built environment distributions, we use spatial statistics commonly used in urban form analysis. We apply our method to a novel dataset comprising of layers on the built environment, nightlighs measurements (a proxy for economic development and energy use), and population density for the world's most populous 15,000 cities.

Motivation & Objective

• Address the lack of scalable, data-driven urban modeling tools that can forecast urban sprawl without relying on scarce local socioeconomic data.
• Develop a machine learning framework capable of performing comparative statics analysis—assessing how changes in macroeconomic factors affect urban development spatially.
• Incorporate physical constraints (e.g., no construction over water) naturally into the generative modeling process.
• Validate model-generated urban forms using domain-specific spatial statistics such as fractal dimension and built-up density.
• Enable global-scale urban simulation for scenario planning in data-scarce regions, particularly in developing countries.

Proposed method

• Formulate urban land use prediction as an image-to-image translation task using conditional GANs, with input maps of population density and nightlight luminosity as conditions.
• Integrate physical constraints (e.g., water bodies) into the GAN loss function to prevent unrealistic urban development in prohibited zones.
• Use backpropagation during training to implicitly compute spatial gradients of built-up area with respect to input factors (population and luminosity), enabling sensitivity analysis.
• Train the model on a novel global dataset comprising built-up areas (DLR), population density (LandScan), and nightlight data (NASA), covering ~15,000 cities.
• Validate generated urban patterns using spatial statistics common in urban science, including fractal dimension and built-up area density.
• Perform spatial sensitivity analysis by computing gradient magnitudes and their propagation distances across global metropolitan regions.

Experimental results

Research questions

• RQ1 Can a deep generative model trained solely on remote-sensing data produce realistic urban land use patterns at a global scale without parametric assumptions or local data?
• RQ2 How can physical constraints such as water bodies be naturally embedded into a deep generative model for urban form prediction?
• RQ3 To what extent do changes in population density or economic proxies (nightlights) influence urban development in spatially distant regions?
• RQ4 What is the spatial range over which sensitivity to macroeconomic factors propagates in urban systems?
• RQ5 Can gradient-based sensitivity analysis be effectively extracted from a GAN-based urban modeling framework without ground-truth labels?

Key findings

• The model successfully generates realistic urban land use patterns that match real-world spatial statistics such as fractal dimension and built-up density.
• Spatial sensitivity analysis reveals that a unit change in population density or luminosity (proxy for economic activity) propagates its influence over distances exceeding 15 km in urban regions.
• Gradient spillover—where sensitivity to input changes extends beyond the immediate region—is quantitatively measurable and non-negligible, with average gradient magnitude dropping below 1% of peak value beyond 15 km.
• The framework enables large-scale, non-parametric sensitivity analysis across ~3,000 major metropolitan regions globally, revealing regional differences in gradient propagation patterns.
• The use of backpropagation to extract spatial gradients provides a scalable, differentiable method for comparative statics in urban modeling without requiring explicit parametric modeling of spatial dependencies.
• Despite the opacity of GANs, the model’s output is validated through domain-relevant spatial metrics, establishing a proof of concept for data-driven urban simulation.