Start Date
4-9-2026 12:00 AM
Description
Over 58,000 residents of Utah's Uinta Basin depend on three highway corridors that each cross high-elevation mountain passes, yet the nearest Doppler radar is approximately 200 km away. BasinWX Roads (basinwx.com/roads) applies a rule-based computer vision pipeline to 46 existing Utah Department of Transportation (UDOT) traffic cameras, integrated with 9 Roadway Weather Information System (RWIS) stations, to provide automated road condition assessments without deploying new hardware. We characterize fundamental limitations of pixel-brightness approaches to snow detection using a labeled dataset of 138 images from four NYSDOT cameras scored by six independent coders. Full-image white pixel analysis achieved only 24.6% accuracy, with 91% of errors attributable to sky-driven false positives. Cropping to the road surface raised accuracy to 38.4%, but failed to detect any heavy snow events (0/27), because road snow appears gray (brightness 100-170) rather than white. Integrating RWIS temperature gating nearly doubled accuracy to 69.4% (n=49), with all 16 temperature-driven reclassifications correct. These results demonstrate that physical context from road weather stations, not pixel analysis alone, is what enables operational snow detection on traffic cameras. The system served an estimated 600-1,250 unique visitors per month through its first full winter (2025-26). Future work includes a machine learning classification pipeline, a labeled UDOT camera dataset, and improved nighttime and low-visibility detection.
Rule-Based Snow Detection Limits and RWIS Integration on UDOT Traffic Cameras
Over 58,000 residents of Utah's Uinta Basin depend on three highway corridors that each cross high-elevation mountain passes, yet the nearest Doppler radar is approximately 200 km away. BasinWX Roads (basinwx.com/roads) applies a rule-based computer vision pipeline to 46 existing Utah Department of Transportation (UDOT) traffic cameras, integrated with 9 Roadway Weather Information System (RWIS) stations, to provide automated road condition assessments without deploying new hardware. We characterize fundamental limitations of pixel-brightness approaches to snow detection using a labeled dataset of 138 images from four NYSDOT cameras scored by six independent coders. Full-image white pixel analysis achieved only 24.6% accuracy, with 91% of errors attributable to sky-driven false positives. Cropping to the road surface raised accuracy to 38.4%, but failed to detect any heavy snow events (0/27), because road snow appears gray (brightness 100-170) rather than white. Integrating RWIS temperature gating nearly doubled accuracy to 69.4% (n=49), with all 16 temperature-driven reclassifications correct. These results demonstrate that physical context from road weather stations, not pixel analysis alone, is what enables operational snow detection on traffic cameras. The system served an estimated 600-1,250 unique visitors per month through its first full winter (2025-26). Future work includes a machine learning classification pipeline, a labeled UDOT camera dataset, and improved nighttime and low-visibility detection.