Vision-Based Air Quality Index Estimation: A Systematic Failure Analysis

This record contains an independent research preprint analyzing the robustness of vision-based Air Quality Index (AQI) estimation under distribution shift. The study evaluates four modeling paradigms—vision-only models, physics-inspired features, statistical fusion, and gated fusion—across satellite, smartphone, CCTV, and urban camera data. While several models achieve strong in-distribution performance (R² up to 0.93), performance degrades severely under strict zero-shot out-of-distribution evaluation, with large error amplification, negative R², and fold-level training instability. Fusion mechanisms frequently degrade performance rather than improving robustness. The manuscript is accompanied by complete code, configurations, and per-sample prediction outputs for independent verification. All reproducibility artifacts are publicly available at:https://github.com/Nesar21/airvision