Automated Road Surface Classification using Real-time Multispectral Sensing

Municipal road networks span over 4 million miles in the United States, yet current assessment methods rely on manual visual inspection or specialized retroreflectivity equipment — approaches that are slow, expensive, and subjective. Spectral degradation in road surfaces precedes visible deterioration, meaning significant damage often goes undetected until it becomes a safety concern. This project develops a low-cost, automated road surface classification system using real-time multispectral sensing mounted on a vehicle or drone platform for continuous, passive road scanning. The system integrates an 8-channel Adafruit AS7341 spectral sensor array covering 405–855nm, an Adafruit I2C multiplexer for parallel triggering, active white LED illumination independent of ambient light conditions, and an ESP32-S3 microcontroller for wireless data broadcast via WiFi UDP. Spectral data is streamed at approximately 20 scans per second to a laptop running a real-time PyQtGraph dashboard, which displays predicted surface class and confidence scores. GPS tagging enables the generation of georeferenced road condition maps across entire networks. A Random Forest machine learning classifier was trained on three surface classes — asphalt, concrete, and brick — achieving 100% cross-validation accuracy (5-fold) across 3,000 labeled samples. Key spectral discriminators include the green band (AVG 550nm), red/orange band (AVG 600nm), and near-infrared band (NIR 855nm), which together capture material composition differences invisible to conventional cameras. Future work includes expanding the dataset to wet, dry, and worn surface conditions, integrating onboard SD card logging, and validating the system through field deployment.