Recently, the marine transportation of hazardous and noxious substances (HNS) has increased continuously. This has raised the risk of HNS spill accidents significantly and emphasized the issue of creating an emergency response system that is specialized to HNS. However, various types and properties of HNS have made it difficult to respond to the accidents. To develop the response system for HNS accidents, some qualitative considerations and statistical approaches are required. However, the previous studies have limitations due to the infrequency of accidents and lack of real-case data. The present study was performed with 103 cases of global HNS accidents from online database. In addition, 58 cases were nonlinearly projected using a self-organizing map, after some steps of data handling. Information about the date and year of accidents, location, cause, type of HNS, amount of spill, type of vessels, and the age of vessels was gathered and learned by SOM after the handling process. The learning parameters were optimized by trial-and-error, and the results of clustering showed 6 types of HNS accidents. The weight vectors of neurons in different clusters showed differences ways to classify HNS accidents. The type of vessel is a particular priority in the classification process.