Date of Award

Spring 2026

Document Type

Open Access Thesis

Department

Geography

First Advisor

Sicheng Wang

Abstract

Sound environments shape how people experience urban spaces, yet conventional noise monitoring relies on isolated sensors that lack semantic context and spatial scalability. This study explores whether Vision-Language Models (VLMs) can infer acoustic conditions directly from street-view imagery, offering a scalable alternative to dense sensor networks. This thesis focuses on three tasks: (1) identifying visible sound sources (vehicles, pedestrians, water features); (2) predicting sound pressure levels (SPL) in decibels; and (3) classifying sound type into Natural, Mechanical, or Human categories. I collected 200 field samples across 11 cities using mobile phones and decibel meters, producing aligned image-spectrogram-dB triplets spanning diverse environments (e.g., snow, night, desert, tropical). Qwen3-VL-4B was fine-tuned with Low-Rank Adaptation (LoRA), achieving 5.4 dB MAE on held-out validation and demonstrating cross-modal alignment between visual scenes and acoustic properties. To ensure interpretability, I employed minimalist structured prompts requiring explicit visual grounding (e.g., "fountain in center"), supplemented by Grad-CAM and attention rollout visualizations that confirm predictions align with physically plausible sound sources. External validation against the US DOT National Transportation Noise Map (4,000+ Manhattan locations) reveals 9.1 dB MAE with systematic bias (+6.7 dB) attributable to the model's detection of non-transportation sources (pedestrians, construction) invisible to traffic-based reference data. While output discretization limits precision for regulatory compliance, the model's directional accuracy and spatial coherence support exploratory soundscape mapping in data-poor areas. This work establishes proof-of-concept for  VLM-based acoustic inference and contributes an interpretable, sensor-free framework for studying urban sound-environment relationships.

Rights

© 2026, Pengyu Chen

Download

Included in

Geography Commons

Share

COinS