Classification of road transport earth retaining structures based on improved ConvNeXt V2

• A real-world dataset captured by a mobile system is developed to classify ERS. • The novel application of AI to classify ERS is introduced as a proof-of-concept. • The ConvNeXt V2 is enhanced for automatic ERS classification. • A web interface is developed for automatic length measurement of ERS with locations. • Future AI research is required to classify more types of ERS with their defects. An earth-retaining structure (ERS) is a type of wall widely used to support and protect soil, rock, or other earth materials in many parts of transport infrastructure. ERSs are constructed to create grade separation, particularly at inclined locations or where earth cuttings are required. Over 42 types of ERSs exist in the areas under the jurisdiction of Transport for New South Wales (TfNSW). For effective structural health monitoring, it is crucial to classify an ERS according to its type and establish its GPS coordinates and longitudinal length. The classification is necessary because the defects that occur in the ERS depend on its type. This study proposes a deep learning-based framework that leverages ConvNeXt V2 to classify ERS types and capture the relevant data from vehicle-mounted video capturing systems. The framework integrates spatial–temporal image sequences with GPS coordinates and longitudinal length to enhance context-aware recognition. An improved ConvNeXt V2 model is fine-tuned to learn rich visual features of ERS embedded with GPS metadata. The model is trained and evaluated on a dataset comprising diverse ERS classes captured across TfNSW road transport areas. The experimental results demonstrate significant improvements in classification accuracy and robustness. This emphasizes the potential of the proposed system for scalable, automated ERS inventory and condition assessment.