The demand for diverse traffic condition data has never been greater, driven by the widespread adoption of intelligent transportation systems (ITS), including autonomous vehicles and advanced driver-assistance systems (ADAS). A key element in ITS is the classification of road surface types (RST), which is essential for various ITS applications. Effective RST classification models must deliver robust and consistent performance across different vehicles, driving styles, and terrains. Recent research has shown the benefits of deep learning techniques for classifying sensor-based RST data, as deep learning methods enable automated feature extraction, removing the need for manual processes. Efficient feature extraction is vital for improving RST classification accuracy. This study examines advancements in deep learning techniques for sensor-based RST categorization and assesses suitable models for feature extraction. To understand their functional structures, we first explored various convolutional neural networks (CNNs) and recurrent neural networks (RNNs). We then developed a sophisticated deep learning model, the residual bidirectional gated recurrent unit with squeeze-and-excitation mechanism (ResBiGRU-SE). This model integrates residual connections and squeeze-and-excitation modules to enhance classification accuracy. Our primary objective was to classify road surfaces into dirt, cobblestone, or asphalt categories. Comparative experiments using the publicly available passive vehicular sensors (PVS) dataset demonstrated that the ResBiGRU-SE model outperforms other state-of-the-art models. The ResBiGRU-SE model achieved an accuracy of 98.41\(\%\) and an F1-score of 98.21\(\%\) in accurately categorizing road surfaces, showcasing its superior performance.