Yamaguchi Lab, The University of Osaka
Light-pollution prediction for Urban Air Mobility
Flying taxis (Urban Air Mobility, UAM) will reflect sunlight into streets and windows. I built a method that predicts this glare across a whole city in seconds instead of hours, so routes and flight times can be compared before anything flies.
3 h → 1.62 s
Evaluation time
10,000 observation points over a five-second window, after one-time city initialization (Core i5-13600K + RTX 3090).
10–54%
Lower RMSE
Versus raw neural prediction, across 18 settings in Rome, Taipei, and New York.
+32–91%
Precision in dense high-rise areas
Relative improvement in New York versus raw neural prediction. Occlusion correction trades off some recall.
Problem
UAM aircraft keep changing position and attitude, so the direction of reflected sunlight — and which buildings block it — changes every moment.
Accurate ray tracing at city scale takes hours. A neural network alone is fast, but it overestimates glare when a building hides the aircraft.
Approach
I split the problem in two. A geometry-aware neural surrogate (a PyTorch 3D-CNN + MLP) learns how an aircraft reflects sunlight in open space; it is trained once and shared across cities.
A geometric occlusion correction handles each city's buildings. Visibility is precomputed into a cache, so prediction only needs fast lookups. Reference data comes from Unreal Engine 5 (Nanite / Lumen) simulations.
How it works
City and flight data
3D buildings, UAM trajectories and attitude, sun direction, observer positions
Neural surrogate
Learns reflected sunlight in open space. Trained once, shared across cities.
Occlusion correction
Per-city visibility cache built from building geometry. Fast lookups at prediction time.
Glare map
Reflected-sunlight exposure per location, for comparing routes and times
My role
I led the method design, simulation environment, data processing, model implementation, experiments, evaluation, and paper writing, building on research directions discussed with my advisor.
Scope
A simulation study under clear-sky conditions. The output is a relative exposure metric and has not yet been calibrated against human discomfort or safety thresholds.
Publications
- PublishedIEEE SmartComp 2026 · Messina, Italy · pp. 136–143
A Simulation-based Framework for Dynamic Light Pollution Prediction in Urban Air Mobility
Ying Chieh Wang (first author), Tatsuya Amano, Hirozumi Yamaguchi
- Published · AwardIPSJ SIG Technical Report, Vol. 2026-ITS-105, No. 38, pp. 1–8 · May 2026
Surrogate-Based Dynamic Risk Mapping of Light Pollution for Urban Air Mobility
Ying Chieh Wang (first author), Tatsuya Amano, Hirozumi Yamaguchi
- Under reviewJournal of Information Processing (JIP) · journal extension
A Geometry-Aware Neural Surrogate for City-Scale Simulation of Light Pollution from Urban Air Mobility
Ying Chieh Wang (first author), Tatsuya Amano, Hirozumi Yamaguchi
- AcceptedDPSWS 2026 (IPSJ DPS Workshop) · demo / poster paper
A Real-Time Urban Air Mobility Simulator Using a Ray-Tracing Surrogate Model
Ying Chieh Wang (first author), Tatsuya Amano, Hirozumi Yamaguchi
Award
Encouragement Presentation Award
IPSJ 105th SIG-ITS Workshop (Intelligent Transport Systems and Smart Community), 2026
Surrogate-Based Dynamic Risk Mapping of Light Pollution for Urban Air Mobility