SkyQI: A Citizen Science Platform for Global Light Pollution Monitoring Using Smartphone Photography and Computer Vision

Abstract

Version 3.0 (2026-10-02) corrects two errors that remained in version 2.0: the overall 95% confidence interval upper bound is 73.8%, not 74.2%, and the 104-image evaluation set comprises 59 Wikimedia Commons night-sky photographs and 45 synthetic test images generated with controlled brightness and Bortle parameters, which earlier versions did not state. Version 2.0 (2026-05-11) corrected the satellite comparison. Version 1.0 reported a correlation of r = 0.577 between SkyQI and VIIRS radiance over 60,681 grid points. That reference was later found to be a synthetic dataset generated by a project script and mislabelled as VIIRS, with per-photo radiance derived from the photos' own Bortle labels, so the figure did not measure agreement with satellite data. Against the real EOG VNL V2.2 2025 composite (84,100 cells at 0.1°), r = −0.175 (n = 103, p = 0.074). The ±1 Bortle accuracy of 65.4% is unaffected. Light pollution affects 83% of the global population, yet comprehensive monitoring remains limited by the cost of specialized equipment. This paper presents SkyQI, a citizen science platform that enables light pollution monitoring using smartphone photography and computer vision. The platform is designed for global use, but the present evaluation is geographically limited to the Indian subcontinent. The system employs connected-component based star detection, regional brightness gradient analysis, and color temperature assessment to estimate approximate SQM-equivalent values and Bortle Dark-Sky Scale classifications from uncalibrated photographs. Evaluation on 104 geolocated night-sky photographs demonstrates 65.4% accuracy (±1 Bortle class, 95% CI: 55.8%–73.8%), with strongest performance for dark sky verification (Bortle 1–2: 75.8%) and urban documentation (Bortle 7–9: 100.0%, n=16), though mid-range classification (Bortle 3–6) remains challenging at 9.1%. Performance is contextualized against external reference data from the EOG VNL V2.2 2025 VIIRS satellite composite (84,100 grid cells at 0.1°), 734 Unihedron SQM readings, and 238 Globe at Night observations. Separately, VIIRS-derived Bortle estimates showed only 51.7% agreement with Globe at Night human observers (mean error −2.53 Bortle classes), highlighting the limitations of satellite-only mapping for ground-level sky quality assessment.