Anonymized Urban Soundscape Prediction Dataset and Reproducible Code for Street-View and Audio Analysis

Description

This dataset contains anonymized processed data and reproducible code supporting a study on urban soundscape prediction using street-view visual features and audio-derived soundscape labels. The release includes the final modeling dataset, ADE20K visual feature statistics, Qwen3-Omni soundscape classification outputs, a human validation subset, data documentation, and reproducible Python scripts. Raw videos, raw audio files, original street-view images, exact GPS coordinates, local file paths, and original file names are not included for privacy and copyright reasons.

Steps to reproduce

1. Download and unzip the dataset package. 2. Install Python 3.10 or later. 3. Install the required packages: pip install -r requirements.txt 4. Run the validation script: python code/00_validate_package.py This script checks the released files, verifies row and column counts, and scans the public CSV files for sensitive markers such as local paths, original file names, raw sample identifiers, and raw model outputs. 5. Run the summary statistics script: python code/01_reproduce_summary_statistics.py This script reproduces label prevalence statistics, visual feature summaries, and human-validation metrics from the anonymized released data. 6. Run the model comparison script: python code/02_reproduce_model_comparison.py This script trains and evaluates baseline classifiers using the anonymized modeling dataset and reports Accuracy, Precision, Recall, F1-score, and ROC-AUC for the three soundscape prediction tasks. The released package contains only anonymized processed tables and reproducibility code. Raw videos, raw audio files, original street-view images, GPX traces, exact GPS coordinates, local file paths, original file names, and raw model response text are not included.

Institutions

• Tongji University

  Shanghai, Shanghai

Categories

Licence