CAGAN for Light Pollution Control

Beijing

London

LosAngeles

0 100

0 25000

The left map shows NTL (Nighttime Light Intensity) for each building type.

The right map shows the city's light pollution.

A larger ATE (Average Treatment Effect) value indicates a greater impact of the building type on the light pollution indicator.

Residential NTL Change 0

Commercial NTL Change 0

Construction NTL Change 0

Grassland NTL Change 0

Ground Truth

Generation

CAGAN

Stage 1: A light pollution image classifier is trained with actual light pollution images, enabling it to establish accurate associations with given residential area light pollution maps.

Stage 2: A conditional variational autoencoder is trained exclusively using the base map without colorization. The generator is then guided by the classifier and discriminator to progressively generate authentic light pollution maps.

Causality Interpretability: Deliberately adjusting the average nighttime light intensity of various building types within residential areas leads to observable changes in brightness across different regions.

Publications

Causally Aware Generative Adversarial Networks for Light Pollution Control.

Yuyao Zhang, Ke Guo, Xiao Zhou. In Proceedings of the 38th Annual AAAI Conference on Artificial Intelligence (AAAI 2024).

Paper link: https://arxiv.org/abs/2401.06453.

Code link: https://github.com/YuyaoZhangQAQ/Light_Pollution_CAGAN.