リモートセンシング画像に基づく都市昼間騒音の空間分布予測とスケール効果分析：成都のケーススタディ

Public health and urban sustainable development are seriously affected by urban road traffic noise. Conventional urban noise monitoring has high cost, narrow coverage, and slow updating; therefore, efficient monitoring is urgently required. In this paper, we merge high-resolution remote sensing imagery with end-to-end deep learning models based on ResNet and ViT architectures to explore the prediction performance of daytime urban noise with imagery of diverse spatial resolutions. The whole urban region of Chengdu was chosen for investigation. We generated an image dataset with nine distinct spatial resolutions via resampling, ranging from 0.25 m to 50 m, to assess resolution-dependent performance. Both ResNet and ViT models were trained and validated on this dataset. Surprisingly, results showed that prediction accuracy exhibited a nonlinear pattern of initially increasing and then decreasing with improving resolution. Imagery at 2-m resolution achieved the highest accuracy. Multi-scale fusion experiments did not lead to significant performance gains, suggesting that scale stacking may introduce redundant or conflicting spectral information without improving generalization. The findings suggest that the 2-m resolution balances detailed information and broader spatial context, making it the optimal choice for predicting urban noise distribution. A two-dimensional Fourier spectral analysis further revealed that the 2 m imagery achieves the most balanced distribution of high- and low-frequency components, providing a physical explanation for its superior performance. This approach provides a novel pathway for environmental noise monitoring and establishes a foundation for studies on scale effects and resolution selection.