PURPOSES : Road surface conditions are vital to traffic safety, management, and operation. To ensure traffic operation and safety during periods of snow and ice during the winter, each local government allocates considerable resources for monitoring that rely on field-oriented manual work. Therefore, a smart monitoring and management system for autonomous snow removal that can rapidly respond to unexpected abrupt heavy snow and black ice in winter must be developed. This study addresses a smart technology for automatically monitoring and detecting road surface conditions in an experimental environment using convolutional neural networks based on a CCTV camera and infrared (IR) sensor data. METHODS : The proposed approach comprises three steps: obtaining CCTV videos and IR sensor data, processing the dataset acquired to apply deep learning based on convolutional neural networks, and training the learning model and validating it. The first step involves a large dataset comprising 12,626 images extracted from the acquired CCTV videos and the synchronized surface temperature data from the IR sensor. In the second step, image frames are extracted from the videos, and only foreground target images are extracted during preprocessing. Hence, only the area (each image measuring 500 × 500) of the asphalt road surface corresponding to the road surface is applied to construct an ideal dataset. In addition, the IR thermometer sensor data stored in the logger are used to calculate the road surface temperatures corresponding to the image acquisition time. The images are classified into three categories, i.e., normal, snow, and black-ice, to construct a training dataset. Under normal conditions, the images include dry and wet road conditions. In the final step, the learning process is conducted using the acquired dataset for deep learning and verification. The dataset contains 10,100 (80%) data points for deep learning and 2,526 (20%) points for verification. RESULTS : To evaluate the proposed approach, the loss, accuracy, and confusion matrix of the addressed model are calculated. The model loss refers to the loss caused by the estimated error of the model, where 0.0479 and 0.0401 are indicated in the learning and verification stages, respectively. Meanwhile, the accuracies are 97.82% and 98.00%, respectively. Based on various tests that involve adjusting the learning parameters, an optimized model is derived by generalizing the characteristics of the input image, and errors such as overfitting are resolved. This experiment shows that this approach can be used for snow and black-ice detections on roads. CONCLUSIONS : The approach introduced herein is feasible in road environments, such as actual tunnel entrances. It does not necessitate expensive imported equipment, as general CCTV cameras can be applied to general roads, and low-cost IR temperature sensors can be used to provide efficiency and high accuracy in road sections such as national roads and highways. It is envisaged that the developed system will be applied to in situ conditions on roads.