Research map / Probabilistic and scientific modeling / Molecules and proteins
Causally Aware Generative Adversarial Networks for Light Pollution Control
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CAGAN, which uses debiased machine learning causal estimates of building-type effects on light pollution to condition a GAN-based (CVAE-GAN style) generator of residential light pollution maps across seven metropolises.
Problem
Light pollution in cities is typically quantified superficially, without identifying causal drivers at fine spatial and land-use scales, and existing tools offer little actionable guidance for illumination planning.
Approach
Combine double/debiased machine learning estimates of how POI building types affect over-illumination, trespass and clutter indices with a conditional variational autoencoder-GAN whose generator is guided by a pollution-level classifier, so generated maps reflect causal relationships and allow counterfactual NTL adjustments.
Key results
- CAGAN achieves KL divergence 7.2380 versus 28.2567 for uniform and 36.1757 for random distributions over pollution levels.
- Generated maps reach PSNR 24.095, SSIM 0.941, LPIPS 0.068 and MAE 0.012.
- DML analysis finds grasslands, commercial centers and residential buildings are consistently high-ATE contributors to light pollution across metropolises.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAGAN instead of a random distribution | Pollution-level distribution generation · Pollution-level distributions | KL: 36.1757 -> 7.2380 | Helped |
| Use CAGAN instead of a uniform distribution | Pollution-level distribution generation · Pollution-level distributions | KL: 28.2567 -> 7.2380 | Helped |
Surprising findings
- Grassland areas show among the largest positive effects on light pollution, attributed to nighttime lighting for visual enhancement and light scattering.
Limitations
- Causal estimates rely on observed confounders, so unmeasured confounding is not addressed.
- Evaluation of generated maps is largely pixel and distribution based, with no human or external validation of counterfactual outcomes.
- Study covers only seven wealthy metropolises, limiting generalization.
Experimental setup
Datasets: OpenStreetMap POI data night-time light remote sensing imagery
Benchmarks: residential light pollution maps from seven metropolises
Metrics: IS PSNR MAE SSIM LPIPS RASE KL divergence ATE
Baselines: Random distribution Uniform distribution
Builds on: CVAE-GAN double machine learning radial basis function light pollution model
Closest papers by meaning
- LuxDiT: Lighting Estimation with Video Diffusion Transformer (NeurIPS 2025 · similarity 0.81)
- LightIt: Illumination Modeling and Control for Diffusion Models (CVPR 2024 · similarity 0.81)
- Do Image Editing Models Understand Lighting? (NeurIPS 2026 · similarity 0.81)
- UrbanDataLayer: A Unified Data Pipeline for Urban Science (NeurIPS 2024 · similarity 0.81)
- DiffusionLight: Light Probes for Free by Painting a Chrome Ball (CVPR 2024 · similarity 0.81)
- ScribbleLight: Single Image Indoor Relighting with Scribbles (CVPR 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.