Road Classification Recognition Method and Interpretability Analysis Based on Deep Learning

Road classification recognition is a key link to improve the handling performance and driving stability of intelligent vehicles, and its fast and accurate recognition ability directly affects the decision-making and path planning of vehicles under complex working conditions. Aiming at the problems of high latency, low accuracy and lack of user trust in traditional methods, this paper proposes an interpretable road surface recognition algorithm based on lightweight deep learning, which takes into account the recognition performance and computational efficiency. First, the DeeplabV3 semantic segmentation model is used to accurately extract the pavement region, significantly suppressing the non-pavement interference, and achieving a pixel accuracy of 93.49% with an intersection and concurrency ratio of 84.55%. Subsequently, MobileNetV3 is utilized to efficiently classify four types of pavements, namely dry asphalt, wet asphalt, sand and snow, with an accuracy of 94.18%. To enhance the transparency of the model, SHAP values are further introduced and key features are visualized in the form of heat maps to explain the prediction basis of the model, which effectively improves the interpretability and user trust.